JurTrans Blog

How would you explain to your client that the court had disregarded an argument about the conditions in which they might be imprisoned because it could not verify the conclusions you had drawn from a ChatGPT translation? Would pointing out that the court had not identified a single mistranslation make that conversation any easier?

In an interim decision delivered on 16 January 2025 (ECLI:NL:RBAMS:2025:326), the Amsterdam District Court disregarded just such an argument. Defence counsel had used ChatGPT to translate a Polish report about Tarnów prison and relied on it in his written submissions, asking the court to put questions to the Polish authorities about the conditions there. The court replied that it could not verify his conclusions without an official translated version issued by the reporting body or a Dutch translation prepared by a sworn translator. Counsel’s explanation at the hearing of how he had translated the report took his argument no further.

For lawyers using AI translation, the absence of a proven linguistic error offers very little comfort here. Even where a lawyer feels confident that a document has been translated accurately, its contents still have to be established on a basis the court is prepared to act upon. Because counsel had already built his written submissions around the ChatGPT translation, the court’s inability to verify those conclusions meant the submission went nowhere. How much preparation are you willing to commit to an AI translation before establishing whether it can serve the procedural purpose you have in mind?

What counsel was asking the court to do

The proceedings concerned Poland’s request for the surrender of a Polish national under a European Arrest Warrant to serve a two-year custodial sentence. Counsel relied on a Polish-language report from Poland’s National Mechanism for the Prevention of Torture, referred to in the judgment as the NMPT. According to his submission, the report described human rights violations embedded in the prison system at Tarnów. Arguing that his client faced a general danger of human rights violations if detained there, counsel asked the court to seek further information before deciding on surrender.

At the hearing, counsel explained that he had translated the Polish report using ChatGPT and that the conclusions set out in his written submissions followed from that translation. The court’s response was brief: “De rechtbank kan dit echter niet controleren”, meaning that it could not verify this, expressly noting the absence of either an official translation issued by the NMPT or a Dutch translation by a sworn translator. The court was being asked to act on counsel’s understanding of a Polish report without either form of verification.

Counsel was asking the court to make further enquiries rather than inviting it to treat the report as conclusive proof and refuse surrender immediately. Even for that preliminary investigative step, however, his account of the report’s contents proved insufficient, and the court passed over the Tarnów argument. While the proceedings continued on other grounds and this interim ruling did not finally determine whether surrender would take place, there is no need to overstate the procedural outcome; losing consideration of a submission on detention conditions is consequential enough in its own right.

What makes the ruling particularly instructive is the court’s treatment of another Polish prison, Barczewo, elsewhere in the same judgment. There, the court was able to quote English-language passages from NMPT reports describing serious allegations of ill-treatment, examine what those passages supported, and put further questions to the Polish authorities. The same chamber was plainly prepared to investigate concerns about Polish detention conditions where it could examine the underlying material for itself. For Tarnów, the obstacle arose at an earlier stage, when the court needed a verifiable basis for what the Polish report actually said.

What, exactly, was the court supposed to rely on?

The institutional standing of the NMPT could not solve the translation problem. An official monitoring body can be responsible for its Polish report without bearing any responsibility for a version someone else generates using ChatGPT. Unless a lawyer relying on an AI-generated version can establish how the author’s findings have been conveyed, the authority of the original body lends a false appearance of weight to wording it has neither supplied nor approved. Why should a court treat the two as interchangeable?

Once a translation has informed written submissions, counsel has already selected what appears relevant and framed the legal argument around it. Any difficulty in understanding the original text is by then woven into the submission itself, where a careful reading of counsel’s prose will not necessarily reveal it. An argument can read with complete internal coherence on the assumption that the underlying report says what counsel believes it says, yet the court still requires a dependable way to test that assumption.

The rest of the Amsterdam judgment illustrates why the precise wording of the NMPT reports required close inspection. When assessing material on other prisons, the court distinguished accounts of isolated incidents from indications of structural problems, and those distinctions governed whether further enquiries to the Polish authorities were warranted. Assessing whether a report makes a firm finding or describes a continuing practice requires a translation that allows the court to examine the text closely, including where the wording may be less helpful to the party relying on it.

I am not suggesting that ChatGPT confused any of those distinctions in the Tarnów report. The translation does not appear in the judgment, and the court identifies no linguistic error. That silence, however, provides no reassurance about how the submission was prepared. Because the court expressly stated that it could not verify counsel’s conclusions, treating the absence of an identified mistranslation as proof that the translation worked would assume that the very check the court could not perform had somehow been completed.

Suppose someone responds to this decision by asking where the mistranslation was. A judge who cannot read the Polish original is in no position to point to a specific linguistic error, which is the very reason the court requires a verifiable translation in the first place. Expecting the bench to demonstrate a translation mistake before counsel has established a satisfactory basis for the document’s meaning puts the burden upside down. In Amsterdam, the court simply declined to proceed on the basis supplied.

The form of the translation was part of the problem

The court’s reference to a sworn translator deserves more than a passing nod to human review. A lawyer who responds to this judgment by asking a bilingual colleague to cast an eye over future ChatGPT outputs could easily meet the same objection at the next hearing. The court specified an official NMPT translation or a sworn Dutch translation, making no mention of informal internal review. Anyone deciding how to prepare a foreign-language report for court needs to consider what form of translation the tribunal will accept, alongside who is qualified to assess its wording.

At the same time, the court’s willingness to quote English-language NMPT material concerning Barczewo shows that its objection to the Tarnów submission cannot be reduced to a rigid rule that every foreign document must be translated into the language of the court by a sworn translator. Nor did the court have occasion to consider what would happen if a sworn translator used AI tools while taking formal professional responsibility for the resulting text. Those wider questions did not arise; what was missing before the court in this case was stated plainly enough.

A sworn translator can, of course, make an error, and any specific challenge to an official or sworn translation would require proper examination. Yet the fallibility of human translators does not explain why counsel’s unverified ChatGPT account should have sufficed in Amsterdam. Pointing to the possibility of error elsewhere leaves the court’s difficulty untouched, just as a lawyer in any other context would never expect uncertainty about hypothetical evidence to establish the reliability of their own.

Could the court have adjourned or sought a suitable translation of its own motion before setting the Tarnów argument aside? The brief passage in the judgment does not canvas that possibility, though where a submission concerns potential ill-treatment in detention, it is a question worth asking. Equally, we do not know why counsel turned to ChatGPT or what time and financial pressures surrounded the preparation of the defence, so attributing the choice to corner-cutting would be unfair speculation. Even so, none of those considerations helps a lawyer who arrives at a hearing hoping the court will find a way around a translation problem.

Counsel had already relied on it

In my recent discussion of VA v JL, I questioned a New Zealand Review Officer’s reliance on ChatGPT-translated correspondence where the published decision did not explain how the accuracy of the English wording had been established. The Amsterdam ruling shows the other side of the coin: when a court expressly confronts the verification problem and finds no dependable basis for checking the AI translation, it disregards the argument built upon it. Although the two sets of proceedings are very different, both point to the same underlying question for practitioners: how do you expect a decision-maker to verify the meaning you attribute to a foreign-language document?

By the time a lawyer advances submissions on a foreign document at a hearing, an assumption about what that document says has already shaped client advice and case strategy. Discovering at the hearing that the court cannot verify the translation undercuts work already done. I would want to know, well before committing an argument to paper, whether the translation on which it rests will withstand the court’s scrutiny.

An initial machine translation may help a lawyer see whether a foreign report is worth investigating further. Once counsel decides to rely on that report in court, however, securing a translation that satisfies the court’s evidential threshold forms part of preparing the submission itself. Waiting until the hearing to find out whether the court will accept a ChatGPT translation leaves the client exposed to a wholly avoidable procedural risk, and having located a potentially valuable report offers scant consolation if the way it is put before the court prevents the argument from being heard.

The Amsterdam judgment does not tell us whether a sworn or official translation of the Tarnów report would have persuaded the court to put further questions to the Polish authorities. It tells us only that the submission advanced on the basis of counsel’s ChatGPT translation was disregarded. That is the practical reality a lawyer using AI translation needs to weigh before the next hearing, with their client’s position firmly in mind.

Before relying on a translated document in court, can you explain both how the translation was produced and why the judge is entitled to rely on it?

#Litigation #ProfessionalResponsibility #LegalTranslation #LegalTech #AIandLaw #Evidence #LegalRisk #LawyersOfLinkedin

In the software development sector one topic discussed a lot these days is what will companies do when all the senior engineers retire if juniors never get the chance to reach seniority because artificial intelligence is doing the foundational work?

In software engineering, typing lines of code is only the visible surface of the profession. Beneath the syntax lies systems architecture and threat modelling, alongside a deep understanding of how a minor change in one module cascades through an entire enterprise infrastructure. Engineers acquire that architectural judgement by spending their early years writing code from scratch and debugging failures line by line.

The legal profession is now confronting this exact same dilemma. In law, much like in software engineering, the text on the screen is only the final output of a much deeper analytical process. Law firms have traditionally relied on trainees and junior associates drafting agreements, reviewing documents, and researching statutory provisions from the ground up to build the legal reasoning required at partner level. Handing those foundational tasks over to generative AI raises an urgent question for the legal sector: if junior lawyers bypass the formative work of constructing legal arguments from scratch, how will they ever develop the senior judgement needed to spot when an algorithm gets the law wrong?

“Fast Metis”, “Slow Metis”, and the trap of Fluent Mediocrity

Across the legal and professional services world, leading scholars are sounding the alarm over how automation disrupts the formation of senior judgement. In an April 2026 study for the Harvard Law School Center on the Legal Profession, titled The Training Crisis, Anthea Roberts and David Wilkins distinguish between two forms of practical wisdom in legal work: “fast metis” and “slow metis”.

Fast metis is the ability to operate AI tools rapidly to generate plausible-looking text. Slow metis is the deep domain expertise acquired over years of wrestling with primary legal sources and resolving complex statutory or contractual ambiguities from first principles. Roberts and Wilkins demonstrate that when high AI speed is combined with shallow domain mastery, the result is what they term “Fluent Mediocrity”: output that reads smoothly to a non-specialist while missing the binding legal constraints that govern the matter. Because automation targets the very foundational tasks that historically trained junior practitioners, every firm that substitutes AI for rigorous foundational work is consuming the capability pipeline the profession needs to function.

Research from Harvard Business School reinforces this distinction. In The Expertise Upheaval (Sigelman, Fuller, Fenlon et al., July 2025), researchers analysed millions of job roles and divided the knowledge economy into “Mastery Roles” and “Growth Roles”. Mastery Roles rely on front-loaded technical rules that remain relatively static, allowing automation to compress onboarding time. Growth Roles, which include legal practice and specialist advisory work, depend on a learning curve that intensifies over time through exposure to nuanced, judgement-heavy problems.

