Beyond AI: The Irreplaceable Human Element (Agustín de la Mora’s keynote @ NAJIT 47th)
One of the highlights of the 47th Annual Conference of the National Association of Judiciary Interpreters and Translators (NAJIT), was the keynote presentation by DE LA MORA founder Agustín de la Mora. Agustín delivered a thought-provoking speech exploring one of the most important questions facing our profession today: What remains uniquely human in an age of artificial intelligence?
Whether you joined us in Atlanta or were unable to attend, we’re excited to share the full transcript of Agustín’s keynote. We hope his message inspires you to embrace the future of our profession while continuing to recognize the value that only humans can bring to the work we do.
Read the full transcript below:
Beyond AI: The Irreplaceable Human Element
Good morning, madam chair, distinguished members of the board, colleagues, honored guests, and fellow guardians of justice. It is truly an honor to be here. I must tell you that 35 years ago when I went to my first NAJIT conference, I heard Dr Rainoff give the keynote speech and saw him up there speaking with such authority, having the attention of all these professionals, and I thought to myself: “I hope one day I’ll be able to do that”. Today I’m saying to myself; “what were you thinking!” This is scary and certainly a daunting task. I am really humbled by the opportunity. So, thank you for the invitation and thank you for the work you do.
Someone once told me that good interpreters are like ninjas, because just like ninjas, if nobody notices us, then we are doing a good job. And when interpretation is done well, the hearing flows. The questions make sense, the answers connect, the process works. And because it works so smoothly, it can feel as if nothing special is happening.
But of course, something very important is happening. Understanding is happening. And for the people who depend on it, those standing in courtrooms trying to navigate unfamiliar systems in unfamiliar languages, that understanding changes everything.
Sometimes people talk about language access as just an administrative function as if it’s just part of the logistics of running a court. But we know it’s much more than that. It’s the mechanism that turns legal rights into lived reality. Without it, rights exist on paper, but not in practice. I think we are all aware that having an interpreter is certainly not a language right. It is an access right.
Imagine if you will, a courtroom where every word spoken is instantly translated by a machine. No pauses, no human presence, just algorithms humming in the background. It sounds efficient, perhaps even inevitable.But is it justice?
Every day, across courtrooms, hearing rooms, and legal settings of every kind, people walk into spaces where the stakes are high and where they may not understand what is being said around them. Imagine being in that position. You hear the words. You recognize that decisions are being made. You sense the authority in the room. But you don’t fully understand what’s happening, what’s being asked, or what it means for your life.
That’s not just confusion. That’s vulnerability.
And then a voice enters; a trained, professional, human voice.
The interpreter.
Suddenly, meaning begins to take shape. Questions become clear. Answers become possible. Participation becomes real.
That’s our work. That’s your impact.
You are not just language professionals. You are, in a very real sense, defenders of meaning. The people who ensure that understanding survives in places where it matters most.
And that’s exactly why the conversation we’re having today matters.
In our legal system, justice begins at the moment understanding becomes possible. And for millions of people in this country, understanding becomes possible because of you.
I said millions, and it is not hyperbole.
The Administrative Office of the U.S. Courts' 2024 Annual Report, states:
"Courts used interpreters in 196,155 court proceedings in the 12 months ending September 30, 2024."
According to a study by the Judicial Council of California, “California’s superior courts recorded more than 2.5 million interpretations between fiscal years 2020–21 and 2023–24.” (2025 Language Need and Interpreter Use Study.)
Language access is not a peripheral issue in immigration court—it is the norm.
Research indicates that roughly 85–89% of immigration court respondents participate in proceedings in a language other than English. At the same time, the immigration court backlog exceeded 3.3 million pending cases in early 2026. If those proportions hold, more than 2.8 million pending cases are likely to require interpreter services, illustrating the enormous and continuing demand for qualified immigration court interpreters.
These numbers are not abstract; they are lives, families, freedoms.
We gather today not only to ask whether artificial intelligence will replace interpreters in courtrooms, but to explore what lies beyond AI. What it cannot replicate, what it cannot understand, and why the human interpreter remains indispensable.
