Skip to content

Interviews

Can AI Video Interview Tools Really Judge Your Personality?

Can an AI video interview read your personality from your face? Here's what the science, vendors, and 2026 law actually say before your next recording.

ResumeFlex guidesIdeas for your next move
PublishedUpdated 13 min read10 sections
Share article
Share on XShare on LinkedIn

On this page10 sections

Key takeaways

  • HireVue stopped scoring faces in 2021 after its own review found nonverbal cues added about 0.25% of predictive power (Fortune, January 2021).
  • A 2019 review in Psychological Science in the Public Interest found facial movements are not reliable, specific signals of emotion.
  • Leading AI models scored identical interview answers lower when delivered in Chinese and Nigerian accents (University of Washington, November 2026).
  • Most AI video interview scoring runs on your transcript (the words you say) not your face (HireVue; SHRM).
  • The EU AI Act has banned emotion recognition in workplaces, including recruitment, since February 2025.

Not from your face, and not from your voice. The companies that once promised to read personality off a raised eyebrow have mostly retreated, because the science behind those claims didn’t survive review. What remains is something much narrower, software that mostly scores the words you say.

If you’ve been asked to complete an AI video interview, that’s both reassuring and a little frustrating. Reassuring, because your nervous habits and your resting face aren’t being decoded into a personality profile. Frustrating, because nobody sends a calendar invite labeled “transcript analysis,” and most candidates still sit down convinced the camera is watching them like a polygraph.

Before recording an interview, ask what the tool measures, how the employer uses the result, and what alternatives are available. The research and vendor policies below help you evaluate those answers.

What an AI video interview tool actually scores today

Modern AI video interview systems score your transcript far more than your face. Automated video interview (AVI) scoring means software converts your recorded answers into text, then compares that text against a job-specific competency model. The full pipeline, from recording to recruiter dashboard, is what we cover in our breakdown of AI-assessed interviews and their limits.

The biggest name in the space made this shift loudly. HireVue stopped using facial analysis in its assessments in March 2020 and confirmed the removal publicly in January 2021, as SHRM reported in February 2021. Its published AI principles now describe its models as “analyzing only the content of a response,” with recruiters and hiring managers making the final call from a ranked list.

The reason was blunt arithmetic, not public relations. Chief data scientist Lindsey Zuloaga told Fortune that nonverbal cues contributed roughly 0.25% of predictive power in most roles, and about 4% even in customer-facing jobs, while CEO Kevin Parker concluded it “wasn’t worth the incremental value” given the concern it caused candidates (Fortune, January 19, 2021). Zuloaga also told SHRM the system was “not analyzing their accent or their diction”, only the transcribed answers.

Incremental validity is the extra predictive punch a feature adds beyond what a system already has, and facial analysis had essentially none. So the “personality from your face” era is mostly over at major vendors, and not because a law forced it. The transcript is what’s left, and the personality claims now ride on it.

Can personality be read from faces and voices at all? Mostly no

The core scientific claim behind facial personality scoring fails its own tests. In 2019, five leading emotion researchers (Lisa Feldman Barrett, Ralph Adolphs, Stacy Marsella, Aleix Martinez, and Seth Pollak) published a 68-page review in Psychological Science in the Public Interest asking whether facial movements reliably and specifically signal emotions. Their verdict, after decades of evidence: reliability is weak, specificity is largely unknown, and context does most of the work (Barrett et al., 2019).

“It is not possible to confidently infer happiness from a smile, anger from a scowl, or sadness from a frown, as much of current technology tries to do when applying what are mistakenly believed to be the scientific facts.”, Barrett et al., Psychological Science in the Public Interest, 2019

The numbers behind that quote are worth seeing. Across 76 studies and 4,487 participants, the average correlation between a facial configuration and its hypothesized emotion was about .31 (which the authors classify as weak) and the “expected” face appeared during the matching emotional episode only about 22% of the time. A scowl can be anger, concentration, effort, bright sun, or indigestion, and the review found no unique mapping in either direction.

Voice runs into the same context problem, and the industry’s biggest players have admitted it. Microsoft retired facial analysis features that “purport to infer emotional states” in June 2022, citing the “lack of consensus on a definition of ‘emotions’” and the “inability to generalize the linkage between facial expression and emotional state” across use cases and demographics (Microsoft, June 21, 2022). Meta had already shut down Facebook’s Face Recognition system in November 2021 and deleted more than a billion face templates, calling it a “company-wide move to limit the use of facial recognition” (Meta, November 2, 2021).

None of this means faces and voices carry zero information. It means the mapping is too loose and too context-dependent to anchor a hiring decision. That’s the science; here’s what happened when a vendor actually measured its own cues.

What the research does support, and how thin it is

Automated video interview personality assessment has been tested, and the results are mixed at best. The most-cited study, Hickman et al. (2022), trained machine learning models on mock interviews to predict Big Five traits, and the SIOP research brief summarizes what they found: models trained on trained observers’ impressions of candidates did better than models trained on candidates’ own self-reports, and the scores correlated with outcomes like GPA in the expected directions.

Look inside the models, though, and the story changes. The strongest cues were verbal (response length, longer words, fewer filler assents like “okay”) while smiles, volume, and speech rate carried comparatively little weight. The study’s own caveats matter just as much: it never tested whether AVI personality scores predict actual job performance, and it didn’t examine bias against underrepresented groups.