In these Growth Roles, automating foundational drafting hollows out the middle of the career ladder. As Dr Christophe Kolb observed in Forbes (September 2026), the first draft of a legal memo or complex document historically performed two jobs at once: it shipped, and it taught. Removing the cognitive struggle of building a text from the ground up imposes a hidden “Apprenticeship Tax”. At Stanford Law School’s Rhode Center on the Legal Profession (Engstrom, Knowlton, and George, November 2026), legal scholars and law firm leaders have raised the same question: when an algorithm generates the first draft and marks up the contract, how does a practitioner ever develop the professional judgement required to supervise the machine?

Equally, clients who instruct legal translators for cross-border corporate deals, property acquisitions, or litigation involving Greece and Cyprus face the exact same structural reality. Putting English words on a page is only the final, visible fraction of what a specialist legal translator actually does. When law firms and corporate boards buy into the narrative that generative AI can handle the “heavy lifting” of translation while a human merely checks the output, they mistake the surface artefact for the underlying legal engineering. Worse still, they rely on a market model that is actively eroding the senior expertise required to keep their transactions safe.

Why “Getting Words on the Page” is the smallest part of legal translation

For a foreign board of directors reviewing a Greek commercial contract, or an English solicitor advising on a Cypriot probate or corporate dispute, the primary obstacle appears to be language: the document is in Greek, and it needs to be in English. Because large language models and neural machine translation engines produce grammatically polished English prose in seconds, it is tempting for them to assume that the machine has done 90 per cent of the job.

Some recent empirical research into how legal translators actually work proves the opposite. Putting words on the page is simply the recording mechanism for a series of diagnostic and comparative legal decisions that happen before a sentence is finalised:

  • Testing legal propositions rather than reading surface grammar: In a December 2025 keylogging experiment using Translog-II software (Sayar Öztürk and Kumcu, 2025), researchers measured the cognitive and temporal effort of experienced translators alongside novices and field specialists as they worked on machine-translated legal documents. The inexperienced participants moved through the machine output quickly with brief pauses, reassured by the smooth grammar on the screen. The experienced translators exerted significantly higher cognitive and temporal effort, recording longer pauses per word and executing deep structural revisions. Because they possessed senior comparative judgement, they were looking past the words on the page to test whether the underlying legal obligations held up across jurisdictions.

  • Coordinating legal context with procedural reality: An October 2026 empirical study by Hu examined what happens when law students and translation trainees are given unrestricted access to AI tools while revising machine-translated legal texts. Despite having AI assistants and authoritative dictionaries at their fingertips, the participants repeatedly failed to correct critical legal misalignments. Having access to a dictionary equivalent or an AI suggestion could not compensate for the inability to distinguish between a term that is valid legal English in the abstract and one that is legally operative in the specific procedural context of the instrument.

Automated systems treat every word and clause in a document as a uniform statistical calculation. A senior human specialist, by contrast, reads a forty-page Greek pleading or regulatory filing with an eye for where the commercial or litigation danger sits. When a single verb or statutory reference carries the entire financial exposure of a deal, a senior lawyer-linguist recognises the stakes immediately, cross-checking Greek civil law authorities and alerting instructing counsel to ambiguities in the Greek original before any commercial step is taken.

The monolingual buyer’s blind spot: Correspondence errors and “Adverse selection”

Why does this distinction between surface words and deep legal judgement create such acute exposure for buyers of legal translation? The answer lies in information asymmetry.

In a September 2025 AI safety report published by Tranquality /LTAC Global (Why Translation Errors Must Be Part of the AI Safety Discussion), Dr Alan Melby and Ryan Foley explain why AI translation failures belong in a different risk category from monolingual AI errors. When a solicitor in London uses generative AI to summarise an English judgment, the solicitor reads English and can verify the summary against the source text. When that same solicitor or a foreign investor receives an English translation of a Greek or Cypriot legal document, they are operating in an asymmetric bilingual scenario. Because they cannot read the Greek source text at a specialist legal level, they have no way of fact-checking whether the English text accurately reflects the Greek original.

Melby and Foley identify correspondence errors (where the translated text is completely fluent and plausible in English while materially misrepresenting the source text) as the primary vector of legal and financial harm in cross-border workflows. Compounded by “fluency bias” (the human cognitive tendency to equate smooth syntax with factual accuracy), correspondence errors remain invisible to the client until a counterparty challenges a clause in court or a regulator rejects a filing.

This information gap triggers what economists call adverse selection. When buyers cannot visually distinguish between a cheap, machine-assisted draft and a rigorously researched legal translation (because both look like neat pages of English words), price becomes the primary basis for comparison. Low-cost providers offering rapid machine output undercut and crowd out thorough legal translation, while the mass market drives down fees and replaces senior specialists with junior post-editors checking machine output at speed. Yet, as the keylogging and revision studies demonstrate, a reviewer who has not spent years translating complex legal instruments from the ground up cannot spot the correspondence errors hidden beneath the fluent surface.

Why seniority cannot be automated

When a law firm, corporate board, or private client instructs a legal translator for a matter involving Greece or Cyprus, the deliverables are English (or Greek) words on a page. What the client is actually purchasing, however, is the decades of “slow metis” that determine which words belong there and what legal consequences they trigger.

That senior judgement is built through formal legal education and academic research, combined with thousands of hours of primary legal translation and direct engagement with Greek and Cypriot statutes, case law, and notarial or corporate practice. As we explain in our guides on Why AI Legal Translation Is a False Economy and Why Clients Choose JurTrans, senior comparative judgement cannot be replicated by a predictive algorithm, and it cannot be safeguarded by a junior reviewer skimming a machine-generated draft.

In software engineering and in cross-border law alike, there is no shortcut to seniority. When the stakes involve enforceable obligations, property rights, regulatory compliance, or judicial proceedings, the only safeguard against fluent mediocrity is a practitioner who has done the hard yards to master both the law and the language from first principles.

At JurTrans, every Greek and Cypriot legal translation is handled directly at a senior level by a specialist who thinks and writes like a lawyer and understands the practical mechanics of how Greek, Cypriot, and English legal systems interact. Whether you are a foreign board overseeing a Greek subsidiary, a law firm coordinating cross-border litigation, or a private client navigating property or inheritance documentation in Greece or Cyprus, our work gives you clear, reliable visibility into your rights, obligations, and commercial exposure. If you need Greek or Cypriot legal documents translated with rigorous legal precision, get in touch with JurTrans to discuss how we can assist.

The Commercial Promise Versus Courtroom Reality

Commercial language service providers frequently present barristers and solicitors with an appealing narrative regarding the maturity of translation technology in international litigation and transactional work. A representative example appears in a 2023 contribution to Barrister Magazine, which contends that machine translation saves considerable time and money across high-volume caseloads, that output accuracy depends largely on database size and customised engine training, and that pairing raw machine output with human post-editing guarantees 100 per cent accuracy for instructing counsel. For legal practitioners under pressure to review foreign disclosure bundles or cross-border transactional documentation within compressed timetables, this hybrid workflow is marketed as a reliable compromise between automated speed and human assurance.

A technorealistic assessment of legal translation technology requires separating low-risk information management from high-stakes legal representation. Neural machine translation engines and generative artificial intelligence models can assist legal teams with initial document triage or internal gisting when counsel simply need to ascertain the general subject matter of a foreign file. Relying on machine output as the foundational draft for evidential documents, statutory provisions, or binding commercial agreements introduces systemic risks that post-editing struggles to remediate. Empirical research in legal translation studies and comparative law, particularly across the Greek civil law and English common law pair, demonstrates that each pillar of the commercial post-editing narrative rests on a flawed conception of how legal language functions across jurisdictions.

Why Database Size and Customised Engine Training Fail to Guarantee Legal Accuracy

The commercial argument for hybrid legal translation rests heavily on the assumption that translation quality is a function of corpus volume and continuous machine training. Proponents suggest that feeding millions of bilingual legal texts into a neural engine, combined with ongoing corrections from human editors, equips the software with a substantive grasp of legal terminology and phrasing. This premise overlooks the fundamental architecture of both neural machine translation and large language models. As Argyri Panezi and I (2026) observe alongside legal technology scholars such as Eliza Mik (2026), automated systems process tokens on the basis of statistical probability and distributional frequency across training corpora. They lack any capacity to comprehend legal meaning, legislative intent, or jurisdictional consequences.

Legal terminology is inherently system-bound, as Susan Šarčević (1997) and Deborah Cao (2007) have long established in comparative legal linguistics. Every national legal order develops its own conceptual architecture shaped by distinct historical and doctrinal evolutions. In their corpus-driven investigation of Greek notarial deeds and their English translations, Vilelmini Sosoni and John O’Shea (2021) examined the transfer of property law concepts between Greek civil law, which draws on Roman-Byzantine traditions and the German Civil Code (Bürgerliches Gesetzbuch), and English common law, which remains rooted in feudal principles of Crown landholding and estates in land. Across the ten most frequent property law terms extracted from twenty Greek notarial conveyancing deeds, Sosoni and O’Shea found zero instances of exact or near equivalence between Greek and English. Half of the core terms exhibited only partial equivalence, while the remaining half represented complete conceptual non-equivalence, requiring careful descriptive paraphrases or functional adaptations grounded in comparative legal analysis.

Empirical testing confirms that expanding a legal database fails to resolve this inter-systemic incongruity. In a recent comparative evaluation, Vilelmini Sosoni and John O’Shea (2026) tested a domain-adapted commercial neural machine translation engine alongside leading generative artificial intelligence models (ChatGPT-4o, Gemini 1.5 Flash, and Claude 3 Haiku) on two authentic Greek legal texts: an extract from Greek planning and environmental legislation (Law 3937/2011 and Law 4280/2014) and a notarial property sale agreement. The domain-adapted neural engine had been trained on a substantial parallel corpus of 10.5 million sentence pairs, of which 5.3 million sentence pairs (50.5 per cent) consisted specifically of Greek-to-English legal and legislative texts.

Despite this extensive domain-specific training, the customised legal engine recorded the worst overall severity index of all systems evaluated across both the legislative text (4.54) and the property sale agreement (4.33). It also produced the highest proportion of Critical accuracy errors, reaching 21.05 per cent in the statutory extract and 10.26 per cent in the conveyancing contract. The generative artificial intelligence models similarly struggled with legal precision, with accuracy errors accounting for between 51.11 per cent and 60.38 per cent of all annotated errors across every model in the legislative text.

The qualitative failures documented in Sosoni and O’Shea (2026) illustrate the specific legal liabilities created when statistical token prediction encounters system-bound concepts:

  • Polarity reversals in statutory provisions: In the legislative text, the domain-trained legal engine inverted the operative force of a mandatory rule by translating the Greek verb ισχύει (“applies to”) as “shall not apply to”. In litigation or regulatory compliance, reversing a statutory prohibition into an exemption alters the entire legal position of a party.