From early machine translation in the 1950s to today’s neural networks, AI has made astonishing progress. Apps can translate menus in real time; chatbots converse across languages, and speech recognition systems transcribe with impressive accuracy.
AI is fast. It can process enormous amounts of data. It can generate text in seconds. It does not get tired. It appears efficient. For administrators, that is appealing. For institutions under pressure to do more with less, it is tempting. But there is something important we must keep in mind:
Efficiency is not the same as justice, and speed doesn’t guarantee understanding.
Producing language is not the same thing as preserving meaning.
Pilot programs are already testing AI transcription and translation tools in courtrooms. The attraction is obvious: speed, cost reduction, scalability, efficiency.
AI excels at:
Handling large volumes of text quickly
Providing instant draft translations
Supporting interpreters with terminology databases
But we know that courtroom language is not just words; it is tone, intent, cultural references, and legal precision. AI struggles with sarcasm, idioms, layered meaning, and dialects. A mistranslation in a restaurant may cause confusion. A mistranslation in a courtroom may cost someone their freedom.
AI systems inherit biases from their training data. They may misinterpret minority languages or marginalized speech patterns. And when mistakes occur, accountability becomes murky. Who is responsible when AI mistranslates testimony? A machine cannot be cross-examined.
Justice demands more than efficiency. It demands accuracy, fairness, and human judgment.
Through accurate and complete interpretation, interpreters ensure defendants, victims and petitioners have an opportunity to understand charges; witnesses are heard accurately, and judges receive precise meaning.
This is not a technical service. It is a constitutional right.
We the interpreters, bring human judgment, cultural competence, and ethical responsibility to the encounter. We can pause proceedings to clarify, detect misunderstanding, and adapt to human emotion. Courtrooms rely on credibility. A defendant must trust that their voice is heard. That trust is human, not mechanical.
One human voice entering that space—carrying meaning across language—is not a convenience, not a luxury, but a constitutional necessity.
Instead of fearing replacement, interpreters can harness AI:
Glossaries generated by machine learning
Real-time draft transcripts for support
Terminology search engines
The future is not interpreter versus machine, but interpreter with machine.
So, do AIs make mistakes? You bet.
Consider the question:
“Did you have anything to drink in the car?”
Seems straightforward, right? Let’s ask AI:
Is the sentence "Did you have anything to drink in the car" ambiguous?
Yes, the sentence “Did you have anything to drink in the car?” can be somewhat ambiguous, especially in legal or police contexts.
It can be interpreted in at least two ways:
Location of drinking
“Did you drink anything while you were in the car?”
This suggests consumption physically occurred inside the vehicle.
Possession/availability in the car
“Did you have any drinks with you in the car?”
This could refer to beverages being present in the vehicle, not necessarily consumed there.
In everyday conversation, most people would probably understand it as asking whether the person drank something while in the car. However, in courtroom interpretation, police interviews, or legal translation, that ambiguity matters.
If the intent is specifically about consumption while driving or riding, clearer English would be:
“Did you drink anything while you were in the car?”
“Did you consume any alcohol in the car?”
If the intent is about possession:
“Did you have any alcoholic beverages in the car?”
This is a good example of why literal translation can create problems in legal interpreting.
The Spanish version:
“¿Tomó algo de Beber en el carro?” leans more toward actual consumption, not merely possession.
The previous answer from the AI makes it clear; it understands the possibility of confusion due to ambiguity. Great, right? We’re all with it; ambiguity is clear. Nevertheless, when I asked AIs directly for a translation this is what we get:
ChatGPT:
Translate this question into Spanish: "did you have anything to drink in the car?'
“¿Tomó algo de beber en el carro?”
That’s it, no explanation, nothing.
How about Gemini? This one tends to be chattier. Let’s see:
Gemini:
Translate this question into Spanish: "did you have anything to drink in the car?'
Here’s the Spanish translation of your question:
“¿Tomaste algo de beber en el coche?”
Would you like me to also give you a slightly more formal version, in case you’re aiming for courtroom or professional interpreting use?
