AI and Machine Translation in Legal Settings: Opportunities, Risks, and the Need for Human Oversight

Artificial intelligence and machine translation (MT) are widely seen as tools that could expand access to justice for non-English speakers in the U.S. legal system. Tools such as Google Translate and AutoML deliver strong accuracy in everyday language, but legal language is specialized, high-stakes, and unforgiving of error. Cristina Llop, a legal expert, shared observations from reviewing live translations between legal-provider staff and users at the Stanford Legal Design Lab's AI and Access to Justice webinar. Her findings show both the potential and the pitfalls of AI-driven language access, and point to what is needed to use these tools safely.

MT can provide fast, low-cost support in several legal tasks. It can help translate declarations and witness statements, convert court forms and pleadings into other languages, make court websites and legal guides more accessible, and support real-time interpretation at court help centers and clerk offices. This matters most in under-resourced courts and legal aid organizations that cannot staff a human interpreter for every case. Even so, fluent-sounding output is not the same as legally precise output, and Llop cautions against relying on it without review.

Llop identified recurring mistranslations with serious consequences. Common legal terms are mistranslated because models lack specialized training data: 'warrant' is often rendered as 'court order,' which understates the seriousness of the document. Contextual errors also appear: 'due date' became 'date to give birth,' 'trial' became 'test,' and 'charged with a battery' became 'loaded with a case of batteries.' Pronoun ambiguity is another problem. Spanish 'su' (your, his, her, their) is frequently mistranslated in legal documents, obscuring ownership, responsibility, or who is accusing whom in restraining order cases, which can put victims at risk. MT can also introduce gender bias, defaulting gender-neutral words such as 'politician' to male forms and translating 'me maltrata' as 'he mistreats me' when the source does not specify gender, potentially distorting evidence in domestic violence cases. Without human review, these errors can go unnoticed.

Perhaps the most troubling finding was the invisible breakdown in communication. Both sides often believed they were exchanging information while key facts were missed or misrepresented. Legal providers lacked critical details, users did not realize their information had been misunderstood, and important matters such as the nature of an abuse claim or financial disclosures were lost. The consequences include people choosing the wrong legal recourse, staff making decisions and offering service menus based on a distorted picture of the person's situation, and ultimately compromised access to justice for vulnerable litigants.

Some institutions have added legal glossaries to MT tools to reduce errors, but Llop's research shows these patches are unreliable. A glossary entry forcing 'address' to mean 'mailing address' was then applied when a clerk asked, 'What issue do you want to address?' A fix for 'will' in legal documents distorted ordinary sentences like 'I will send you this form.' Some fixes worked, such as 'due date' and 'pleading,' but failed to correct grammatical structure or pronoun usage. Glossaries address vocabulary, not context, so they cannot substitute for comprehensive training, oversight, and quality control.

AutoML lets legal organizations train MT models on their own specialized data. Llop's team collected more than 8,000 legal translation pairs from official sources, corrected AI output manually, fed the improved translations back into the model, and iterated. Quality improved, but significant issues remained. The model struggled with conversational context, such as switching a pronoun to 'he' when a later message about 'my wife' omitted gender. It also overfit to common legal phrases, inserting 'petition' where 'form' was correct. These results show why human review remains essential even after targeted training.

Text-based translation can be refined over time, but real-time voice-to-text in legal offices is harder. People do not dictate punctuation, yet pauses and commas can change legal meaning, as in 'I'm guilty' versus 'I'm not guilty' when a comma is misplaced. Minor misspellings and grammar errors can also shift meaning. Speech patterns compound the problem: filler words ('um,' 'huh,' 'oh'), quick speech, soft voices, heavy accents, and sentence structures shaped by indigenous or regional dialects all reduce accuracy. Because legal system users come from diverse linguistic backgrounds, these conditions are common rather than exceptional.

Llop's research highlights a critical gap: there are no national standards, training datasets, or quality benchmarks for legal translation AI. A centralized effort to collect high-quality legal translation pairs, localize legal terminology by state, and issue guidelines for AI use in courts and legal aid organizations would allow models to be trained more effectively while protecting legal accuracy. Such infrastructure would give courts and providers a shared baseline for evaluating tools rather than relying on each organization's ad hoc fixes.

Human review is not optional. English-speaking staff can improve AI output by using plain language and short sentences, avoiding vague pronouns such as 'his,' 'her,' and 'their,' and confirming meaning before finalizing a translation. Human-in-the-loop workflows should be standard for any high-stakes legal communication, with trained reviewers checking terminology, pronouns, gender, and context. Oversight makes AI a supporting tool rather than an unchecked authority.

AI and machine translation can meaningfully support language access, but they are far from perfect. Without human expertise, legal-specific training data, and national standards, AI-generated translations could compromise the very access to justice they are meant to expand. Llop's work calls for human-in-the-loop AI translation, better training data for legal contexts, and national standards for AI language access. In law, a mistranslation can change lives, so AI must be designed with legal precision and human oversight from the start.