  • Temporal shifts in contractual warranties: In the Greek property sale agreement, the vendor warranted that the real estate was presently free of encumbrances and third-party claims (ελεύθερα). The domain-adapted engine rendered this present warranty of title as a future obligation (“shall be free and exempt”), altering the allocation of transactional risk at completion.

  • Conflation of civil and criminal legal institutions: Every neural and generative system evaluated in the study mistranslated the Greek civil enforcement term κατάσχεση (civil attachment or distress levied against property for debt recovery) as “seizure”, importing a criminal or public-law concept into a private conveyancing instrument.

  • Conceptual flattening of property rights: Distinct Greek civil law concepts were routinely reduced to vague everyday expressions. The term κατοχή (physical detentorship under the Greek Civil Code, as distinct from legal possession or νομή) was flattened into “occupancy” or “control”, while οροφοκτησία (horizontal ownership of an entire floor under Law 3741/1929) was stripped of its statutory specificity and rendered simply as “ownership”.

These findings converge with empirical research by Stavroula Paraskevi Vraila (2026), who analysed 400 translations of system-bound legal terms into Greek across DeepL, Google Translate, GPT-5.3, and Claude Opus 4.5. Vraila demonstrated that because commercial models rely overwhelmingly on Anglophone training corpora or English pivot representations, they exhibit high rates of English-mediated terminological displacement, averaging 57.9 per cent across neural engines and 28.5 per cent across generative models. Fundamental Romanistic property distinctions between νομή (possession) and κυριότητα (ownership) were systematically collapsed by the majority of engines into imprecise common law approximations, substituting the doctrinal categories of one legal tradition for those of another.

The Post-Editing Myth: Deceptive Fluency and the Absence of Productivity Gains

When commercial providers concede that raw machine translation cannot be trusted in isolation, they invariably present human post-editing as the safeguard that restores 100 per cent accuracy while preserving time and cost savings. This claim ignores a well-documented cognitive hazard in neural translation workflows known as deceptive fluency. Earlier statistical machine translation produced visibly broken syntax that immediately alerted a reviewing lawyer or linguist to underlying mistranslations. Contemporary neural engines and large language models are optimised through alignment techniques such as reinforcement learning from human feedback to generate smooth, idiomatic target prose (Ouyang et al., 2022; Vraila, 2026). As Elisabeth Wiesmann (2019) and Karolina Stefaniak (2020) have shown, grammatical fluency conceals severe semantic departures from the source text, inducing automation bias in human reviewers who become far less likely to detect subtle omissions, fabricated qualifiers, or shifted legal obligations.

Empirical process research refutes the claim that post-editing delivers efficiency or superior accuracy in specialised legal translation. In an experimental study involving professional translators working from Greek into English on property law texts, Vilelmini Sosoni, John O’Shea, and Maria Stasimioti (2022) measured temporal, technical, and cognitive effort using keylogging and eye-tracking data to compare translation from scratch against the post-editing of neural machine translation. The investigation revealed zero productivity gains from post-editing in legal texts. Furthermore, working over raw machine output introduced substantial cognitive friction: translators primed by the machine draft committed more verity and style errors than when they constructed the legal translation directly from the Greek source text.

The quantitative metrics from Sosoni and O’Shea (2026) explain why repairing machine output consumes so much professional effort. Across the four neural and generative systems tested on the Greek legislation and the property sale agreement, Translation Edit Rate (TER) scores ranged between 56.56 per cent and 74.42 per cent. A human reviewer confronted with a machine draft in which well over half to nearly three-quarters of the text requires structural modification must spend extensive time unpicking plausible-sounding errors before reassembling the syntax to match English legal drafting conventions.

In the contractual text analysed by Sosoni and O’Shea (2026), style errors alone accounted for between 41.38 per cent and 43.75 per cent of all errors produced by the generative artificial intelligence models. Vraila (2026) similarly found that stylistic flattening affected between 18 per cent and 85.7 per cent of legal outputs across systems, eroding the formal register conventions that carry performative force in legal instruments. Unravelling a flattened, conceptually displaced machine draft demands prolonged verification against the source text, eliminating the promised commercial speed while increasing the probability that a camouflaged terminological error survives into the final court bundle or closing set.

The Human-in-the-Loop Fallacy, Professional Liability, and Client Confidentiality

The commercial promotion of hybrid workflows relies heavily on the reassuring label of the “human-in-the-loop”. In their examination of legal liability and ethical governance in automated translation, Argyri Panezi and John O’Shea (2023) demonstrate that this phrasing misrepresents the operational and legal reality faced by the specialist translator. When a lawyer-linguist translates a pleading or contract from scratch, they exercise full epistemic and professional control over every interpretive decision from the opening recital to the execution clause. When placed at the end of an automated pipeline to post-edit output generated by an opaque neural architecture, the linguist forfeits control over the cognitive baseline of the draft while remaining expected to assume full contractual and professional liability for the accuracy of the delivered text.

Practising lawyers who handle bilingual legal documentation recognise these hazards in their daily work. In a survey-based empirical study examining how translating lawyers in Greece adopt and perceive machine translation, Kyriaki Kourouni and John O’Shea (2025) found that two-thirds of respondents avoided machine translation altogether, electing to translate legal texts manually. Among the minority of legal practitioners who did use machine translation, 61.55 per cent restricted its use to occasional, low-stakes tasks, such as obtaining a rough gist of a foreign text (9.28 per cent) or drafting routine correspondence with international clients (8.25 per cent). Practitioners rejected the notion that automated output could serve as a reliable basis for court submissions or binding transactional instruments.

The survey by Kourouni and O’Shea (2025) also highlighted severe vulnerabilities surrounding data protection and professional ethics. Although commercial articles frequently caution against free web-based translators, 73.33 per cent of the surveyed lawyers who used machine translation relied on free Google Translate. While 35.72 per cent of respondents acknowledged that uploading client texts to external servers risked breaching confidentiality obligations, far fewer connected the use of automated translation to professional negligence, malpractice exposure, or breaches of the duty of care owed to clients (Wahler, 2018; Giampieri, 2023). Even where paid application programming interfaces offer contractual assurances regarding data retention, feeding sensitive corporate disclosures or litigation strategies into third-party generative models raises unresolved questions under data protection regimes and professional conduct rules.

Systemic Language Pollution and the Efficiency Paradox in Cross-Border Practice

The widespread marketing of machine translation as a routine component of legal case management carries systemic consequences that extend beyond individual translation errors. In their analysis of artificial intelligence and legal epistemology, Argyri Panezi and John O’Shea (2026) introduce the concept of “language pollution” to describe how unverified machine translations, fabricated citations, and algorithmically mediated legal transplants degrade the linguistic and doctrinal integrity of the legal system. Because law is constituted through authoritative texts and precise interpretive traditions, injecting statistically generated approximations into legal workflows contaminates the shared informational environment relied upon by counsel, transactional parties, and the judiciary.

Rather than streamlining cross-border representation, uncritical reliance on automated language tools triggers an efficiency paradox in legal practice (Evans, 2025; Panezi and O’Shea, 2026). When clients, overseas instructing solicitors, or commercial agencies use generative models or hybrid machine workflows to translate and interpret foreign law, they frequently construct entirely false legal premises that specialist counsel must subsequently dismantle.

Panezi and O’Shea (2026) document a telling empirical demonstration of this phenomenon. When asked in June 2025 and again in November 2025 to identify the Greek legal rules governing the construction of boundary fences between neighbouring properties in Athens, ChatGPT produced confident, structured legal memoranda citing Articles 1003 to 1010 of the Greek Civil Code. Upon inspection against the actual statutory text, those provisions have no bearing on the erection of boundary walls: Articles 1003 to 1005 regulate emissions of smoke, soot, and noise; Article 1006 addresses the risk of collapsing structures; Article 1007 governs digging near a boundary; Article 1008 deals with encroaching tree roots and branches; Article 1009 regulates fruit falling onto a neighbour’s land; and Article 1010 addresses construction encroaching onto neighbouring land in good faith. When pressed for supporting authority, the model fabricated non-existent case law from the Greek Supreme Civil and Criminal Court (Areios Pagos) alongside invented planning statutes. When machine-generated translations or summaries of this kind enter a cross-border file, experienced practitioners lose substantial time correcting fabricated statutory frameworks and re-educating clients before genuine legal work can proceed.

The Lawyer-Linguist Standard in Cross-Border Legal Practice

The empirical record from Greek-to-English legal translation research refutes the commercial narrative that larger databases, customised engines, and human post-editing can replicate the work of a specialist legal translator. Statistical models cannot bridge the conceptual divide between civil law and common law systems because resolving terminological incongruity requires comparative legal reasoning rather than probabilistic token matching. Post-editing fluent machine prose fails to save time in complex legal genres and exposes legal teams to camouflaged polarity reversals, displaced property and procedural concepts, and unquantified professional liability.

Cross-border litigation, corporate acquisitions, and real estate conveyancing involving Greece and Cyprus require a translation partner grounded in the substantive law of both jurisdictions. At JurTrans, every instruction is handled directly by a specialist lawyer-linguist with deep familiarity with the nitty-gritty of the legal systems involved, thinking and writing like a lawyer to ensure that complex civil law concepts are rendered into precise, authoritative English. Through rigorous attention to detail and direct human scholarship from the first word to the final review, JurTrans delivers translations that are entirely free of fabricated content and fit for the demands of the Bar and international legal practice.

Editorial note: This article examines empirical research on translation methodology and technology risks for general informational purposes; nothing written herein constitutes legal advice.

Ten Years from Google Translate to ChatGPT: Why Machine and AI Legal Translation Carries Severe Consequences for Clients and Lawyers

By John O’Shea (LL.B. (Hons), LL.M.)

1. A Decade of Technological Evolution, The Same Underlying Hazard

Across commercial law firms, corporate legal departments, and private litigation, a quiet assumption has taken hold that pasting foreign-language correspondence, witness statements, statutory provisions, or commercial agreements into a machine translation engine or generative AI chatbot carries no real consequence. Users frequently rationalise this practice by telling themselves that the output is merely for preliminary orientation, internal comprehension, or an informal check before instructing external counsel.

Empirical research into legal workflows confirms how deeply entrenched this attitude has become. In a survey of practising lawyers in Greece that I co-authored with Kyriaki Kourouni in Revista Tradumàtica (2025), practitioners reported turning to machine translation primarily to grasp the basic meaning of foreign texts (9.28%), manage cross-border client communications (8.25%), review internal documents (7.22%), and examine draft commercial contracts (6.19%). Strikingly, whilst 35.72% of the surveyed lawyers recognised that uploading legal material into free online translation tools breaches client confidentiality, 0% of respondents expressed any fear of being sued for professional malpractice or legal liability as a consequence of relying on machine output.

