Interviewing
Can You Take an AI Interview in Your First Language?
You can sometimes take an AI interview in your first language. Whether an interview is held in another language, and whether the answers are judged in that language, are two separate questions, and a platform's language count answers neither. Read the invitation, then ask the recruiter which language the answers will be judged in before you start recording. If the job genuinely requires spoken English, answering in English is the signal that matches the work.
The takeThe advice most people meet here is to practise their English harder, which answers a question nobody asked. The open question is which language the employer will read the answers in, and that is a fact a recruiter holds and can give you in one line. Sending that message costs a minute. Preparing for three weeks in the wrong language costs considerably more.
Where Olive fits
Open a role and see what the work shows
An Olive assessment is not an interview, and nothing drawn from how you sound enters its findings: you work a real assignment openly with an AI assistant, think aloud in speech or in writing, and a person writes or edits six findings on what happened. You are granted the same report the employer reads, free.
Assess your shortlistWhich Language Will the Platform Take?
The invitation and the recruiter can tell you, and a vendor's language count cannot. At least one vendor advertises a count of supported languages, and a count is not the same fact as which language your answers get read in. Alex, one live-interviewer product, claims support for 30 languages on its own About page, next to a cumulative count of a million candidates interviewed by phone, video or text 1. Both numbers are the vendor's own and unaudited.
Ask which language the answers get read in, not how many the menu lists.
The one question you are asking is really two. The first is whether the session itself runs in your language. The second is whether the answers are evaluated in that language or in a translation of it. A count of supported languages does not separate those two, so the invitation email and the platform's own candidate help pages are the first places to look.
Hiring teams meet a version of the language question from the other chair. How a hiring manager judges communication after an AI-translated interview is written for them, for the case where a candidate answered through a live translation layer, and it shows what an employer weighs then.
Ask Which Language the Answers Are Judged In
Send the recruiter one line, in writing, before the first recording starts: which language should the answers be in, and is that the language they will be judged in. That is two questions because accepting an answer in Portuguese and evaluating it in Portuguese are separable steps, and the recruiter is the person who can find out which one applies here.
HireVue says how a candidate experiences its assessment is configurable by the company deploying it 9, so on that platform, at least, the employer is the party who knows.
What happens after you speak matters as much as the language you speak in. HireVue's 2025 explainability statement describes a three-stage pipeline: speech to text, then a language model, then a score benchmarked against expert human raters. The same document says the AI relies only on the transcribed words, and not on facial expression, background or tone of voice 7. That is a vendor describing its own product rather than an audit, and it covers HireVue's AI-scored interviews only. It also says nothing about what a human recruiter watching the same recording notices.
The first stage is the one to carry: on HireVue, an answer scored by AI is scored from a transcript, and a transcript is a machine's best guess.
The same machinery shows up when the interview is explicitly a language test. What an AI scoring your spoken English is actually measuring takes that case on its own terms, and a transcript sits underneath both.
The language you answer in and the transcript that reaches the reviewer are two different facts, and only one of them is yours to control.
How Much Worse Will a Recognizer Hear You?
Measurably worse on average, unevenly, and not by a number you can look up for yourself. DiChristofano and colleagues put recordings of 2,713 speakers from the Speech Accent Archive, all reading the same paragraph, through the Amazon, Google and Microsoft speech-to-text services in 2021 2. Word information lost was 0.14 lower for first-language English speakers.
That measure runs from zero to one and is not word error rate, so it does not convert into a share of words the service got wrong. Regressions across the three services found the loss rising with a later age of starting English, which is nearer to the question people actually ask 2.
It is one read paragraph, many of the recordings made years before they were transcribed, on services that have been updated since.
A Dutch study points the same way in another language. Feng and colleagues tested one research recognizer on a corpus built to be difficult, and reported 26.1% word error rate for native speakers on read speech against 55.9% for non-native speakers 3. The native group there is children, teenagers and older adults rather than typical working-age speakers, so the gap is the finding and the individual figures belong to that corpus. It is Dutch, one system, and one year, which is the honest size of it.
The same study grouped its non-native adults by Common European Framework level. Read-speech error rates came out at 57.2% at A1, 56.5% at A2 and 57.0% at B1, and the authors write that they do not see a reduction in error with an increase in level 4. The groups at each level were small, and the levels were the corpus's labels rather than tested at recording.
That is not a reason to stop studying the language, but it is a reason to stop reading a machine transcript as a verdict on how well you speak.
Unrecognisable is the wrong word for any of it. Five recent systems were tested on read sentences from speakers whose first languages were Arabic, Chinese, Hindi, Korean, Spanish and Vietnamese. Whisper and AssemblyAI reached mean match error rates of 0.054 and 0.056, which the authors describe as approaching human accuracy 5. Match error rate is not word error rate, so those figures do not line up with the Dutch ones, and all six first languages are well resourced in recognizer training data.
Read sentences from a corpus built for pronunciation research are close to the easiest case available, and a spontaneous interview answer, on whatever microphone you have, is close to the hardest.
The risk is real and uneven rather than universal.
Whether an AI interview tool can score you down for your accent gathers what the record shows on the scoring side, which is where a misheard word becomes a scoring input.
