You think you’re saving money. You think you’ve received a proper translation. The English looks excellent. And you may already be heading towards a legal or financial problem you cannot see.

This is the real danger of AI legal translation.

It is fast. It is cheap, sometimes free. It produces polished, authoritative-looking English in seconds. And if you cannot read the original document, you may have absolutely no reason to suspect anything is wrong.

Until the document reaches a lawyer, a court, a public authority or the other party to a transaction.

AI legal translation can be a false economy because the money saved at the translation stage may be dwarfed by the cost of undetected errors, delays, disputes and corrective work.

At JurTrans, we have spent nearly three decades translating legal documents between Greek and English. We have also spent years researching what happens when legal translations go wrong.

The evidence makes uncomfortable reading.

The problem isn’t that AI translation looks bad. It’s that it looks good.

There was a time when machine translation errors were often obvious. The English was awkward, the grammar unnatural and the meaning sometimes plainly absurd.

Today’s systems are much more convincing.

And that is precisely the problem.

In our 2026 research, Translating Normative Legal Texts: A Comparative Evaluation of Neural Machine Translation and Generative Artificial Intelligence Models in the Greek–English Language Pair, Vilelmini Sosoni and I examined the performance of a specialist neural machine translation system and three generative AI models when translating Greek property law documents into English.

The texts included planning legislation and a property sale agreement.

The findings were revealing.

The English was generally fluent. Yet accuracy errors predominated across the systems tested, and most of those errors were classified as major.

Read that again.

The systems could produce English that read well while making errors that mattered.

The generative AI models did not systematically outperform the specialist machine translation system either.

These findings come from a defined sample of legal texts, not a survey of every AI translation ever produced. They do not establish a universal percentage of AI translations that fail.

But they establish something important: fluency is not a reliable measure of legal accuracy.

If you are a lawyer who cannot read Greek, or a property buyer trying to understand a Greek contract, what exactly are you assessing when you read the English?

You are judging how convincing it sounds.

You are not necessarily judging whether it accurately represents the law.

And the two things are not the same.

Sometimes one word is enough to change the law

Consider a finding from an earlier study I conducted with Vilelmini Sosoni and Maria Stasimioti.

In Translating Law: A Comparison of Human and Post-Edited Translations from Greek to English, published in 2022, we examined the translation of Greek property legislation and contractual material.

One machine-translated passage contained the words “shall not apply”.

The relevant Greek provision actually stated that the provisions did apply.

That is not a stylistic preference. It reverses the legal effect of the sentence.

Another example involved the Greek expression τίτλος κτήσης. The machine translation offered “deed of conveyance”. A human post-editor recognised a problem but changed it to “property title”, which was also incorrect in context. The appropriate rendering was “title deed”.

This study concerned neural machine translation rather than today’s generative AI systems, but it illustrates the underlying difficulty: an English phrase can be perfectly comprehensible and still be legally wrong.

It also challenges a common proposed solution: let AI do the translation, then ask someone to check it.

Who is doing the checking? Do they understand the relevant legal systems? Can they identify an incorrect legal concept rather than simply improve an awkward sentence?

Our small 2022 study found no productivity gain from post-editing in the tasks examined. Some machine translation errors remained uncorrected, while other errors were introduced during editing.

Human translators are not infallible either. That is why specialist competence, appropriate research and meaningful quality assurance matter.

Putting a human somewhere in the process is not the same as having a qualified specialist take responsibility for the translation.

Two serious errors. €18,890 in reported costs.

Long before generative AI became widely available, I presented research at a conference in Seville examining what poor legal translation actually costs clients.

The paper, Quality Metrics in Legal Translation: Taking a Client-Side Perspective, examined cases in which translations had to be corrected after problems emerged in real transactions.

One case involved a Russian purchaser buying a property on a Greek island for €324,000.

The original translation comprised approximately 23,000 words.

Only two serious errors were identified.

Yet the client reported €18,890 in planning-law fines, alongside delays affecting the transaction.

Think about that.