A decade of case law across common law and civil law jurisdictions demonstrates that this complacency is misplaced. Ten years ago, litigants who placed unverified Google Translate outputs before tribunals watched their evidence get rejected and their claims fail because statistical machine translation produced visibly broken syntax. Today, ten years down the line, the transition to neural machine translation (NMT) and large language models (LLMs) such as ChatGPT has failed to eliminate the underlying jurilinguistic hazard. By replacing obvious syntactic incoherence with polished, authoritative-sounding prose, modern artificial translation has made substantive errors far harder to detect, expanding the fallout to encompass both the client whose substantive rights are compromised and the legal practitioner whose professional conduct comes under judicial or disciplinary scrutiny.

2. Why Fluent AI Output Traps Both Clients and Counsel

To understand why modern AI translation creates acute legal exposure, one must look at how language models process legal text. In earlier generations of machine translation, defective output announced itself through garbled grammar that immediately warned a careful lawyer or company director to treat the text with caution. As Juliette Scott and I demonstrated in our Legal Translation Vulnerability Scoping framework (Research Handbook on Jurilinguistics, Edward Elgar, 2023), neural networks and generative AI replaced that visible awkwardness with an illusion of fluidity.

Machine learning systems optimise for probabilistic plausibility rather than comparative legal equivalence. When translating between legal orders that do not share the same conceptual architecture, such as the continental European civil law tradition of Greece and the common law systems of England and Wales or the United States, surface fluency masks severe doctrinal distortions. In our chapter Words Matter: How AI Pollutes the Language of the Law (International Handbook of Legal Language and Communication, Springer Nature, 2026), Argyri Panezi and I analysed how English-dominated training corpora introduce systematic jurisdiction bias and technology-mediated legal transplants into translated legal texts.

Empirical studies of Greek-to-English and English-to-Greek AI outputs illustrate how easily these systems flatten or displace statutory concepts:

  • Criminal and Regulatory Classification: In Greek criminal and regulatory compliance matters, automated systems routinely collapse the statutory tripartite division of offences under the Greek Penal Code (κακούργημα, πλημμέλημα, and πταίσμα) into generic English words such as “crime” or “offence”, or import Anglo-American common law concepts by rendering δόλος (dolus) as mere “intent” and αναστολή ποινής (suspension of sentence) as “probation”.

  • Civil and Corporate Acts: In civil and commercial contexts, neural translation systems have rendered a private-law legal transaction or juridical act (δικαιοπραξία) as “jurisdiction” (δικαιοδοσία), and reduced a formal criminal complaint subject to strict statutory time limits (έγκληση) to an informal “report”.

  • Appellate and Procedural Finality: As Daniel Webber and I documented when translating the Greek procedural codes (Translating Greece’s Procedural Codes, ELETO, 2023), even traditional legal dictionaries and automated tools routinely flatten the distinct procedural stages of Greek court decisions into the single English phrase “final judgment”. Under the Greek Code of Civil Procedure, however, an οριστική απόφαση disposes of the action before the first-instance court whilst remaining subject to ordinary appeal, a τελεσίδικη απόφαση denotes a judgment where ordinary appellate remedies have been exhausted or expired, an ανέκκλητη απόφαση refers to a decision that is by law not amenable to ordinary appeal, and an αμετάκλητη απόφαση signifies an irrevocable judgment no longer open to cassation before the Hellenic Supreme Court (Areios Pagos). A translation that renders all four categories simply as a “final judgment” misleads foreign counsel on the fundamental question of whether a judgment can still be challenged or enforced.

Furthermore, as Pierre Moor (2026) has observed in the context of legal interpretation, statutory and contractual provisions frequently contain deliberate normative ambiguity or open-textured concepts whose scope must be resolved by a court rather than a translator. When a generative AI model encounters an ambiguous source provision, it silently selects the most statistically probable English phrasing, committing the reader to a single substantive interpretation without alerting counsel that the original text admitted more than one reading.

Nor can this problem be solved simply by asking a reviewer to “post-edit” machine output. As Vilelmini Sosoni, Maria Stasimioti, and I demonstrated in our empirical comparison of from-scratch human translation and machine translation post-editing (Journal of Language and Law, 2022), the grammatical smoothness of neural output primes even trained reviewers to overlook major terminological and accuracy errors, creating a false sense of security whilst shifting full professional liability onto the human who signs off on the text.

3. Ten Years of Machine Translation in the Courts: From Early MT to Modern AI (2016–2026)

If technological progress had solved the risks of automated legal translation, judicial decisions criticizing defective machine translations would have faded into history after the mid-2010s. Instead, the case law from 2016 to 2026 shows a continuous line of evidentiary failures across tax tribunals, commercial courts, and statutory tribunals.

The 2016 Benchmark: Rejected Evidence and £243,564 in Tax Liability

Ten years ago, two decisions before the UK First-tier Tribunal (Tax Chamber) illustrated the immediate financial and procedural cost of relying on unverified translations in cross-border proceedings:

  • KIP UK Ltd v Revenue and Customs [2016] UKFTT 821 (TC): In a customs duty dispute, a party sought to rely on English translations generated through Google Translate. The Tribunal upheld the opposing party’s formal objection to the evidence, ruling that the machine translation “did not produce any meaningfully comprehensible summary” upon which a judicial body could rely.

  • Essential Telecom Ltd v Revenue and Customs [2016] UKFTT 475 (TC): In a cross-border value-added tax dispute involving £243,564, HM Revenue and Customs relied on unverified translations of Italian Guardia di Finanza investigation reports concerning an Italian counterparty. The UK Tribunal refused to admit the disputed evidence, characterising it as double hearsay and finding that the defective translations had caused “many wrong assumptions about the corporate or non-corporate form of the counterparty”.

Ten Years Later: NMT and ChatGPT in Commercial and Statutory Litigation

Move forward a decade to the era of neural machine translation and generative AI, and litigants are encountering the exact same judicial wall, now compounded by AI fabrication:

  • Euroboor BV v Grafova: In commercial litigation before a United States court, a party submitted a Google Translate version of a Dutch court judgment on which its legal argument relied. Despite a decade of neural network upgrades since 2016, the American court struggled to decipher the judicial reasoning in the machine-translated Dutch judgment, demonstrating that raw MT remains incapable of conveying complex foreign judicial reasoning to a common law bench.

  • Lozano González v Roberge: Before a Canadian tribunal, a litigant relied on a ChatGPT translation of foreign statutory provisions to support their legal case. Upon scrutiny, the tribunal found that ChatGPT had not merely translated the legislation inaccurately; the chatbot had effectively rewritten the statutory provisions on which the litigant’s argument depended.

This phenomenon of statutory distortion aligns directly with the empirical test reported in Panezi & O’Shea (2026), where ChatGPT, queried in English regarding a civil dispute under Greek law, fabricated confident English translations and doctrinal interpretations of Articles 1003 to 1010 of the Greek Civil Code. Across the globe, more than 2,000 judicial decisions have now documented the submission of AI-fabricated legal authorities and distorted statutory texts, a trend that has begun to trigger direct commercial litigation against AI vendors themselves, such as Nippon Life v OpenAI (filed in March 2026).

4. Consequences for the Client: Substantive Loss and ‘Terminology Drift’ in Commercial Disputes

For corporate and private clients, the danger of casual or uncoordinated translation extends well beyond a single rejected exhibit. Commercial relationships and cross-border investments evolve over many years. Agreements are negotiated and signed, board minutes and financial disclosures under Greek Law 4548/2018 are recorded, correspondence is exchanged when commercial friction arises, and extrajudicial notices are served when the relationship breaks down.

A recent decision of the Hellenic Supreme Court, Judgment No. 221/2026, provides a textbook demonstration of how translation choices made at different stages of a commercial venture can become a central forensic issue in high-value litigation.

The dispute arose out of a long-running wind-energy project in Greece between a Greek project developer and a Spanish corporate investor. Although the wind-energy project was situated in Greece, the commercial relationship between the parties was governed by contracts drafted in English. When the commercial relationship broke down, the Spanish investor served an extrajudicial notice declaring that it had decided to “terminate the … Contracts and withdraw from the Project”, proposing to transfer its shares in the project company for a nominal consideration pursuant to Article 5.2 of the Framework Agreement.

When the dispute reached the Greek courts, the legal characterisation and validity of that notice turned heavily on how the English contractual terms had been translated into Greek across the documentary record:

  1. Inconsistent Rendering of the Core Contractual Mechanism (“Withdrawal”): In translating the Framework Agreement, the Greek developer rendered the heading “Withdrawal” as Υπαναχώρηση (the Greek civil law concept of statutory or contractual rescission/withdrawal). The Spanish investor’s translation of the heading used the more general word Απόσυρση, yet within the operative clauses of the Spanish investor’s own translation, the phrase “withdraw from their obligations” was rendered using υπαναχώρηση, while a third variant, αποχώρηση (departure/exit), also appeared in the arguments. The Court of Appeal expressly noted this terminological inconsistency across the translations before concluding, from the contractual mechanism of the share transfer, that “withdrawal” operated in Greek law as contractual υπαναχώρηση.

  2. Disputed Translation of Commercial Thresholds (“Feasibility”): Under Article 3.5 of the Project Agreement, the investor’s right to withdraw was conditional upon the project’s “technical and financial feasibility” proving insufficient. One translation rendered “feasibility” as βιωσιμότητα (viability/sustainability), whereas the Court of Appeal held that σκοπιμότητα (expediency/feasibility in context) was the appropriate rendering within the architecture of the clause. Ultimately, the courts found that the investor’s notice failed to specify whether the alleged inadequacy was technical, financial, or both, and failed to identify the supporting data, resulting in the dismissal of the investor’s cassation appeal before the Supreme Court.

Hellenic Supreme Court Judgment No. 221/2026 illustrates what happens when translation is treated as an isolated lexical exercise rather than a file-level legal strategy. When a corporate client or legal team feeds a newly received extrajudicial notice, board resolution, or commercial letter into an AI tool or commissions an ad hoc translation without cross-referencing the terminology established years earlier in the underlying framework agreement, terminology drift enters the file. Because generative AI translates probabilistically, it will readily render an ordinary-looking English word like “withdrawal”, “feasibility”, “material”, or “default” using one Greek synonym in 2021 and a different Greek term in 2026. Once those documents are placed side by side before a court, the client has to pay counsel to litigate avoidable ambiguities created by their own translations.

5. Consequences for the Lawyer: Judicial Criticism, Disciplinary Exposure, and Malpractice

While clients bear the commercial and substantive fallout of defective translations, recent decisions from 2025 and 2026 demonstrate that practising lawyers face acute reputational, procedural, and disciplinary exposure when AI-translated material enters a legal file.