Yes
Here’s a more formal rendering of your question in Spanish, suitable for courtroom or professional interpreting contexts:
“¿Consumió usted alguna bebida dentro del automóvil?”
This phrasing uses consumió (consumed) and automóvil (automobile), which are more formal and precise than tomar and coche.
Again, no mention of ambiguity, no hesitation. And that’s the interesting part, no hesitation whatsoever.
DuckAI:
Translate this question into Spanish: “did you have anything to drink in the car”
¿Tomaste algo para beber en el coche?
Claude:
Translate this question into Spanish: “did you have anything to drink in the car”
The Spanish translation is:
¿Tomaste algo para beber en el coche?
Alternative translations depending on context:
“¿Bebiste algo en el carro?” (more direct)
“¿Tomaste algo para beber en el coche?
Note: “Carro” is more commonly used in Latin America, while “coche” is more common in Spain.
DeepL:
Translate this question into Spanish: “did you have anything to drink in the car”
«¿Bebiste algo en el auto?»
Alternatives
"¿Has tomado algo de beber en el auto?"
«¿Has bebido algo en el auto?»
None of the IAs asked for clarification or even mention ambiguity. Certainly, none of them asked me: “is this sentence about someone who was actively drinking or had drinks in the car?”. AI systems tend to choose one interpretation and move forward. A professional interpreter immediately recognizes that there are multiple possible meanings—and that the ambiguity itself matters.
That ability—to notice uncertainty, to recognize when meaning is unclear is deeply human. It is part of professional judgment.
Because human communication is rarely neat or precise.
People hesitate. They interrupt themselves. They imply things without saying them directly. They rely on shared assumptions. They use humor, metaphors, tone, and cultural references.
And when people are under stress, or have experienced trauma, communication becomes even more complex. Stories may be fragmented. Details may come out of order. The way something is said may not match what someone expects.
That does not mean the person is unreliable. It means they are human.
And understanding that distinction can affect how testimony is perceived, how credibility is assessed, and ultimately, how decisions are made.
Convincing falsehoods are more dangerous than obvious errors. A human interpreter can say: The interpreter…
is unfamiliar with the term
requests counsel to repeat the question
needs to make a correction for the record
AI often cannot. Instead, it may generate an answer that sounds fluent, confident, and authoritative, even when it is fundamentally wrong.
And that is extraordinarily dangerous in a courtroom. Because in legal settings, people do not lose points. They lose freedom. They lose credibility. They lose asylum claims. They lose parental rights. They lose cases. Sometimes they lose years of their lives.
Interpreters are trained to recognize uncertainty, to pause, and to clarify. That humility, the ability to admit “I don’t know” is a safeguard of justice. Machines lack that humility.
As machines become better at producing language, what remains uniquely human?
To answer that, we have to clear up a common misunderstanding:
As you know, lots of people think interpreting is just about replacing words, taking something said in one language and swapping it into another. Patricia Michelsen-King, used to tell a story about one of our colleagues who during a hearing informed the Court that she was not informed of the nature of the hearing and that she could not understand the testimony of the witness who was using very technical vocabulary and therefore she was not able to interpret, According to the story, the judge said: “just interpret the words, you don’t need to understand” Now, I don’t know if the story is apocryphal, but I am sure many of you have heard different versions of it.
But you know it’s not that simple.
Interpretation is not about words. It’s about meaning.
Words are just the surface. Meaning lives underneath. shaped by context, culture, intention, tone, and experience.
In some cultures, direct eye contact signals confidence. In others, it can be seen as disrespectful. Silence might suggest hesitation in one setting, and careful thought in another. The behavior doesn’t change. The meaning does. And that’s where your role becomes so critical.
You’re not just working between languages. You’re working between systems of meaning. Between expectations. Between cultures. And often, between people and institutions that hold very different kinds of power.
Because every courtroom is shaped by power. Authority, hierarchy, procedures; all of it creates an environment that can feel overwhelming, especially to someone who doesn’t fully understand what’s happening.
Interpreters stand right in the middle of that.
You notice when someone is nodding without understanding. You sense when something isn’t landing the way it should. You know when language is becoming a barrier instead of a bridge.
