Interviewing With an Accent: What the New AI Scoring Research Means for You

Say a one-way interview link lands in your inbox. You record an answer about a missed deadline, you know it’s a good answer, and every word of it would look fine written down. Now imagine the software scoring it decides you mumbled.

Researchers at the University of Washington found something close to that when they played leading AI models the same interview answers in different accents. UW published the results on November 4. If you interview in accented English, whether it’s your second language or your first, the study gives you a concrete reason to ask how you’re being scored. It also undercuts some advice you’ve probably been given.

The same words scored lower in some accents

The team’s paper, presented at the EMNLP conference in October, modeled its hiring test on asynchronous AI interviews, where you record answers to fixed prompts and a system scores them before anyone shortlists. Its main analysis covered Gemini 2.5 Pro, Gemini 2.5 Flash, GPT-Audio-1.5, and Qwen3-Omni, each rating delivery and content from 1 to 7. Scripts were identical across American, British, Indian, Chinese, and Nigerian accents, so the words couldn’t explain a gap.

American and British speakers generally ranked highest, Indian speakers landed in the middle, and Chinese and Nigerian speakers came last. Two details matter more than the ranking. The models repeatedly said Chinese and Nigerian speakers “mumbled” or were “difficult to understand” while transcribing that speech very well, and the paper found transcription errors didn’t explain the lower delivery scores. The models also didn’t reliably follow instructions to disregard delivery when judging content, so content scores could slide along with it.

This isn’t a test of English ability, either. English is the official language in Nigeria and the teaching language in Indian schools, as UW points out, and both accents still scored below American and British.

The limits are real: the team tested general-purpose models, not a named hiring product, and says real deployments may differ from its setup. Its human check used 52 speakers, only four of them British, and the main comparisons use scripted speech. So nobody should claim this describes the tool your next employer uses. But AI interviews are no longer rare: Greenhouse’s 2026 report says 63% of US job seekers have experienced one, and 70% of those evaluated by AI said it wasn’t clearly disclosed beforehand.

Plain mishearing is a separate risk. Information Age reported in 2025, covering University of Melbourne research, that HireVue reported transcription error rates under 10% for US English speakers and up to 22% for non-native speakers with accents from other countries. A system can mishear you, or hear you perfectly and still mark you down.

Human interviewers have this habit too

You may have read that candidates overestimate how much their accent affects interviewers, and that most people only register it when they can’t follow you. The research is less comforting. A meta-analysis in Personality and Social Psychology Bulletin pooled 139 effect sizes covering 4,576 participants and found standard-accented candidates were rated more hireable, with an effect size of d = 0.47. The gap was larger for communication-heavy jobs, for women, and for foreign rather than regional accents.

The most useful finding is the one that sounds worst. How understandable a candidate was relative to others did not significantly change the size of the bias. Being easier to follow didn’t reliably buy your way out of it.

That should change where you spend your money. Skip accent reduction courses sold as a quick hiring fix. The UW models did score voices lower the further they sat from American English, but that isn’t a distance you’ll close in the weeks before an interview. Work on being easy to follow, since that helps your answers land with any listener, and stop there.

There’s one more place to push. A Psychological Science study of 5,367 TED Talks found less engagement for speakers with nonnative accents, and in its experiments the same speaker with the same content was rated less warm and less friendly with a Persian accent than with an American one. That’s one accent, far from hiring, so treat it as a hint. Warmth is still something you can act on in the first minute. Use the interviewer’s name and respond to something they actually said before you launch into your answer.

Ask how you’ll be scored before you record

When a one-way link arrives, ask about it. You won’t sound strange: Greenhouse found 46% of US candidates want the option to request a human interview. Something like this works:

“Thanks for the invitation. Before I record, could you tell me whether these answers are scored by AI or reviewed by a person? I speak English with an accent and want to be judged on what I say. If AI scores them, could we do a short live call instead?”

A high-volume employer will probably say no, and you record anyway, having lost nothing. At a smaller company, a recruiter may simply call you. Some places also give you notice rights, covered in our guide to AI hiring laws and candidate rights. If a disability affects your speech, say so and ask for an accommodation in those words.

Give a one-way interview less to misread

You usually won’t know whether a system is reading a transcript, listening to your voice, or both. Record a test clip in the room and on the microphone you’ll use, and listen back, because worse audio gives transcription more chances to go wrong.

Then build your answers for a reader. Put your main point in the first sentence and use the exact terms from the job posting, because if a transcript is what gets scored, those words have to be there. Say the full name of a tool or method once before you shorten it. Slow down for the first sentence of each answer in particular, when a listener has the least context, and let the rest run at your natural pace.

None of this repairs a biased scorer. It only removes the avoidable reasons to lose points.

In a live interview, ask instead of guessing

The most expensive mistake is answering a question you half understood. Checking takes five seconds, and a confident five minutes on the wrong question can cost you the round. Learn these until they come out without effort:

• When the wording tripped you up: “Could you rephrase that? I want to answer what you’re actually asking.”
• When you caught most of it: “Just to check, you’re asking how I’d handle a late vendor when the launch date can’t move?”
• When it’s an idiom: “I’m not familiar with that expression. What do you mean by it?”
• When you need a moment: “Give me a few seconds to think about that.”
• When a word disappears: “I’m missing the English word, but it’s the mechanism that stops a queue from overflowing.”

That last one is worth practicing most, because describing a concept shows you have it, and the word may come back two sentences later.

Drop “sorry for my English,” which asks the listener to start hunting for mistakes. And don’t go quiet for long stretches while you translate in your head, because silence is hard to read on a video call. Short spoken fragments are fine, and thinking out loud is a skill you can drill.

When the feedback says “communication”

If a rejection mentions communication, ask for specifics: “Could you share an example of where my communication fell short of what the role needs?” Many companies won’t answer. The ones that do give you something to work with, because “your answers ran long” is fixable and a vague worry about clarity after a strong interview might be something else.

In the US, the EEOC’s guidance says a decision may legitimately rest on accent if it “interferes materially with job performance,” and that the employer must show both that the job requires effective spoken English and that the accent materially interferes with it. Rules differ elsewhere, and bias is hard to prove from one rejection, so keep dated notes and talk to an employment lawyer if you see a pattern.

Practice against a transcript

A friend can tell you how you sounded. A transcript shows you where your sentences actually went. Record a mock answer, read it back, and look for the places where the structure fell apart. Ignore misrecognized words, since speech recognition can stumble on accents and that error isn’t yours.

ExtraBrain is useful for this drill. Its live transcription captures your practice answer, and the local Parakeet option keeps the speech-to-text step on your computer. You can quote a passage from the transcript into chat and ask which sentence is hardest to follow, then cut it down in your own words. Import your resume and project notes as reference documents and your behavioral interview practice will pull from your real examples.

Use it in an actual interview only where the rules allow AI assistance. On calls after you land the job, ExtraBrain’s session summaries list key points, decisions, and follow-ups, which helps when most of your attention went to following the English. Get consent first where the law or the other people require it.

Look at the next step on your calendar. If it’s a one-way link, send the recruiter that question today. If it’s a live call, say “Could you rephrase that?” out loud until it feels as ordinary as your own name.