How Do You Judge Communication After an AI-Translated Interview?

What does a translation layer actually change?

Less than it felt like at the time. The layer changes the surface: word choice, idiom, rhythm, and the small talk that used to make the first five minutes easy. It does not choose which trade-off to defend, name the constraint nobody mentioned, or produce a second answer that holds when a follow-up moves the target. Most of what the round existed to measure is still there.

Six things come through intact, and they are the ones worth writing down:

• The decision, and the reason given for it before you asked for one.
• What they demanded first: which figure, which document, which person.
• What they refused to guess at, and what they said they would have to check.
• The correction, when a follow-up made the first answer wrong.
• The question asked back, and whether it was about the constraint or about the format.
• The order of operations, and what each step would settle.

What does not come through is fluency, and fluency is what you are most likely to grade by accident. In a University of Chicago experiment, native English listeners rated identical trivia statements as less true when the speaker had an accent: 6.95 for a mild accent and 6.84 for a heavy one, against 7.59 for native speakers, on a 14 cm scale where a higher number means more truthful. Every speaker was reciting statements the experimenter wrote, so prejudice about the source of the claim could not explain it, and listeners told that processing difficulty was the cause corrected for a mild accent but not a heavy one ¹.

That is the reason to run the ladder rather than trust the impression. The same follow-up questions that expose whether someone understands the answer they just gave work through an interpreter, because each one asks for a different fact rather than a better sentence.

Does the job actually require unmediated English?

Answer that in writing before you judge anything. In the United States, the EEOC's enforcement guidance on national origin discrimination, issued 18 November 2016, holds that an English fluency or proficiency requirement "is permissible only if required for the effective performance of the position for which it is imposed," and that the degree of fluency lawfully required varies from one position to the next, assessed case by case ².

The same guidance sets a two-part standard before an accent may factor into a decision: evidence that effective spoken communication in English is required to perform the duties, and evidence that this person's accent materially interferes with their ability to communicate in spoken English ². A discernible accent is not evidence of either. The regulation behind the guidance is blunter still. National origin discrimination includes denying equal employment opportunity because an individual has "the physical, cultural or linguistic characteristics of a national origin group" (29 CFR 1606.1), and when the Commission investigates a selection procedure for adverse impact on the basis of national origin it applies the Uniform Guidelines on Employee Selection Procedures ³. Those guidelines define a selection procedure as any measure or procedure used as a basis for an employment decision, running from paper and pencil tests "through informal or casual interviews and unscored application forms" ⁴. Your round is one of those. This is public legal fact rather than advice, and the specific call belongs with counsel.

So write the requirement down and be honest about which side of the line the role sits on.

• Genuinely required. An on-call incident commander who runs a live English bridge at 3am. A revenue cycle specialist who phones a US payer and argues a denial in real time. A consultant who takes questions from a client's board with no deck to hide behind.
• Usually not. A role whose output is written and reviewed, whose standups are already captioned, and whose team runs in three languages on a normal Tuesday. Here the interview's language was a courtesy, not a duty.

If it is required, name it in the posting, test it as its own short segment on a real task, unaided, with the same prompt for everyone, and say in the invite that the segment exists. Testing it openly is also the only version you can defend later. The move is the same one that separates a condition from the construct in ADA accommodations on an AI-based assessment: what changes is how the person reaches the task, and the task still measures what the job needs.

Where does the translation pipe put errors, and whose are they?

In two places, and neither of them is the candidate. Speech recognition transcribes first, a model translates second, and both stages fail unevenly across speakers and language pairs. The damage arrives as a dropped clause, a technical term swapped for a general one, or a hedge flattened into a flat claim. Grade any of that as imprecision and you are grading the pipe.

