Can AI interpretation make serious mistakes at a live event?

Yes. AI interpretation can make serious mistakes, and so can human interpreters, only different ones. A human interpreter who is unsure hesitates or speeds up, so listeners can hear it. An AI voice sounds equally confident whether the sentence is right or wrong, which makes a wrong word harder to spot, especially for listeners who do not understand the original language.

Why does AI interpretation get names and brands wrong?

Most AI interpretation mistakes begin before the translation, in speech recognition. If the system hears something other than what was said, the translation is a correct translation of the wrong word. A brand name pronounced in an unexpected way may be wrong every time it is said. Preparing the event with documents that contain the names and brands helps, but no preparation covers what nobody expected.

How can I reduce AI interpretation errors before my event?

Test under the same technical conditions as the event: the same microphones and AV setup, in the real room. Watch the audio quality indicator at the sound check. Send the agenda, presentations, technical papers and a list of speakers, names and brands so the terminology can be prepared. Test the listener languages that matter most, and keep the microphone close to the speaker. Ask all speakers to speak clearly and not too fast. This reduces the risk but does not remove it.

What should an event agency put in its contract when it uses AI interpretation for a client?

Promise the effort, not the result: AI interpretation prepared with the client's material and tested under the conditions of the event, without a guaranteed level of accuracy. State that it may contain errors and omissions and that the spoken original applies where the exact wording matters. Where errors could have serious legal, medical, financial or safety consequences, recommend human interpreters for those parts. Such a clause sets expectations but cannot take away rights the law gives the client.

When should I use human interpreters instead of AI interpretation?

Where an error could have serious legal, medical, financial or safety consequences, such as binding negotiations, the signing of agreements and safety briefings, plan human interpreters for those parts of the event. Some uses, such as emergency services and clinical decisions, are excluded altogether by the Terms of Nubart TRANSLATE.

Rosa Sala

CEO of Nubart

Can AI Interpretation Make Serious Mistakes? What to Tell Your Guests and Your Client

Q: What should I tell attendees when an event uses AI interpretation?

One visible sentence where guests choose their language is enough, for example: "Interpretation at this event is provided by AI. It may contain errors, is not certified and has no legal force. For the exact wording, the spoken original applies." Institutions such as the United Nations, the European Parliament and the German Bundestag use a similar disclaimer for human interpretation.

Event organizer with a folder of documents standing in an empty conference hall before an event.

In more than 30 years as a human interpreter, no client ever asked me what would happen if I made a serious mistake. Nobody asked me to have my own clients confirm in writing that interpreters are human and can get things wrong. Now I run Nubart TRANSLATE, an AI simultaneous interpretation system for live events, and event organizers ask me exactly that: if the AI makes a serious mistake at an event we run for a client, where does that leave us? Yes, AI interpretation can make serious mistakes. So can human interpreters, only different ones. This article explains where the AI mistakes come from, what you can do about them before the event and what to put in the agreement with your own client.

Human and AI interpreters make different mistakes

Interpreters make mistakes. We mishear a number, choose the wrong word or lose the end of a long sentence. Listeners hear it in the voice: a human interpreter who is unsure hesitates, or speaks very fast to catch up with the speaker. An AI voice sounds equally confident whether the sentence is right or wrong. A wrong word is therefore harder to spot, especially for listeners who do not understand the original language. I describe the other differences in AI versus human interpretation.

Where AI mistakes come from

Most of them begin before the translation. If the speech recognition hears something other than what was said, the translation is a correct translation of the wrong word. Take a brand name. The speaker says "Nike" but pronounces it in a way the system does not expect and does not connect with the brand. The result will probably be wrong every time the speaker says it. Preparation helps: if the brand is in your documents, Nubart adds it to the preparation, and it is usually recognized. But nobody can foresee every pronunciation, and no preparation covers what nobody expected. A speaker who talks about cars may suddenly compare Nike and Puma as an allegory. If the system was not prepared for it, repeating the comparison repeats the mistake.