Judicial Rebuke for Overlooking AI Provenance: AA v Secretary of State for the Home Department [2026]

In AA v Secretary of State for the Home Department [2026] UKFTT PA747122024 (IAC), an appellant’s solicitors filed an English witness letter in support of an asylum appeal before the UK First-tier Tribunal (Immigration and Asylum Chamber). The English text read fluently and appeared to support the appellant’s factual account. However, at the end of the witness letter sat a conspicuous sentence offering to turn the letter into a formal affidavit, insert notarisation language, or make the wording more neutral.

First-tier Tribunal Judge CJ Williams immediately recognised the sentence for what it was: a conversational follow-up prompt generated by a chatbot. Upon questioning by the judge, the witness explained that he had written his account in Arabic, used ChatGPT to translate it into English, and then asked his daughter to compare the English output against the Arabic original to check that the meaning was the same.

At paragraph [28] of the judgment, Judge Williams delivered a pointed observation directed at the legal representatives who had placed the document into the court bundle:

“It is unfortunate that it fell to the Tribunal to notice this.”

For litigation solicitors and counsel, AA exposes two distinct professional vulnerabilities:

  • Failure of Basic Document Control: Detecting the leftover ChatGPT prompt at the foot of the witness statement required no knowledge of Arabic. It required only that the legal team read the English exhibit carefully from start to finish before filing it with the Tribunal.

  • The Fallacy of Casual ‘Human Involvement’: When lawyers defend the use of AI translation, they frequently point to the presence of a “human in the loop”. In AA, the witness’s daughter had reviewed the ChatGPT translation against the Arabic original. Yet as the case demonstrates, the mere fact that a bilingual relative, client employee, or junior staff member has looked over an AI translation tells a court nothing about their competence to identify subtle shifts in evidentiary certainty, omitted qualifications, or legal terminology. If a legal team misses an obvious English trace of ChatGPT on the face of a document, it has no credible basis for assuring a judge that the underlying foreign-language evidence has been reliably verified.

Professional Misconduct Findings Based on Unverified ChatGPT Translations: VA v JL [2025]

Even more concerning for legal practitioners is the New Zealand disciplinary decision in VA v JL [2025] NZLCRO 123, which arose from a client’s complaint regarding her lawyer’s handling of an employment and commercial dispute.

A central issue before the Legal Complaints Review Officer was whether the lawyer had misled her client by advising that commencing High Court proceedings could be used to halt or disrupt proceedings before the Employment Relations Authority. Crucially, the disputed advice was contained in chat messages exchanged between the lawyer and the client on 23 July 2021 in a foreign language (anonymised in the decision as “Language F”).

The client submitted a fuller record of the foreign-language chat exchange accompanied by English translations that the client stated she had generated using ChatGPT. In the ChatGPT-translated version, the lawyer was recorded as stating: “Yes, it will work” and “It does work! This is feasible!” The lawyer denied sending the disputed message and provided a shorter record with her own translations.

In making findings of unsatisfactory conduct against the lawyer, the Review Officer preferred the completeness of the client’s documentary record at paragraph [112] and expressly relied at paragraphs [208] and [209] on the statements “Yes, it will work” and “It does work! This is feasible!” when assessing the legal advice the lawyer had given. Yet nowhere in the published decision did the Review Officer explain how he established that the client’s ChatGPT translation accurately reflected the nuances of what had been written in Language F.

In legal advice concerning litigation strategy, the boundary between telling a client that a procedural route is theoretically possible and assuring the client that it will succeed turns on fine linguistic distinctions of modality and qualification. VA v JL serves as a stark warning to practising lawyers on two fronts. First, any foreign-language message you send to a client may subsequently be run through ChatGPT by that client and placed before a disciplinary tribunal as evidence of professional misconduct. Secondly, decision-makers who cannot read the source language may conflate the authenticity of a foreign-language record with the accuracy of its fluent English AI translation, treating unverified chatbot phrasing as proof of what a lawyer advised.

Confidentiality Breaches and the Vacuum of Vendor Liability

Beyond evidentiary and disciplinary perils in court, lawyers who upload client contracts, corporate disclosures, or pleadings into public machine translation or generative AI platforms face direct regulatory and contractual exposure. As Argyri Panezi and I examined in our paper for the International Conference on Machine Learning (Workshop on Generative AI and Law, ICML, 2023) and as the High Court of England and Wales emphasised in Ayinde v London Borough of Haringey [2025] EWHC 1383 regarding lawyers’ non-delegable supervisory duties over AI, practitioners remain bound by strict statutory duties of professional secrecy, UK and EU General Data Protection Regulation (GDPR) obligations, and contractual Non-Disclosure Agreements (NDAs).

When an unverified AI translation causes financial loss or misses a statutory condition, the terms of service of every commercial AI and machine translation provider disclaim all warranties of accuracy and fitness for purpose. As I explored at the joint EULITA and PEEMPIP conference in Athens (Is Anyone Liable for Legal Translation?, 2024), technology vendors accept zero liability for the legal consequences of their outputs. The entire burden of professional negligence, regulatory sanction, and financial loss falls upon the instructing lawyer or corporate user who chose to rely on the machine.

6. Frequently Asked Questions

Has modern generative AI solved the legal translation errors seen in earlier tools like Google Translate?

No. While neural machine translation and generative AI models produce much smoother English sentences than statistical machine translation did ten years ago, empirical research shows that high fluency masks severe accuracy and terminological errors. Recent court decisions such as Lozano González v Roberge (where ChatGPT rewrote statutory provisions) and Euroboor BV v Grafova (where a US court struggled to understand a Google Translate rendering of a Dutch judgment) confirm that modern AI tools still fail to preserve legal meaning in contentious and transactional settings.

What are the risks to a lawyer who submits or relies on an AI-translated document in legal proceedings?

A lawyer who relies on or files an unverified AI translation risks judicial criticism for failing to check document provenance (as seen in AA v Secretary of State for the Home Department [2026] UKFTT PA747122024), file-level terminology drift across related contractual and litigation documents (as highlighted in Hellenic Supreme Court Judgment No. 221/2026), breaches of client confidentiality and GDPR obligations when uploading sensitive material to third-party servers, and full professional liability for any resulting loss. Furthermore, as VA v JL [2025] NZLCRO 123 illustrates, lawyers also face disciplinary risk when unverified ChatGPT translations of client communications are relied upon in professional conduct proceedings.

Is having a bilingual colleague or family member check a ChatGPT translation sufficient for legal purposes?

No. Reviewing an AI-generated translation against a foreign-language original requires specialist competence in both comparative law and legal translation. Empirical studies of machine translation post-editing (Sosoni, O’Shea & Stasimioti, 2022) demonstrate that the polished surface fluency of neural outputs anchors human reviewers, causing them to miss critical system-bound legal errors, omitted qualifications, and false equivalences between civil law and common law terminology.

7. Protecting Both Client and Counsel Through Specialist Lawyer-Linguist Expertise

The lesson of the past decade from KIP UK Ltd and Essential Telecom in 2016 to Hellenic Supreme Court Judgment No. 221/2026, AA v SSHD, and VA v JL in 2025 and 2026 is unmistakable: legal translation shapes the operative force of contracts, the admissibility of evidence, and the professional standing of legal advisers. Treating it as a zero-consequence mechanical task to be delegated to a probabilistic chatbot exposes both client and counsel to avoidable legal hazard.

Founded in 1997 in London and led by lawyer-linguist John O’Shea (LL.B. (Hons), LL.M.), JurTrans has spent nearly three decades providing specialist Greek-to-English and English-to-Greek legal and financial translation for international law firms, corporate legal departments, financial institutions, and private clients. Our practice is grounded in formal legal training, nearly thirty years of hands-on transactional and litigation experience across Greece, Cyprus, and the United Kingdom, and authoritative legal scholarship, including co-translating Greece’s foundational statutory codes (The Greek Penal Code, The Greek Code of Civil Procedure, The Greek Code of Criminal Procedure, and the Greek Code of Private Maritime Law) within the Lex Graeca series.

When you instruct JurTrans on a commercial contract, corporate finance transaction, regulatory filing, or cross-border litigation bundle, we protect both client and instructing counsel by ensuring:

  • From-Scratch Specialist Translation: Every document is translated from scratch by a qualified specialist without reliance on public machine translation or generative AI, eliminating the risk of algorithmic hallucinations and common law terminology bias.

  • File-Level Terminological Consistency: In multi-year commercial relationships and contentious proceedings, we cross-check new notices, pleadings, and exhibits against the terminology already established in your underlying agreements and court file.

  • Source-Text Detective Work: During 2024, across 946 Greek legal documents translated into English at JurTrans, we identified source-text errors, ambiguous statutory citations, or internal inconsistencies in approximately 95% of the originals. Rather than silently guessing or flattening an ambiguity, we preserve the integrity of the text and flag source-text defects directly for instructing counsel.

  • Strict Confidentiality and Accountability: All documents are handled under strict professional confidentiality and formal Non-Disclosure Agreements, backed by professional accountability that no AI vendor provides.

To discuss a Greek or Cypriot legal or financial translation requirement in complete confidence, contact John O’Shea at JurTrans via jurtrans.com or email info@jurtrans.com.

The words “It does work! This is feasible!” appeared in a New Zealand legal disciplinary decision concerning a lawyer accused of misleading her client about litigation strategy. The Legal Complaints Review Officer referred to them when examining the advice the lawyer had given. He ultimately found that the client had been misled and made findings of unsatisfactory conduct against the lawyer.

There is something particularly interesting about those words, though. They were not originally written in English. They appeared in a translation of chat messages that the client said had been produced using ChatGPT. The Review Officer knew this. He expressly recorded it in his decision. But how did he establish that the English accurately reflected what the lawyer had written in the original language?

That is the question I find most interesting about VA v JL [2025] NZLCRO 123.

A dispute about what the lawyer said

The case arose from a complaint about the handling of an employment and commercial dispute. The client maintained that her lawyer had advised her that High Court proceedings could be used to stop or disrupt proceedings before the Employment Relations Authority. An important part of the evidence consisted of chat messages exchanged between the client and the lawyer on 23 July 2021. Both parties produced records of the exchange, but the records differed. The client’s version contained messages in a language anonymised in the decision as Language F, accompanied by English translations she said ChatGPT had produced. The lawyer supplied a shorter version, with what the Review Officer believed were her own translations. Crucially, the client’s version included substantive advice that did not appear in the lawyer’s record. The lawyer denied sending the relevant message.

Which record was more complete?

At paragraph [112], the Review Officer explained why he preferred the client’s record. He considered her records otherwise meticulous, thought deletion of messages more likely than fabrication, and noted that some messages in the lawyer’s version appeared to be replies to missing communications. On the balance of probabilities, he found the client’s record more complete. But that finding did not establish that its English translation was accurate.

Two different evidential questions

For me, this is the central problem with the way the translated evidence was handled. The Review Officer had two questions to consider. The first was which party had supplied the more complete account of the correspondence. The second was what the original-language messages actually meant. Those are not the same question. It is entirely possible for a chat record to be authentic and complete, yet for its English translation to be inaccurate. A translation might omit a qualification, introduce a degree of certainty absent from the original, or convey the general sense while altering an important distinction.