Say Something If the Transcript Mangles Your Answer
Raise it with the recruiter while the process is still open. A transcript you can see and do not recognise is a concrete, checkable thing to report, and it is a different request from asking to be judged more generously. Name the question, name the words that came out wrong, and ask what happens to that answer now.
At least one platform already treats this as a defined case. HireVue's statement gives two conditions for an answer to be scored by AI at all: the candidate consented, and the system detected enough content in the response. It names speech the system cannot understand clearly enough, along with a short answer or a microphone problem, as reasons a response is instead flagged for human review 8.
A small proportion of candidates is how HireVue puts the volume, with no figure attached, and flagged for a person is not the same as marked well.
Whether a transcript that got your words wrong counts against you takes the same problem from the moment you notice it.
If the job genuinely requires spoken English, answering in English is the signal that matches the work, and it is the version of you that shows up on day one. In the US, the EEOC's 2016 enforcement guidance on national origin discrimination states that an English fluency or proficiency requirement is generally permissible only if it is required for the effective performance of the position 6. That is guidance rather than law, and it was written about stated fluency requirements, not about machine-transcribed interviews.
What it ties the bar to is the job, which is the part worth knowing if a fluency standard looks broader than the work.
Before the recording opens you can still ask which language the answers should be in, who reads them afterwards, and whether a live conversation is available instead. After you submit, most of those questions have already been answered for you.
A minute of asking beats weeks of rehearsing in the wrong language.
Common questions
Can I answer in my first language if the platform shows a language menu?
Sometimes. A vendor's list of supported languages does not say whether the questions, your answers and the evaluation all happen in the language you pick. The invitation and the platform's candidate help pages are where that lives, and the recruiter can answer what neither of them covers. Ask before the first recording rather than after it.
Does answering in English count against me if it is not my first language?
The measured risk sits in the transcript rather than in the choice. Where the role needs spoken English, answering in English is the signal that matches the job, but speech recognizers make more errors on non-native speech, so the transcript your answer becomes can be rougher than what you said, and where a platform scores the transcript, that roughness can reach the score. That is a reason to speak at your own pace and to check the transcript if you are shown one, rather than a reason to avoid the format.
Will a machine transcript judge how fluent I am?
A transcript is not a fluency test and should not be read as one. In one Dutch study, word error rates for non-native adults did not fall as their labelled proficiency level rose, which suggests that recognizer's errors did not track the speaker's fluency 4. What an employer does with the transcript is a separate question again, and one worth putting to the recruiter directly.
What should I do if the transcript got my words wrong?
Report it to the recruiter while the process is still open, naming the question and the words that came out wrong. On HireVue, speech the system cannot understand clearly enough is one of the stated reasons a response is flagged for human review rather than scored by AI, so a garbled answer is a defined case on at least one platform. Do not quietly re-record unless the employer says you can.
Is an employer allowed to require English fluency?
For US employers the EEOC has written the test down. Its 2016 enforcement guidance says an English fluency or proficiency requirement is generally permissible only if it is required for the effective performance of the position, and it flags uniform fluency bars applied across dissimilar jobs 6. That is guidance rather than law, and it was written about stated requirements rather than machine-transcribed interviews. Check the rules that apply where you are applying.
References
- 1. About Alex alex.com The vendor's own claim of 30 supported languages and a cumulative million candidates interviewed by phone, video or text, which is this article's evidence that non-English AI interviews exist as a product setting.
- 2. Global Performance Disparities Between English-Language Accents in Automatic Speech Recognition arxiv.org The 2,713-speaker measurement across Amazon, Google and Microsoft speech-to-text, its 0.14 word-information-lost gap for first-language English speakers, and the finding that loss rose with a later age of starting English.
- 3. Quantifying Bias in Automatic Speech Recognition arxiv.org The Dutch read-speech word error rates of 26.1% for native and 55.9% for non-native speakers, this article's evidence that the non-native penalty is not an English-only phenomenon.
- 4. Quantifying Bias in Automatic Speech Recognition arxiv.org Table 4's read-speech error rates by Common European Framework level and the authors' statement that they see no reduction with an increase in level, which supports the claim that a transcript is not a fluency meter.
- 5. Automatic Speech Recognition for Non-Native English: Accuracy and Disfluency Handling arxiv.org The read-speech match error rates of 0.054 and 0.056 described as approaching human accuracy, which stops this article from claiming accented speech is unrecognisable.
- 6. EEOC Enforcement Guidance on National Origin Discrimination eeoc.gov The agency's stated position that an English fluency or proficiency requirement is generally permissible only if required for the effective performance of the position.
- 7. 2025 Explainability Statement hirevue.com The three-stage pipeline of speech-to-text, language model and benchmarked score, and the vendor's statement that its AI relies only on the transcribed words, which is why this article treats the transcript as the scoring input.
- 8. 2025 Explainability Statement hirevue.com The two conditions for AI scoring and the statement that speech the system cannot understand clearly enough is routed to human review, which supports the advice to report a mangled transcript.
- 9. 2025 Explainability Statement hirevue.com The vendor's statement that how a candidate experiences the assessment is configurable by the company deploying it, which is this article's only sourced basis for telling a candidate to put the language question to the employer.
9 sources, numbered by first appearance. How Olive sources claims
General guidance for hiring teams. What works at one company and one volume may not transfer to yours.
Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.