Speakers who are hard to understand

The same happens when the speech itself is hard to follow. An accent is only one cause. Others are mumbling, many "ums" (only some of them can be filtered out), very fast speech, very long sentences and a low voice. The speech recognition struggles with all of these, and so does a human interpreter. There is no fix during the speech.

There is also a social side. Nobody tells a celebrated chef that his sentences are confusing, or a well-known German theologian that the Spanish in which he wants to lecture in Spain is not good enough. I have interpreted both. A human interpreter cannot repair that, and neither can an AI.

What helps before the event

You often cannot test with the real speakers. A head of government will not give you 30 minutes of his diary, and at an international conference with 20 speakers you may not know most of them. What you can do is test under the same technical conditions as the event: the same microphones and the same AV setup, in the real room, with a colleague speaking. Then at least no technical cause makes the recognition harder. At the sound check, watch the audio quality indicator on the speaker's screen. It shows how clearly the speech recognition can understand the speaker. A low value can come from the equipment, but also from mumbling or a strong accent, so it works as an early warning.

Nubart offers a free test for up to 30 minutes for this. The free test does not include event preparation. If you want to hear your own terminology prepared, ask about the paid prepared test. A good result in one listener language says little about the others, because translation and voice output differ by language, so test the languages that matter most. Testing creates some confidence, but it does not remove the risk. A speaker with a cold can pronounce a word differently on the day. Say so in the agreement with your client.

Send Nubart your agenda, presentations, technical papers and a list of speakers, including the names and brands that will come up. Nubart prepares the terminology for your event from them, much like briefing a human interpreter. This improves the chance that names and terms come out right. It does not guarantee it. A microphone close to the speaker helps more than most organizers expect.

If you organize the event for a client, invite the client to the test. Then they hear for themselves what AI interpretation sounds like, and their expectations come from experience, not from a paper.

What to tell your guests and your own client

Your guests

One visible sentence is enough, in the place where guests choose their language: on the event page or next to the QR code. For example: "Interpretation at this event is provided by AI. It may contain errors, is not certified and has no legal force. For the exact wording, the spoken original applies." The Terms of Nubart TRANSLATE require customers to tell participants that the interpretation is provided by AI. A sentence like this does that.

Institutions have used the same idea for human interpreters for decades. The United Nations, the European Parliament and the German Bundestag publish a disclaimer saying that interpretation serves communication and that only the original speech is authentic.

Your own client

If you organize the event for a client, promise the effort, not the result. You can promise AI interpretation that is prepared with the material your client sends you and tested under the conditions of the event. Do not promise a level of accuracy. The Terms of Nubart TRANSLATE do not guarantee one either, so you could not pass such a guarantee on.

Put this into your offer or contract in plain words. For example: "The interpretation is provided by an AI system (Nubart TRANSLATE). It is prepared with the material you send us and tested under the conditions of the event. It may contain errors and omissions, and we do not guarantee a specific accuracy. Where the exact wording of a statement matters, the spoken original applies." If parts of the event could have serious legal, medical, financial or safety consequences, add that you recommend human interpreters for those parts. Your client signs the offer or contract anyway, so a separate form is not needed. A clause like this sets expectations. It cannot take away rights that the law gives your client.

When a human interpreter is the better choice

Where an error could have serious legal, medical, financial or safety consequences, plan human interpreters for those parts of the event. Examples are binding negotiations, the signing of agreements and safety briefings. The Terms of Nubart TRANSLATE exclude some uses altogether, such as emergency services and clinical decisions. The reverse case is different: human interpreters as the main channel and AI as the backup if they cannot make it. I describe it in What to do when your interpreters can't make it to the event.

Nubart TRANSLATE is built for in-person and hybrid events. On the Nubart TRANSLATE product page you can request a free test for up to 30 minutes and try it with your own speakers and material.