In this case, the Review Officer offered reasons for preferring one record over the other. But the published decision does not explain how he established that the ChatGPT-generated English accurately represented the underlying messages. There is no identified independent translation, no discussion of linguistic verification and no indication that the Review Officer was himself able to understand Language F. Perhaps the translation had been checked. Perhaps there was some other basis for treating it as reliable. The decision does not tell us.

There is also a more subtle difficulty with the reasoning in paragraph [112]. One reason the Review Officer gave for preferring the client’s record was that some of the messages were not what the client would have wanted to hear. But how could he assess what the messages would have meant to the client without first establishing what they actually said? That assessment appears to depend, at least in part, on the meaning conveyed by the English translation. In other words, the translated content was being used to help assess the reliability of the record, without an explanation of how the translation itself had been verified.

I find that an uncomfortable evidential gap.

Why the precise wording mattered

This was not simply a case where an AI translation was included among the documents and then largely ignored. At paragraph [208], the Review Officer expressly returned to the July correspondence. He attributed to the lawyer the statements “Yes, it will work” and “It does work! This is feasible!” He considered those statements when assessing whether the lawyer had allowed her client to believe that commencing High Court proceedings would achieve a particular procedural result. At paragraph [209], he included the July exchange among the evidence concerning the advice the client had received.

The precise meaning of those words matters.

There is a difference between telling a client that a particular course of action is legally possible and assuring the client that it will achieve the desired outcome. There may also be important distinctions between expressing confidence, identifying a possibility and giving a qualified opinion about litigation strategy. Those distinctions can be conveyed differently in different languages. They can also be altered in translation. I am not suggesting that ChatGPT necessarily mistranslated the lawyer’s message. We have no basis for concluding that it did. But equally, the English wording cannot establish its own accuracy. The Review Officer did acknowledge that the disputed message included a warning about the costs and stress of pursuing proceedings. He did not treat it as an unqualified recommendation to litigate. Nevertheless, he relied on its translated content when assessing the advice given.

If the meaning of a disputed communication contributes to a finding about a lawyer’s professional conduct, surely the reliability of the translation deserves some attention.

The Review Officer recognised translation difficulties elsewhere

There is another aspect of the decision I find revealing. At paragraph [106], the Review Officer expressly acknowledged that he did not have English translations of certain earlier messages from the client. He therefore discussed the lawyer’s replies, which were in English. At paragraph [114], dealing with a separate instruction, he described an expression as an inexact translation and interpreted the client’s instructions in their wider context. So translation difficulties were not entirely absent from his reasoning. Yet when it came to the disputed ChatGPT-translated exchange, there is no corresponding explanation of how the accuracy of the English wording was established. The distinction is important. Recognising that a translation exists, or recognising that another translation may be inexact, is not the same as verifying the meaning of the original-language evidence.

Does this call the disciplinary outcome into question? There is an important qualification to this criticism. The Review Officer did not rely exclusively on the ChatGPT-translated messages. His findings were based on a substantial body of evidence, including written legal advice, the conduct of the proceedings and a later translated conversation supplied by the lawyer herself. It would therefore be wrong to suggest that the entire disciplinary outcome depended on the accuracy of one ChatGPT translation. We cannot know whether an independently verified translation would have affected any of the findings. Nor does the decision establish that the ChatGPT translation was inaccurate.

My concern is narrower, but I think it is important. The Review Officer expressly relied on statements appearing in a translation when deciding what advice a lawyer had given. Yet his published reasons do not explain how the accuracy of that translation was established. That is a problem with the transparency of the evidential reasoning, irrespective of whether the ultimate findings were justified by other evidence.

Who checks the evidence when the decision-maker cannot read the original?

In my recent discussion of AA v Secretary of State for the Home Department, I examined a case in which a judge noticed an obvious indication of ChatGPT use in a witness letter that the solicitors had apparently overlooked. That case raised questions about lawyers’ responsibility for checking AI-assisted evidence before submitting it to a court. VA v JL raises a different question. Here, the decision-maker was expressly aware that ChatGPT had been used to translate the evidence. The problem was not detecting AI involvement. It was establishing the reliability of the translated material before relying on its meaning.

Of course, courts and other legal decision-makers routinely deal with evidence written in languages they do not understand. They must be able to rely on translations. I am not suggesting that every translated document requires an independent expert assessment, or that the mere involvement of ChatGPT automatically renders a translation unreliable. But where the content of a communication is disputed, the parties offer different records, and the meaning attributed to particular words contributes to findings of professional misconduct, the method by which that meaning is established deserves scrutiny.

The mere fact that a translation reads fluently in English does not establish that it accurately communicates the source. Nor does preferring one party’s documentary record establish the accuracy of that party’s translation. In AA, the question was why the lawyers failed to notice what the judge noticed. In VA v JL, the question is what the Review Officer did with the knowledge that he was relying on an AI-generated translation.

If a decision-maker relies on translated words to determine what a lawyer said, shouldn’t the decision explain how the meaning of those words was verified? That seems to me a reasonable question to ask when translated evidence becomes part of the basis for a finding of professional misconduct.

 

#LegalTranslation #AIandLaw #Evidence #LegalTech #ProfessionalResponsibility

The judge spotted ChatGPT. Why didn’t the lawyers?

Imagine receiving a witness letter in support of your client’s case. It appears to confirm facts on which you intend to rely. The English reads reasonably well. The document is added to the evidence bundle and filed with the court. There is just one problem. The last sentence offers to turn the letter into a formal affidavit, add notarisation language or make the wording more neutral. That is not what you would normally expect to find at the end of a witness’s account of events. It reads remarkably like an invitation from a chatbot to continue the conversation. Would you have noticed it?

In AA v Secretary of State for the Home Department [2026] UKFTT PA747122024 (IAC), the judge did. The solicitors who filed the letter apparently had not.

The case raises questions about the use of AI translation in legal proceedings. But for me, its most interesting aspect is not a demonstrated failure of machine translation. It is the responsibility of lawyers for checking the provenance, preparation and reliability of the documents they place before a court. And there is an interesting tension in the judgment between the judge’s criticism of the solicitors and the way the Tribunal dealt with the use of AI translation itself.

A witness letter with an unexpected ending

The case concerned an Egyptian national appealing against the refusal of his asylum claim. He maintained that he faced persecution because of his political activities and claimed association with the Muslim Brotherhood. Among the evidence was a letter from a witness identified as YR, who said he had known the appellant in Egypt and had participated in political activities with him. At the hearing, First-tier Tribunal Judge CJ Williams asked how the witness’s letter had been prepared. The witness initially explained that he had written it with someone’s assistance and that his daughter had translated it into English. But the judge had noticed something peculiar at the end of the letter: a sentence offering to adapt the document into a formal affidavit, insert notarisation language or use more neutral wording. That prompted the judge to ask whether ChatGPT or another AI system had been used. YR explained that he had written the letter in Arabic, used ChatGPT to translate it into English and then asked his daughter to compare the English version with the Arabic original to check that the meaning was the same.

There is a small but interesting distinction between those accounts. In the initial explanation, the daughter had translated the letter. In the subsequent explanation, ChatGPT had translated it and the daughter had checked it. One account might simply be shorthand for the other. The judgment does not suggest that the distinction was itself decisive. Nevertheless, for lawyers assessing the provenance of evidence, the difference between translating a document and checking someone else’s translation is not insignificant.

The judge’s criticism

Paragraph 28 contains the observation that, in my view, lawyers should concentrate on: “It is unfortunate that it fell to the Tribunal to notice this”. The judge was referring to the discovery of the apparent AI-generated text in the letter, something the appellant’s solicitors had apparently failed to identify before submitting it. The judge also criticised the absence of a statement of truth from a witness who was intended to give oral evidence. These are separate issues. One concerns the preparation and checking of a document. The other concerns the form in which witness evidence was presented. Neither requires a finding that the underlying Arabic was mistranslated. That is precisely why the episode is worth examining. Lawyers routinely receive documents prepared by clients, witnesses, foreign counsel, administrative staff and other third parties. They cannot

personally create or verify every piece of information those documents contain. But when does reasonable reliance on material prepared by others become an inadequate check on evidence submitted in their professional capacity? And how far should lawyers be expected to investigate the use of AI in producing that material?

Two very different kinds of checking

For me, AA exposes a distinction that is frequently overlooked when lawyers discuss AI translation. The first kind of checking concerns the document itself. Does it contain extraneous material? Is it complete? Does it appear to be the witness’s evidence rather than a draft, template or chatbot response? Has the final version been read before filing? Does it satisfy the relevant procedural requirements? In this case, the telltale sentence was in English. Detecting it required no knowledge of Arabic and no particular expertise in translation. That is what makes the judge’s criticism difficult to dismiss. The second kind of checking is much harder. Does the English version accurately and completely communicate what the witness wrote in Arabic? Has anything been omitted, added, softened or made more emphatic? Have factual assertions retained their meaning? Has the translation introduced language that changes the character or certainty of the evidence? Answering those questions requires more than reading the English fluently. It requires an assessment against the original by someone competent to perform that task. A lawyer who does not read Arabic may be able to identify a chatbot’s leftover instruction. The same lawyer may be entirely unable to identify a subtle mistranslation that materially changes the witness’s account.

The fact that one check has been performed tells us very little about whether the other has. The problem with ‘human involvement’

There is another aspect of this case I find revealing. The witness said his daughter had compared the Arabic and English versions to ensure that the meaning was the same. That may have been a useful check. There is no finding that she was unable to perform it or that the translation was inaccurate. But should the mere fact that a human has reviewed an AI-generated translation be sufficient to reassure a lawyer relying on it? I would suggest not, without knowing more about the nature of the review. Who carried it out? What linguistic competence did that person possess? Did they compare the complete texts? Were they checking for accuracy, omissions and alterations of meaning, or simply whether the English appeared to reflect the general sense of the Arabic?

Those are quite different exercises.

In a legal setting, the distinction matters because the translation is not merely facilitating informal communication. It is being used as evidence. The judgment does not determine whether the daughter’s review was adequate. Nor does it establish what further verification, if any, the solicitors undertook. It does, however, provide a useful example of why the presence of a human somewhere in an AI-assisted process should not automatically be equated with meaningful quality assurance.

Judicial tolerance is not judicial approval

There is an important qualification to all this.

