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.