And you operate within a framework of ethics that guides how you respond.
Accuracy, impartiality, confidentiality, these principles aren’t abstract. They guide real decisions, moment by moment. When to ask for clarification. When to preserve ambiguity. How to convey meaning faithfully, even when it’s difficult.
That kind of judgment cannot be automated.
And it matters because legal systems depend on accountability. When an interpreter makes an error, there’s responsibility. There is a standard. There is a process for addressing it.
With machines, that becomes much less clear.
And that is one of the reasons your role is not peripheral. It is foundational.
Language access is part of what makes constitutional rights real. The ability to understand proceedings, to consult with counsel, to participate meaningfully in one’s own case, it all depends on communication. Without that, justice is incomplete.
Let me be clear: having an interpreter is not a language right. It is an access right.
Now, none of these means there is no place for technology. There absolutely is.
AI can help with preparation, research, terminology, and access to information. It can support your work in meaningful ways.
But there’s a line that matters.
Technology can assist. But it cannot replace judgment, it can generate language; but it cannot guarantee meaning; it can produce answers; but it cannot take responsibility for them.
The future is not about rejecting AI. And it is not about surrendering to it either.
It’s about understanding where it fits, and where it does not.
Because justice is, and will remain, a human endeavor.
It requires people who can listen, interpret, evaluate, and take responsibility. It requires people who can hold complexity, recognize uncertainty, and act with integrity.
It requires defenders of meaning.
Interpreters must follow ethical codes. They are accountable to professional standards. AI systems, by contrast, are accountable only to their developers, and often not at all to the people whose lives they affect.
The path forward is not rejection of AI, but responsible integration. Interpreters must lead in shaping how AI is used in justice. That means:
• Establishing clear ethical boundaries.
• Demanding transparency in AI systems.
• Using AI as a support tool, never as a substitute for sworn testimony.
Finally, we have to consider that using AI in legal matters goes beyond mistranslations, lack of recognition of ambiguity or not asking for clarification.
Consider the following case study of Mata v. Avianca:
Mata v. Avianca, Inc., No. 1:2022cv01461 - Documento 54 (S.D.N.Y. 2023)
Roberto Mata sued Avianca Airlines for personal injuries, claiming he was struck in the knee by a metal serving cart during a 2019 flight from El Salvador to New York.
When Avianca sought to dismiss the case (partially citing statute of limitations under the Montreal Convention), Mata’s attorney, Steven Schwartz, used ChatGPT to conduct legal research to support his arguments.
ChatGPT completely "hallucinated" (invented) six fake judicial opinions, complete with fabricated case names (e.g., Martinez v. Delta Airlines), citations, and internal quotes.
The presiding judge (U.S. District Judge P. Kevin Castel) could not find the cited cases and ordered Schwartz and his colleague, Peter LoDuca, to provide the actual opinions.
Instead of realizing the mistake immediately, the attorneys used ChatGPT a second time to verify the fake cases and filed affidavits reinforcing the non-existent citations.
In June 2023, Judge Castel sanctioned Schwartz and his law firm (Levitow, Levidow & Oberman). The attorneys were fined $5,000 and were forced to notify their real clients of the misconduct.
I think they got off easy, right? A cheap fine and they were forced to notify the real clients of the misconduct.
Did it have any repercussion? Absolutely:
The case became the legal industry’s most infamous warning about the dangers of using generative AI for legal research without human verification.
It led to numerous courts—including the Southern District of New York—updating local rules to mandate that lawyers certify that no portion of their filings was drafted by generative AI or explicitly verifying all AI-generated content.
Think carefully about what that means: if AI can invent law itself, can it truly be trusted to preserve sworn testimony accurately?
Just this morning I was talking to James Plunkett, my mentor, and he said that something similar happened in his courtroom. One of the attorneys presented some kind of argument, and when the judge said, “Well, let me verify it”, and took out her book, it turned out that the citation used in the argument was not in the book. Maybe it was a wrong addition or was in a newer version? Until they found out that the citation used was made up by AI.