Paragraph 29 records that the Home Office presenting officer did not argue that the use of AI for translation, with some human involvement, warranted reducing the weight given to YR’s evidence on that particular ground. The judge ultimately found that the evidence of the two supporting witnesses did not particularly strengthen the appellant’s claim. The appeal was dismissed following an assessment of credibility and the evidence as a whole. But the Tribunal did not find that ChatGPT

had produced an inaccurate translation. Nor did it establish a general rule that ChatGPT translations checked by a family member are acceptable evidence. That distinction is important. It would be easy to present AA as another case where AI translation went wrong and a litigant suffered the consequences. The judgment does not support that conclusion. Equally, it would be misleading to interpret the absence of a translation-related penalty as judicial endorsement of AI-assisted translation. What the decision demonstrates is narrower, but arguably more interesting: a court can be critical of the way AI-assisted evidence was prepared and submitted without treating AI involvement as an automatic reason for discounting that evidence.

A different warning from the cases already emerging

In my earlier discussion of Lozano Gonzάlez v Roberge, a Canadian tribunal confronted a ChatGPT translation of legislation that had effectively rewritten the statutory provisions on which a litigant relied. In Euroboor BV v Grafova, an American court struggled to understand the reasoning in a Dutch judgment supplied in a Google Translate version. Those cases raised questions about what happens when translated legal material fails to communicate the meaning on which a legal argument depends.

AA raises a related but different problem. Here, the visible failing was not a demonstrated mistranslation. It was the failure to identify an obvious indication of chatbot involvement before the document reached the Tribunal. That brings the issue much closer to the everyday work of lawyers managing evidence. It is one thing to ask whether AI can translate accurately enough for use in proceedings. It is another to ask what a legal team must do before relying on a document that someone else has translated using AI. The first question concerns technological performance and linguistic reliability. The second concerns professional judgment, document control and accountability. The answers are connected, but they are not the same.

Who takes responsibility?

Nothing in AA establishes professional misconduct by the solicitors, or a particular sanction for submitting an AI-translated document. Nor should a decision of a UK First-tier Tribunal be treated as laying down a universal standard for lawyers in other jurisdictions. But the wider questions transcend the procedural rules of any single court. Suppose a witness sends a lawyer an English statement translated from another language using ChatGPT. The lawyer cannot read the original language. The witness assures the lawyer that a relative has checked the translation. The English looks plausible. The document is then filed. What exactly is the lawyer relying on? The witness’s assurance? The relative’s linguistic ability? The apparent fluency of the English? The reputation of the AI system? And what evidence is there of the verification process if the translation is subsequently challenged?

I do not think the answer is that every lawyer must personally translate every foreign-language document or that AI can never form part of a translation workflow. The question is whether the lawyer has sufficient grounds for treating the version filed as a reliable representation of the underlying evidence. That is a question about the process as well as the product.

A warning worth taking seriously

The AI-generated sentence in AA was conspicuous enough for the judge to notice. Its presence was embarrassing, but it also served a useful purpose: it made the use of ChatGPT visible and prompted questions about the document’s preparation. Many machine-translated documents contain no such obvious clue. They look like perfectly ordinary English documents. Some may be entirely accurate. Others may not be. The difficulty is that the absence of a conspicuous error tells us nothing reliable

about the accuracy of the underlying translation. For lawyers, that is the uncomfortable point. The judge’s observation in AA should not be read as a finding that AI translation is inherently unreliable It should be read as a reminder that someone must take responsibility for what reaches the court.

If the legal team misses an obvious trace of ChatGPT that can be detected simply by reading the English, what basis does it have for confidence in the parts of the document that can only be checked against the original language?

That is a question worth asking before the next witness letter is filed.

 

#Litigation #ProfessionalResponsibility #LegalTranslation #LegalTech #AIandLaw #Evidence #LegalRisk #LawyersOfLinkedin

When a Greek court delivers a judgment, the parties involved naturally want to know what the court has decided and what happens next. For international clients who do not speak Greek, obtaining a clear and accurate English translation is often essential to understanding the outcome of the proceedings and discussing the implications with their Greek lawyers. The same applies to lawyers practising outside Greece who advise clients involved in Greek litigation. They may need to examine the judgment to understand the court’s reasoning, identify the issues decided and discuss the available options with the Greek legal team.

However, translating a Greek court judgment presents some distinctive challenges. Greek judicial decisions frequently contain lengthy and complex legal reasoning expressed in highly formal language. Procedural concepts may have no straightforward English equivalents, while apparently familiar terms can conceal important differences between legal systems. These problems become particularly significant when the judgment may be challenged on appeal or when the parties need to understand its consequences.

And with artificial intelligence now making it possible to generate an English translation or summary of a lengthy judgment in seconds, there is a further question: how much confidence can international clients and their lawyers place in the resulting text?

Understanding the structure of a Greek court judgment

Greek court judgments follow established drafting conventions which reflect the structure of Greek judicial reasoning and the procedural rules governing the proceedings. Depending on the court and type of proceedings, the judgment will ordinarily identify the court and parties, record relevant procedural information and set out the court’s reasoning before stating its decision. Two sections are especially important when translating a Greek judgment. The first is the σκεπτικό, the court’s reasoning, often introduced by the formula ΣΚΕΦΘΗΚΕ ΣΥΜΦΩΝΑ ΜΕ ΤΟ ΝΟΜΟ (having considered the matter in accordance with the law). This part of the judgment explains the legal basis on which the court has reached its conclusions. It may contain lengthy discussions of statutory provisions, references to case law and an examination of the parties’ arguments. The second is the διατακτικό, the operative part, commonly introduced by the expression ΓΙΑ ΤΟΥΣ ΛΟΓΟΥΣ ΑΥΤΟΥΣ (for these reasons). The operative part records the court’s decisions on the matters before it, including the disposition of claims and any orders made.

For a foreign reader, the relationship between these sections is fundamental to understanding what the judgment actually says.

The linguistic complexity of Greek judicial decisions

The language of Greek judgments presents difficulties that extend beyond specialised legal terminology. Greek judicial writing frequently employs long sentences containing multiple subordinate clauses. The syntax can be intricate, with qualifications and exceptions embedded within the same sentence as the principal legal proposition. In addition, some judicial formulations retain features of the older formal language associated with καθαρεύουσα, particularly in established expressions and conventional legal phraseology.

For translators working from Greek into English, these characteristics demand close attention to the grammatical and logical relationships within the text. The translator must establish precisely which proposition is being qualified, what a particular pronoun or participle refers to and whether a statement reflects the court’s conclusion or the position of one of the parties. These distinctions can become difficult to follow when a sentence extends over much of a paragraph. English legal writing generally permits a different approach to sentence construction. A lengthy Greek sentence may therefore need to be divided or restructured in English to make the reasoning intelligible. However, restructuring must preserve the original logical relationships. Careless division of a sentence can detach a qualification from the proposition to which it relates or alter the relationship between the court’s findings and its conclusions. This is where the ability to think and write like a lawyer becomes particularly valuable. Understanding the legal argument allows the translator to reconstruct it in clear English while preserving its precise meaning. The objective is an English text that a lawyer can read naturally and follow without losing any of the legal substance contained in the Greek original.

When is a Greek court judgment final?

One of the most revealing examples of the difficulties involved in translating Greek judgments concerns the terminology used to describe their procedural status. Consider the expression τελεσίδικη απόφαση. A reader encountering the English expression “final judgment” might reasonably assume that the judgment can no longer be challenged. Greek procedural law, however, recognises distinctions that make this apparently straightforward translation potentially misleading. In our paper Translating Greece’s Procedural Codes, presented at the 2023 conference of the Hellenic Society for Terminology, Daniel Webber and I examined the treatment of procedural terminology in existing Greek-English legal dictionaries. We found that all four dictionaries examined rendered τελεσίδικη απόφαση using variants of “final judgment” or “final decision”, without adequately reflecting the procedural distinctions recognised by Greek law. An οριστική απόφαση may be final in the sense that it disposes of the matter before the court, while remaining open to ordinary legal remedies. A τελεσίδικη απόφαση has reached a different procedural position, generally because the relevant ordinary legal remedies have been exhausted or the periods for exercising them have expired. Greek procedural law also recognises the category of ανέκκλητη απόφαση, concerning judgments which by their nature are not amenable to the relevant remedies, as well as αμετάκλητη απόφαση, which involves the exhaustion of available legal remedies, including the relevant avenues of recourse to higher courts.

For international clients, these distinctions can be important to understanding the procedural history of their case. They also illustrate the difficulties faced by foreign lawyers who may be familiar with the English word “final” but unfamiliar with the different categories recognised under Greek law. The translation needs to communicate the source terminology accurately without creating a misleading impression about the status of the decision. Determining whether a particular judgment can still be challenged, or what consequences follow from its procedural status, is a matter for the lawyers responsible for the proceedings.

Errors in the original Greek judgment: an overlooked translation problem

There is another difficulty which frequently arises in legal translation and which is easily overlooked by the end user: the original Greek document may itself contain errors. At JurTrans, our experience of translating Greek legal documents has made us particularly conscious of this problem. In an article published in January 2025, I reported that during 2024 we translated 946 Greek legal documents into English and found that approximately 95% contained at least one error of some kind. These figures relate to JurTrans’s own workload, which included various categories of legal documents. They should not be interpreted as a representative statistical study of the quality of Greek court judgments. Nevertheless, they illustrate how frequently source-text problems arise in our translation work. Some errors are relatively straightforward, while others require considerably more research to identify and understand. A judgment may refer to an incorrect article of legislation or contain an ambiguous statutory citation. There may also be inconsistencies in how a particular legal term is used. The translator must determine whether the apparent inconsistency has a reasonable explanation or whether there is a genuine problem in the Greek text.

In The legal translator as detective, I described an example from our translation work involving a reference to an article of the Greek Code of Civil Procedure. The provision cited in the document appeared to have no connection with the procedural steps being described. After checking the surrounding provisions and searching the Code, I eventually discovered that the reference related to an earlier version of the legislation. The source document had failed to make that historical context clear. Having identified the issue, I suggested that the instructing lawyer clarify the Greek wording. He agreed. The incident is a good illustration of the research that may be necessary to understand a legal text correctly. Given the successive amendments made to Greek legislation, consulting the current version of a statutory provision may sometimes be insufficient to establish the meaning of a historical reference.

Where a judgment contains an apparent error, the translator must preserve the integrity of the source text while drawing the issue to the attention of the instructing lawyer where appropriate. A careful translation process can therefore reveal problems which might otherwise remain unnoticed, particularly by clients who cannot consult the Greek original.

Why relying on AI-generated summaries of Greek judgments is risky

The ability to obtain an instant summary of a lengthy court judgment is understandably attractive. International clients may want an immediate answer to whether the proceedings have been successful, while foreign lawyers may need to understand the broad significance of a decision before speaking to their Greek counterparts. Generative AI systems appear to offer a convenient solution. A document can be uploaded and a short English summary requested in a matter of seconds. The difficulty lies in the reliability of that summary. Court judgments often depend on distinctions which can be lost when a lengthy text is compressed into a short narrative. A statutory exception may substantially qualify a legal proposition. The court may accept part of a party’s argument while rejecting another part. An apparently straightforward conclusion may depend on a particular interpretation of the facts or the application of a specific procedural rule. If a summary overlooks one of these qualifications, the reader may be left with an inaccurate understanding of the decision. There is also the risk that an automated system will fail to distinguish correctly between the parties’ submissions and the findings of the court. Greek judgments may reproduce the arguments advanced by litigants before explaining whether those arguments have been accepted. An AI-generated summary which attributes a rejected submission to the court would fundamentally misrepresent that part of the judgment.

