Machine Translation vs. Generative AI in Legal

Our team has a long track record of supporting global law firms and corporate legal departments through highly complex, multilingual matters. In almost every recent strategic meeting or case consultation we have attended, a familiar question arises among legal professionals: Can we use Generative AI to translate our foreign-language data sets and reduce our eDiscovery costs?

It is an understandable inquiry. As data volumes expand exponentially, the pressure to find fast, automated solutions for managing cross-border discovery expenses is greater than ever. Clients frequently look at the bottom line and question why they need a bilingual or multilingual attorney with a higher billing rate when they could theoretically plug foreign evidence into an AI tool, translate it to English, and have an English-speaking attorney review the output. However, jumping straight into application without evaluating the underlying technology creates a significant knowledge vacuum. In litigation, treating “AI” and “Machine Translation” as interchangeable terms introduces severe operational and reputational risks. To safeguard your evidence, it is vital to understand the distinct mechanics, strengths, and weaknesses of these two modern pillars of automated translation.

Understanding the Technical Distinctiveness: NMT vs. GenAI

To make an educated buying and strategic decision, legal teams must distinguish between how these technologies process language:

• Neural Machine Translation (NMT): This is the technology behind established enterprise tools like DeepL, Amazon Translate, and Microsoft Translator. NMT uses artificial neural networks to evaluate entire sentences at once, optimizing for predictable, repeatable output and a strict mapping between the source and target texts.
• Generative AI & LLMs: These are large language models like ChatGPT, Claude, or Gemini. Rather than relying on strict source-to-target mapping, they generate translations as part of a broader text generation process driven by contextual reasoning.

While Generative AI produces highly fluid, creative, and polished English text that reads beautifully, it introduces a dangerous paradox for legal discovery: looking polished is not the same as being legally accurate.

Because LLMs are generative by nature, they prioritize natural sentence flow. In doing so, they can easily alter critical legal nuance, shift emphasis, or smooth over technical discrepancies in a way that Neural MT is far less likely to do.

Video Briefing: Behind the Scenes of Automated Translation

Before integrating an automated tool into your next cross-border investigation or review workflow, it is critical to look at how these systems handle real-world data. In the video below, we sit down with Remu Ogaki of Hilgers PLLC, an attorney specializing in multilingual litigation and technology-assisted review (TAR), to break down what is actually happening on the back end.

The “One Answer” Pitfall: The Ultimate Litigation Risk

The single greatest hidden weakness of any automated translation or AI model is that it only provides one answer. It calculates a singular output based strictly on what is statistically most likely to be accurate.

In international litigation and white-collar investigations, foreign language documents frequently contain phrases with multiple “correct” translations. From a statistical standpoint, a machine may see two distinct phrasing options as nearly identical in probability. However, from a legal standpoint, those options can have radically opposite implications for your case.

The Discovery Risk: As Remu Ogaki notes in the session, multiple statistically accurate translations can have “radically different legal importance where the meaning of one translation may have a dramatic effect on the litigation and the other one may have a dramatic opposite effect […]. You may be leaving evidence that could be very helpful […] on the table by not knowing that this other possible translation is out there, or having an extremely damaging piece of information go out without you knowing that it’s there and handed to the other side in a production.”

The Strategic Value of Human Context

Ultimately, while advanced machine translation and GenAI models have an important, highly effective place in the discovery pipeline when used in the right context, they lack strategic judgment. They cannot understand the unique landscape of a specific lawsuit, nor do they know when they are wrong. A bilingual or multilingual attorney does not just read the words on a page; they understand the multi-layered context, regional idioms, and legal weight of the text. They possess the unique ability to flag when a document has dual meanings and warn you of potential pitfalls before it impacts your strategy.

In an increasingly commoditized world of subscription platforms and “do-it-yourself” translation apps, failing to properly assess the service level behind your workflows invites immense risk. To preserve terminology consistency, maintain strict data security, and safeguard the integrity of your evidence, leveraging expert human editors, project managers, and specialized linguistic review teams remains absolutely vital.