So, these are not hypothetical concerns. Reuters documented the Mata v. Avianca case extensively.Researchers studying legal AI systems have found persistent hallucination problems in advanced legal language models.In evaluations, even sophisticated systems generated inaccurate or fabricated information at concerning rates.
Interpreters, translators, and educators must be at the table when courts consider AI adoption. You may think, “Good luck with that”, but we can still try. Because only those who understand the stakes of language can safeguard justice.
The question is not whether AI will enter the courtroom—it already has. The question is whether justice will survive its entry. And the answer depends on us.
In the courtroom of the future, let AI assist, but let interpreters ensure justice speaks in every language, with accuracy, fairness, and humanity.
Let me close with this.
Years from now, people will look back on this period as a time of rapid technological change. They’ll talk about artificial intelligence, automation, and new ways of working.
But I think they’ll also ask something deeper.
When machines became capable of producing language, who made sure that meaning was not lost? Who made sure that people were still truly understood? Who ensured that justice remained something more than just efficient?
I believe the answer will point at professionals like you.
Because in the end, going beyond AI is not about rejecting technology, it’s about remembering what it cannot do.
It cannot replace judgment. It cannot replace responsibility. And it cannot replace the human capacity to understand another person.
Justice still depends on those things.
And it still depends on you.
Bill Wood (founder of DS-Interpretation, Inc.) said it best. He said, “Interpreters are not going to be replaced by technology, but some interpreters will be replaced by interpreters who use technology.”
Thank you for the work you do. Thank you for being defenders of meaning. And thank you for carrying the human element forward—beyond AI.
Agustín Servin de la Mora is the president of DE LA MORA Institute of Interpretation. He was born and raised in Mexico City, Mexico, and has been a professional freelance and staff interpreter for the last 30 years. He is one of the Supervisor Raters for the National Center for State Courts and has been a lead rater for the federal and consortium oral exams for court interpreters. He was the lead interpreter for the Ninth Judicial Circuit for over a decade and served as a member of the Project Advisory Committee responsible for the creation of the National Standards for Healthcare Interpreter Training Programs for the NCIHC. He was a member of the Florida Court Interpreter Certification Board and a voting member of the Technical Committee of the National Consortium for Interpreter Certification. He is a federally certified court interpreter. He has been a consultant for the Administrative Offices of the State Courts, conducting orientation seminars and advanced skills workshops for interpreters in over 20 states. He has been featured as a speaker and presenter in several national conventions, including those of NAJIT, ATA, IMIA, and NASCA. His innovative methodology, Weightlifting for Court Interpreters™, is valued by interpreters of all fields for its focus on deliberate practice and performance improvement. He is also the author of the acclaimed book Weightlifting for Court Interpreters: Mastering Performance Through Deliberate Practice.
References
• Hale, S. (2011). Community Interpreting. Palgrave Macmillan.
• Morris, R. (2010). “Court Interpreting: The Trial of the Century.” The International Journal of Speech, Language and the Law.
• O’Connell, M. (2020). “Artificial Intelligence in Legal Translation: Opportunities and Risks.” Journal of Law and Technology.
• European Commission (2022). Study on the Use of AI in Judicial Systems.
• American Bar Association (2023). Ethics and AI in Courtrooms.
• García, B. (2019). “Bias in Machine Translation and Its Impact on Minority Languages.” Language and Society Review.
Reuters (2023). Coverage of Mata v. Avianca case.
• Legal AI Research (2024–2025). Studies on hallucination rates in advanced language models. Bipartisan Policy Center. (2020). Language access in the immigration system: A primer.https://bipartisanpolicy.org/article/language-access-in-the-immigration-system-a-primer/
• Transactional Records Access Clearinghouse (TRAC). (2026). Immigration Court Operations: February 2026 Update.https://tracreports.org/whatsnew/email.260324.html
• National Association of Judiciary Interpreters and Translators (NAJIT). (2025). What's going on in immigration courts?https://najit.org/proteus/whats-going-immigration-courts-winter-2016-17/
With the collaboration of: ChatGPT, Claude, Duckai, Gemini, DeepL and Google.