These concerns are consistent with findings in the growing academic literature on automatic legal-document summarisation. In a 2024 study of large language models and generative systems for summarising court judgments, Aniket Deroy, Kripabandhu Ghosh and Saptarshi Ghosh examined systems using judgments from the supreme courts of the United Kingdom and India. Although generative systems performed relatively well according to conventional automatic evaluation metrics, closer examination identified inconsistencies in their summaries. The researchers concluded that further improvements were needed before such systems could be considered sufficiently reliable without detailed human checking. More recently, a 2026 review of legal-document summarisation research highlighted the difficulty of preserving the argumentative structure of judgments and the limitations of evaluation methods which primarily measure similarities in wording rather than legal accuracy.

Neither study, though, specifically evaluates Greek court judgments, but both identify problems relevant to the task of summarising complex judicial reasoning. The challenge is compounded when a Greek judgment is summarised directly into English. The system is then required to interpret the Greek legal text and produce a compressed account in another language, potentially concealing errors from readers who cannot independently verify the original.

For lawyers considering the implications of a judgment, a summary cannot provide the level of detail available in an accurate translation of the full text. The actual wording of the court’s reasoning may be central to understanding the decision and discussing possible further steps with Greek counsel.

What does research tell us about AI translation from Greek into English?

The limitations of automated translation in the Greek-English legal context have also been the subject of empirical research. In September 2026, Vilelmini Sosoni and I published a chapter in Springer’s International Handbook of Legal Language and Communication entitled Translating Normative Legal Texts: A Comparative Evaluation of Neural Machine Translation and Generative Artificial Intelligence Models in the Greek-English Language Pair. The study compared a neural machine translation system trained on legal-domain material with three generative AI models. We evaluated translations of two Greek property law texts, comprising planning legislation and a property sale agreement, using automatic translation-quality metrics alongside detailed human error analysis. Our findings were revealing. Accuracy errors predominated across the systems examined, and most were classified as major in severity. Although the generated English generally displayed adequate fluency, the generative AI models did not systematically outperform the specialist neural machine translation system.

These findings concern normative legal texts rather than court judgments. Their significance for the present discussion lies in the difficulties the systems encountered when translating specialised Greek legal concepts into English. The results reinforce the need to distinguish between an English text that reads fluently and one that faithfully communicates the legal meaning of its Greek source. A system may produce entirely natural English while selecting an inappropriate equivalent for a Greek procedural concept or failing to preserve an important qualification. Recognising such problems requires detailed knowledge of the source legal system and the ability to evaluate the translation against the original. Technology certainly has useful applications in legal translation, particularly in supporting research and maintaining terminological resources. However, the limitations identified in empirical studies call for considerable caution when using AI or machine translation to communicate the substance of court judgments.

At JurTrans, our approach to documents of this nature is based on establishing the meaning of the Greek text through specialist legal-linguistic analysis and producing an English translation whose accuracy can be carefully verified against the source.

Preparing Greek judgments and appeal documents for international readers

The usefulness of a Greek-English legal translation also depends on understanding the circumstances in which it will be read. For a Greek lawyer representing an international client, the primary objective may be to enable that client to understand the court’s decision and participate meaningfully in discussions about the case. Lawyers practising outside Greece may require more detailed access to the reasoning, particularly where the litigation has implications for their client’s commercial activities or wider legal affairs. In either situation, the translator benefits from understanding the purpose of the assignment. Where a judgment refers extensively to previous decisions or procedural documents, access to those materials may assist with the interpretation of terminology and the maintenance of consistency across the translations. The English text should also enable the reader to locate the corresponding material in the Greek original. This requires attention to the document’s structure, the identification of the parties and the reproduction of references to legislation and earlier decisions. Where only selected passages are translated, the extent of the translation should be clearly identified so that the reader is aware of the material covered. However, where the purpose is to understand the judgment as a whole, a complete translation is generally more useful than an abbreviated account which may omit context essential to the court’s reasoning. The appropriate scope should be agreed with the instructing lawyer, who remains responsible for determining how the translated document will be used in connection with the proceedings.

The value of specialist Greek-English legal translation expertise

At JurTrans, we have specialised in Greek-English legal and financial translation for three decades. Our work involves examining legal documents within their procedural and substantive context, with particular attention to terminology whose meaning depends on the structure of the Greek legal system. That experience is supported by formal legal education, university-level teaching experience and sustained involvement in academic research into legal translation.

Through our work on the Lex Graeca project, including the translation of major Greek codes (both substantive and procedural) into English, we have also developed extensive experience in researching Greek legal terminology and establishing consistent English renderings of concepts which are often inadequately covered by existing reference works.

We have translated many millions of words over those decades and have countless satisfied clients.  These figures reflect the depth of experience accumulated across a broad range of legal assignments.

Our approach to translating Greek court judgments is informed by detailed familiarity with the intricacies of Greek law and an appreciation of how the English translation will be understood by lawyers accustomed to different legal systems.

We approach each text with the analytical mindset of a lawyer, examining the reasoning and procedural context before settling on the most appropriate English terminology. Equally important is close attention to detail. Statutory references are checked where necessary, potential inconsistencies are investigated and questions arising from the Greek original are raised with the instructing lawyer when clarification is required. This is supported by specialist legal reference resources, access to current legal databases and a commitment to the confidentiality of clients’ documents.

For international clients and their legal advisers, the objective is to provide an English translation which communicates the substance of the Greek judgment faithfully and allows the legal issues to be discussed on an informed basis.

 

Frequently asked questions about translating Greek court judgments

Does a Greek court judgment need to be translated in full?

The appropriate scope depends on why the translation is required. Where an international client or foreign lawyer needs to understand the complete reasoning and outcome of the proceedings, a full translation is generally the most useful option. Translations of selected passages may also be appropriate, provided their scope is clearly identified and agreed with the instructing lawyer.

 

Can AI summarise a Greek judgment accurately?

AI-generated summaries may provide preliminary orientation, but their reliability should not be assumed. Research into automated judgment summarisation has identified factual inconsistencies, while studies of Greek-English legal translation have found substantial accuracy errors in automated outputs. Where a judgment’s meaning is important, the original text and an accurate professional translation provide a more dependable basis for legal discussion.

What should be provided when commissioning a Greek court judgment translation?

Providing the complete Greek document and, where relevant, earlier decisions or existing translations can help the translator understand the procedural context and maintain consistent terminology. It is also useful to explain who will read the translation and whether particular cross-references or presentation requirements need to be preserved.

Greek court judgment and appeal translation services at JurTrans

Greek court judgments and appeals can be demanding documents to translate. Their legal meaning depends on the procedural context, the structure of judicial reasoning and the specialised terminology of Greek law.

For international clients and the lawyers advising them, an accurate English translation provides the basis for understanding what the Greek court has decided and discussing its significance with the Greek legal team.

JurTrans provides specialist Greek-English translation of court judgments, appeal documents and related litigation materials, drawing on decades of practical experience and academic research into legal language.

If you require an English translation of a Greek court judgment or documents relating to an appeal, please contact JurTrans or email info@jurtrans.com.

JurTrans is a specialist legal translation company and does not provide legal advice. This article addresses legal translation and communication issues for general information. Questions concerning legal rights, procedural remedies and the consequences of court decisions should be addressed to the lawyers responsible for the proceedings.

 

 

“This paper has shown that GenAI is not fully capable of identifying terms that may be ambiguous for readers; consequently, its ability to define or explain such elements is limited. Likewise, its aptitude to ‘choose familiar, legally accurate words’ is also compromised, especially in view of the training-based nature of its responses.

Indeed, the six chatbots under scrutiny did not show the ability to interpret legal terms in context and legal subtleties accurately. Users need certainty and clarity when addressing legal concepts.

Therefore, chatbots are currently unsuitable for the easification or disambiguation of legal communication and making it understandable and usable for its intended audience.

Although AI may seem to provide sensible responses, inexactness and approximation come to the fore when analysing outputs in detail. This means that AI-generated responses are not reliable and consistent, as they can differ from one query to another.”

Giampieri, P., “Can GenAI identify and disambiguate legal polysemous terminology? a critical analysis in light of ISO 24495-2:2025 on legal communication” (2026) Iura & Legal Systems

 

#LegalTranslation #LegalConcepts #GenAI #Accuracy #LegalLanguage

“In legal contexts, translation does not merely convey information; it can shape the operative force of a document. If a machine translation changes the scope of a clause, weakens a condition, or substitutes a general term for a legally defined one, the translated text may carry a different normative effect from the source text.

This problem is especially serious in legally binding texts such as contracts, regulations, and judgments, where the translated version may be relied upon for legal interpretation or enforcement.

Machine learning systems are not inherently aware of these consequences, because they optimise for probabilistic plausibility rather than legal equivalence. accuracy is inseparable from the goal of achieving an equivalent, uniform legal effect. Unlike general or literary translation, legal translation operates within tightly constrained systems of meaning shaped by jurisdiction-specific norms, institutional practices, and codified terminology.

A mistranslation in this domain is not merely a semantic error; it can alter rights, obligations, and legal outcomes, potentially leading to financial loss, contractual disputes, or miscarriages of justice.”

 

#LawyersOfLinkedin #AI #MT #AlteredMeaning

 

Pinčić, Domagoj & Bajcic, Martina. (2026). Challenges of the Legal Translation Task in the Context of Machine Learning. 10.5281/zenodo.22765851.

“Legal translation is widely recognised as a high-stakes, domain-specific activity in which linguistic accuracy is inseparable from the goal of achieving an equivalent, uniform legal effect. Unlike general or literary translation, legal translation operates within tightly constrained systems of meaning shaped by jurisdiction-specific norms, institutional practices, and codified terminology. A mistranslation in this domain is not merely a semantic error; it can alter rights, obligations, and legal outcomes, potentially leading to financial loss, contractual disputes, or miscarriages of justice.”

Pinčić, Domagoj & Bajcic, Martina. (2026). Challenges of the Legal Translation Task in the Context of Machine Learning. 10.5281/zenodo.22765851.

#lawyersOfLinkedin #rights #AI #MT


Contact


If you would like further information about the specialised services we provide, or need a quote for a translation please feel free to contact us.

Name
Email
Message

Thank you for your message! We will get back to you as soon as possible.
There has been an eror in the form. Please validate the required fields.
© Copyright 2014 JurTrans