Two serious errors in a document of 23,000 words might suggest an overwhelmingly accurate translation if you looked only at the number of mistakes.

But error counts tell you very little about what those mistakes might cost.

The same research examined a prospective purchaser of a €750,000 property in Athens whose documentation contained extensive translation problems. The client reported €22,700 in planning-related fines.

In another case, an investor’s translated documents were rejected by a permitting authority, and the client reported €4,500 in associated losses and costs.

These were client-reported cases involving earlier translation failures, not generative AI. They cannot be used to claim that AI caused those particular losses.

What they demonstrate is something more fundamental.

The financial consequences of a legal translation error are not necessarily proportional to the number of errors in the document.

One mistranslated legal concept can matter more than thousands of correctly translated words.

That is why judging translation quality by how good the English looks is so dangerous.

The cost often emerges after the translation has left the translator’s desk

This was also a central concern in my subsequent research with Juliette Scott.

In our 2021 paper, How Legal Documents Translated Outside Institutions Affect Lives, Businesses and the Economy, we examined what happens to translated legal documents after delivery.

This is a stage of the translation process that receives far too little attention.

A translated contract is sent to a prospective investor. A property document reaches a conveyancing lawyer. A court judgment is used to assess the status of proceedings. A patent translation becomes part of an application for legal protection.

And then the translation takes on a life of its own.

People make decisions based on it.

Sometimes those decisions involve very substantial sums of money.

Our research explored how translated legal documents can subsequently be challenged, contested or discredited, with repercussions for businesses, legal rights, financial interests and legal certainty.

The problem is no longer confined to the words on the page.

It becomes a problem for the people relying on them.

Take patent translation.

One case discussed in the literature concerned a Russian patent in which “median particle diameter” was mistranslated as “average particle diameter”. The patent owner later attempted to correct the translation, but the correction was refused and the patent was invalidated.

Two apparently similar words. A very different legal result.

Consider court judgments, too.

Greek law distinguishes between different stages of a judgment’s finality. A decision that is final in one procedural sense is not necessarily beyond all forms of challenge.

An English translation that obscures the distinction between a final and an irrevocable judgment may leave a foreign reader with the wrong understanding of the judgment’s procedural status.

Or consider Greek property rights.

Concepts relating to ownership, division of property, planning status and the precise nature of an interest in land do not always have straightforward English equivalents.

A familiar-sounding English legal term can suggest rights or procedures that do not correspond to those established under Greek law.

These are not simply vocabulary problems. They are problems involving legal systems, rights and consequences.

Our work with Juliette showed why the real-world impact of legal translation must be assessed beyond the point of delivery.

The problem may only become visible when the document is challenged — and by then, the damage may already have been done.

AI makes the traditional problem harder to detect

Now combine those earlier findings about real-world consequences with our more recent research into automated translation.

The result is troubling.

We already knew that inaccurate legal translations could delay transactions, complicate litigation, undermine rights and generate substantial financial losses.

We now have empirical evidence that automated systems can produce fluent, convincing translations while making major accuracy errors.

Put those two findings together.

A client uploads a Greek contract to an AI system.

Seconds later, they receive perfectly readable English.

They assume they have an accurate translation because nothing looks wrong.

They forward it to their solicitor, business partner or financial adviser.

Perhaps nobody in that chain can read the original Greek. Perhaps nobody notices that a legal qualification has disappeared, that a property right has been misidentified or that an obligation has been expressed incorrectly.

The document looks professional.

The assumption of accuracy becomes the basis for further decisions.

You haven’t necessarily eliminated the cost of translation. You may simply have transferred that cost into a future legal problem.

And crucially, you may have no way of knowing that you have done so.

The cheapest translation may carry the largest hidden bill

When clients compare translation prices, they naturally look at the immediate cost.

A professional specialist charges a fee. An AI system may appear to charge almost nothing.

On that narrow comparison, AI wins.

But that is not a meaningful comparison if the translation is going to be relied upon for a legal or commercial purpose.