How uneven the first stage is has been measured. Across five commercial systems from Amazon, Apple, Google, IBM and Microsoft, transcribing 19.8 hours of structured interviews with 42 white and 73 black speakers, the average word error rate was 0.35 for black speakers against 0.19 for white speakers, and the gap held on a subset of identical phrases spoken by both groups ⁵. That study measured race in American English rather than non-native speech, so the transferable finding is the narrow one: recognition error is a property of the speaker as much as of the system, and a garbled answer says nothing about a person until you know which stage garbled it.

The second stage errs too, and it errs on prepared text. In a 2019 study of 100 sets of emergency department discharge instructions containing 647 sentences, machine translation rendered 594 sentences (92%) accurately into Spanish and 522 (81%) into Chinese; 2% of all sentences in Spanish and 8% in Chinese carried errors with potential for clinically significant harm ⁶. Written, edited, reviewed text, and the accuracy still moved with the language pair. A live interview is the harder case, not the easier one.

Four habits keep the pipe out of your notes:

• Mark it unreadable, not wrong. When an answer arrives incoherent, say so and ask for it again. An unreadable answer written down as a weak one is a measurement error you introduced.
• Put figures and names in writing. Share a document. Ask for the number typed rather than spoken. Digits survive translation; a spoken decimal often does not.
• Send a glossary with the invite. Ten terms the round will use, in English, to every candidate.
• Confirm the term, not the sentence. "When you said the base case, did you mean the 2024 actuals or the plan?" costs fifteen seconds and settles most of it.

Rewrite the round so the evidence isn't a sentence

Ask for things that survive a bad translation: a number, a choice between two named options, a source, an order of operations. "It depends on the discount rate" carries the same information through an interpreter that it carries in English. Then push once on the reason, and once more on what would have to change for the answer to flip.

Five roles, and what to hand over in each:

• Financial analysis. Hand over a two-page packet in which one growth figure is not supported by the filing behind it. Ask which claim they checked, what they found, and what it moved in the recommendation.
• Software engineering. Show a failing test and a stack trace. Ask which file they would open first and what they expect to find in it. A stack trace is the same document in every language.
• Management consulting. Give two entry approaches and one budget. Ask which they would drop first, and what number would have to move for them to swap.
• Healthcare revenue cycle. Give three denials with their codes. Ask which one they would concede, and why that one rather than the other two.
• Marketing. Give them the most on-message statistic in the brief. Ask what population it was measured on, and what they would do if the answer were forty customers.

Write the answers into the same shared document for every candidate, in the same columns, before anybody debriefs. That is what makes a translated round comparable to an untranslated one: the same questions, the same rubric, the same record. The mechanics are the ones behind a structured interview about AI use, where the rubric gets written first, and they carry more weight here, because the part you are most tempted to write from memory is the exact part the layer distorted.

What if the candidate never said they were using it?

Silence here is a gap in your invitation, not a lie. An ordinary interview invite says nothing about whether captions, an interpreter or a translation tool are allowed, so a candidate who brought one broke no rule anybody wrote. Write the line for next time, offer the choice to everyone, and judge this candidate on the answers you already collected.

The line, short enough to paste into the invite:

• This round can run in English, or with live translation or captions. Say which you would like when you book.
• The questions and the rubric are identical either way.
• Where the role carries a language requirement, it is named here, along with the segment that tests it.

Asking up front does more than close the gap. A candidate hides the tool because needing it looks like a cost to them, which is also why disclosure demanded after the call is worth so little: it arrives as an accusation with a form attached. Offering it in writing turns the medium into a choice you made available rather than a secret somebody kept.

It also stops the failure this question usually becomes, where the medium stands in for a doubt about the person. Whether the doubt is about identity, about AI use in the answers, or about candidates recording the interview and running it through a model, each is a separate question with its own answer, and none of them is settled by how the words reached you.

If the job's real question is whether this person catches a confident wrong number before it ships, no interview was ever going to answer it. A translated round, like an untranslated one, catches a candidate describing what they would check rather than checking it. A work sample settles that, in either language.

See how it works >