AI or a live interpreter
Machine interpreting or an interpreter in the booth – when does each hold?
Text from an AI often looks right. That does not mean the listener heard what was said. Here is the difference, as a class-A conference interpreter describes it.
A machine looks right at first glance
The microphone becomes text, and the text becomes a voice. Hear “low mortality” as “high mortality”, and the next sentence can say the opposite.
A live interpreter in the ISO booth
An interpreter can mishear too. The mistake is usually caught before the room hears it, and the same term is kept when the script breaks.
How the sound passes through the machine
This comes from a class-A conference interpreter with full training and certification.
AI has become very good at producing text quickly, and at first glance the text looks convincing. In practice the microphone picks up the voice, the background noise and other sounds. Speech recognition turns that sound into text. A language model finds a wording in the other language. Out comes the text, or a synthetic voice reading it aloud.
Dialect, accent, unclear speech, mumbling, a fast pace and specialist terms make the system hear the wrong word. It does not follow the context, so it confuses words that sound alike, or picks a common word where a rare technical term belonged. It is especially weak on dialect, and when the source language is not the speaker’s first language. That produces a lot of misunderstandings.
A human interpreter can mishear too. She weighs the context and usually clears the misunderstanding before it is said in the room.
When the words only look right
The translation is fast, and it often sounds right. Idioms, negations, unclear references and sentences that only make sense together often come out wrong. It happens especially when someone speaks freely. Spoken language is not written language, and the models are trained on writing.
The same term also turns into different words along the way, and then it gets unclear. An interpreter prepares, builds the terminology and stays with the same word throughout. Specialist material needs specialist knowledge.
Norwegian «tillatelse» can be permit, permission or licence, depending on what is meant. «Det er vanskelig å huske nøyaktig når det skjedde» came out as «It is difficult to swing exactly when it happened.» It looks trivial. Someone who only hears the English has no way to recover what was said.
And the errors amplify each other. If the system hears «lav dødelighet» as «høy dødelighet» — low mortality as high mortality — the next sentence can be grammatically correct and still say the opposite. The mistake in the speech recognition drives the translation.
As an example: the AI cost is x. The loss from an error is y. The loss against a third party, a conflict, or reputation is z.
When the script breaks
Events rarely follow the script. People leave their notes, speak fast, have a strong accent, interrupt, answer off the cuff, point at a slide or make a joke. A trained interpreter adapts, prioritises and keeps the thread, also when it turns emotional or diplomatically sensitive.
She notices when a wording becomes ambiguous in the other language, and adjusts a little so the misunderstanding does not land. It is not only translation. It is an adjustment so the meaning can be understood.
Simultaneous interpreting from a trained interpreter gives natural, fluent speech. You hear it in a keynote, a panel, negotiations and ceremonies. The human delivery holds attention and trust through long meetings. AI interpreting is still monotonous, and it is not shaped to the content. It is hard to follow for long.
A human interpreter notices when she does not understand, and finds a way through that does not wreck the exchange. AI delivers something anyway.
A professional conference interpreter carries intention, argument, tone, irony, humour, what sits underneath, and cultural references. She handles unstructured speech and spontaneous departures, uses intonation and adjusts the pace, reads the mood in the room and checks the interpreting as it goes. That is why she avoids many of the misunderstandings AI produces.
When the speech leaves the room
With AI interpreting, the speech can go to a third party, be stored, be used to train the model, or fall under another jurisdiction. It depends on the vendor and the contract. A recording can hold sensitive information, and the handling varies. That covers confidential board meetings, acquisitions, legal matters, health data and politically sensitive content. A human interpreter can work under a duty of confidentiality, in controlled conditions.
When the machine is enough
The machine can hold in informal, low-risk situations. It is useful for glossaries, preparation and drafts, and for live captions when the error rate is stated plainly. Captions still mishandle context and self-repair. They are an aid, not a substitute when an error cannot be taken back.
Some systems can be switched on without an interpreter booked in advance. They can scale to many people and several languages, and some produce a transcript or minutes. It depends on the system. For preparation the machine is often strong. In the meeting itself, when someone leaves the script, it is the live interpreter who corrects.
Questions people ask
Can AI replace a live simultaneous interpreter?
Not when an error cannot be taken back. AI can hold in informal, low-risk situations, and as support for glossaries, drafts and captions where the error rate is plain. Courts, boardrooms, negotiations and demanding meetings still need a live interpreter who can correct in real time.
When are live captions enough?
As an aid. Rarely as the only language path when errors have consequences. Captions still mishandle context and self-repair.
Can AI prepare, and a live interpreter take the meeting?
Yes, that is often what holds. AI for glossaries, terminology and briefing. A live interpreter in the moment when errors have consequences.
Why can’t the machine just correct itself?
Simultaneous interpreting goes out continuously to the people listening. When the context arrives later, a human interpreter can reshape the next sentence. A pipeline that has already sent the wrong text or audio cannot take back what was heard.
What if AI hears “low mortality” as “high mortality”?
The error is amplified. The next sentence can be grammatically correct and still say the opposite, because the mistake in speech recognition drives the translation. A human interpreter can mishear too, but weighs the context and usually clears the misunderstanding before it is said in the room.
Why did “huske” come out as “swing”?
AI confuses words that sound alike, or picks a common word. «Det er vanskelig å huske nøyaktig når det skjedde» came out as «It is difficult to swing exactly when it happened.» It looks trivial, but someone who only hears the English cannot recover the meaning. The same fog applies to terms: Norwegian «tillatelse» can be permit, permission or licence, and AI often switches the word along the way. An interpreter prepares the terminology and keeps the same word.
Where does the speech go in AI interpreting?
It can go to a third party, be stored, be used to train the model, or fall under another jurisdiction, depending on the vendor and the contract. That covers confidential board meetings, acquisitions, legal matters, health data and politically sensitive content. A human interpreter can work under a duty of confidentiality, in controlled conditions.
What does the cost look like, as an example?
As an example: the AI cost is x. The loss from an error is y. The loss against a third party, a conflict, or reputation is z.
If the interpreter is going in a booth, see booths and receivers, or send a request.