The real cost includes the possibility of a document being rejected, a transaction being delayed, additional lawyers’ fees, corrective translation work, disputes over contractual meaning or rights being misunderstood.

The Seville cases showed how consequential such problems can become.

And there is another uncomfortable question.

If you use AI to translate a document and then commission a specialist to verify it properly, how much have you actually saved?

A proper legal review may require comparing every relevant passage against the source, researching the applicable law, checking terminology and correcting errors.

In some circumstances, producing a reliable translation from the original may be more sensible than attempting to repair an untrustworthy one.

The 2022 study with Sosoni and Stasimioti is a useful reminder that post-editing does not automatically deliver the time savings people assume.

Cheap at the beginning does not necessarily mean economical overall.

What you are paying for when you choose JurTrans

At JurTrans, we specialise in Greek-English legal and financial translation.

We understand that a Greek legal document cannot always be translated simply by selecting an English phrase that appears to mean roughly the same thing.

The task involves identifying legal concepts, understanding the source legal system, researching terminology, recognising ambiguities and considering how the translated document is going to be used.

If a source document contains an apparent error or an ambiguous legal expression, that may need to be investigated or brought to the client’s attention.

If there is no satisfactory one-to-one equivalent between Greek and English legal terminology, that requires professional judgment.

A fluent sentence is not enough.

Nor is an impressive technological interface.

A specialist legal translator is there to identify problems that the client may not even know exist.

That is a substantial part of the value.

At JurTrans, we use appropriate specialist tools and legal research resources to support our work, but we do not use generative AI to perform client translations.

For almost 30 years, our focus has been on delivering translations that accurately communicate the source document and are appropriate for their intended legal or commercial purpose.

That is quite different from asking a machine to produce something that looks like a legal translation.

Is AI legal translation ever worth using?

For an informal, preliminary impression of a document, automated translation may sometimes be useful, provided the user understands its limitations and addresses any confidentiality concerns.

But there is a very substantial difference between obtaining a rough impression of what a document discusses and relying on a translation to make legal or financial decisions.

If the translation is going to be filed with a court, used in a property transaction, submitted to an authority, relied upon in litigation or used to understand binding contractual rights, apparent fluency is not sufficient.

The accuracy of the translation must be established through an appropriate process.

And that process requires genuine expertise.

Frequently asked questions

Why is AI legal translation a false economy?

AI legal translation may reduce the initial cost of producing an English text, but undetected errors can create much greater costs through delays, disputes, rejected documents and corrective work. Research into Greek-English automated legal translation has found major accuracy errors despite generally fluent output.

Can I trust AI to translate a Greek legal document?

AI may convey some passages accurately, but it cannot be assumed to preserve every legally significant distinction. If the document will be relied upon for a legal or commercial purpose, it should be assessed by a suitably qualified specialist.

Why use JurTrans instead of ChatGPT for legal translation?

JurTrans provides specialist Greek-English legal translation supported by legal knowledge, experience and professional research. Unlike a general-purpose AI system, a specialist translator can examine legal context, identify ambiguous terminology and take professional responsibility for the translation delivered.

The most expensive mistake is the one you never noticed

The attraction of AI translation is obvious.

You upload a document, receive an answer immediately and apparently avoid the cost of hiring a professional.

But what exactly have you received?

An English text that reads well?

Or an accurate translation on which you can safely rely?

Our research shows why you cannot simply assume that the first is the second.

And our work examining the real-world consequences of translation errors shows why that distinction matters.

You may believe you have saved money on a legal translation. In reality, you may have acquired a problem whose existence you cannot detect and whose eventual cost you cannot predict.

That is not economy.

It is risk disguised as a saving.

Before trusting an AI-generated legal translation, ask yourself a simple question: can you afford the consequences if it is wrong?

For specialist Greek-English legal translation, visit JurTrans or contact us at info@jurtrans.com

#AITranslation, #LegalTranslationRisks, #JurTrans, #GreekEnglishTranslation, #TranslationAccuracy