There’s a second, deeper problem: the target itself is soft. Personality tests are modest predictors of job performance to begin with, and the published estimates look flattered, Kepes and colleagues’ 2015 publication-bias analysis found conscientiousness validity is overestimated by roughly 30% in the journal literature, with bias-adjusted correlations closer to .12–.13 (Kepes et al., 2015, in PMC). So even a perfect cue reader would be predicting a weak signal.

That combination (thin cues aimed at a soft target) is exactly why “AI personality assessment” reads as marketing shorthand rather than measurement. A tool can be perfectly consistent without being right. Consistency at scale is what these systems genuinely deliver; accuracy is a separate claim nobody has earned yet.

Claim by claim: what tools say versus what science and regulators accept

The gap between vendor language and evidence is easiest to see side by side.

What a tool may claim to measure What the science supports Where the law stands in 2026
“Personality” or “enthusiasm” from facial expressions No. Facial movements show weak, unspecific links to emotion (average r ≈ .31; expected face appears ~22% of the time) Banned in EU workplaces, including recruitment, since February 2, 2025; dropped by HireVue (2021) and Microsoft (2022)
Traits or affect from voice tone Barely. Nonverbal and paraverbal cues added about 0.25% of predictive power at the largest vendor EU’s Article 5 ban covers emotion inferred from biometric data; Illinois requires notice, explanation, and consent
Traits from word choice and answer structure Partly. Verbal cues carried most of the measured signal in Hickman et al. (2022), but validity against job performance is untested Auditable under NYC Local Law 144; Illinois’ 2026 rules demand disclosure
Honesty or attention from eye contact and gaze No credible public evidence for personality claims from gaze alone No outright US ban; the EU emotion-inference ban can apply; NYC notice duties still cover the tool
Job fit from competency-aligned answers Defensible. This is standard, auditable selection practice, though audit quality varies widely NYC requires an independent bias audit; the EU’s high-risk regime for employment AI starts December 2, 2027

Read that middle column honestly and the picture is consistent: the only cell with real evidentiary legs is the last one, and it’s not personality. It’s content.

Where bias creeps in: accents, style, and everyone outside the average

The “fairest possible scoring” argument collapses when you test real voices. In a University of Washington study presented at EMNLP in Budapest on October 25, 2026, four leading AI models (Google’s Gemini 2.5 Pro and 2.5 Flash, OpenAI’s GPT-Audio 1.5, and open-source Qwen 3 Omni) scored identical interview scripts differently depending on accent, ranking Chinese- and Nigerian-accented speakers lowest for both delivery and content (UW News, November 4, 2026).

The details sting. The models transcribed the accented speech accurately, then claimed those speakers “mumbled” anyway, an implicit hierarchy of who “sounds professional.” Human raters showed the same pattern, and wider. Co-lead author Orevaoghene Ahia’s framing: people now trust audio AI as third-party evaluators in contexts “where fairness really matters.”

Accent is the measured example; style mismatch is the broader risk. Senior author Yulia Tsvetkov warned that high average scores can hide groups scoring at 20%, and the same warning applies to any candidate whose delivery sits outside the training data, pace, energy, a neurodivergent communication style included. The accent evidence underneath that risk is measured, replicated across 52 human raters, and hard to wave away.

Candidates sense the problem. Pew Research Center’s survey of 11,004 U.S. adults, published April 20, 2023, found 70% oppose employers analyzing facial expressions and 73% believe workplace face recognition would misread expressions (Pew Research Center). A striking 66% said they wouldn’t even want to apply to an employer that used AI in hiring decisions.

What the law actually says in 2026

Regulators have moved faster on this technology than the science has. For a US job seeker, three layers stack up, and the EU adds a stricter one.

  • Illinois has regulated this since January 1, 2020. The Artificial Intelligence Video Interview Act requires employers to notify you before the interview, explain in writing how the AI works and what traits it analyzes, obtain your consent before AI review, and delete the video within 30 days if you ask (Ogletree Deakins’ analysis).
  • Illinois tightened the screws from January 1, 2026. HB 3773 makes it a Human Rights Act violation to use AI in employment decisions in ways that discriminate, and the state’s draft rules explicitly cover “video interview analysis of expressions/voice” (Hinshaw & Culbertson, 2026 employer guidance).
  • New York City audits the tools themselves. Local Law 144, enforced since July 5, 2023, bars automated hiring tools unless an independent auditor checked them for bias within the past year, and the notice you get 10 business days before use must explain how to request an alternative process or accommodation (NYC Department of Consumer and Worker Protection).

The EU went further. Article 5 of the AI Act bans AI systems that infer emotions in workplaces and educational settings (recruitment explicitly included) with narrow medical and safety exceptions, and violations risk fines up to €35 million or 7% of global turnover (EU AI Act prohibited-practices tracker). Since February 2, 2025, “we score your enthusiasm from your face” isn’t just bad science in the EU; it’s an illegal practice, and broader high-risk obligations for employment AI now begin December 2, 2027, per the implementation timeline as updated in August 2026.

Notice the pattern. US law mostly demands disclosure, audits, and human escape hatches, while the EU bans the face-and-emotion layer outright. Neither regime certifies that the remaining transcript-based scoring is valid; that gap is yours to probe and the employer’s to defend.

What this means for you before you record

Your words are the assessment; everything else is mostly theater. Some practical adjustments, ordered by how much they matter:

  • Write for the transcript. Structure, specific examples, and concrete numbers are the scored signal; a rambling answer hurts you far more than a tight smile ever could. Why employers use one-way interviews explains what they’re hoping your answers will show.
  • Audio beats lighting. A decent mic and a quiet room protect your transcript; studio lighting is optional once facial scoring is out of the pipeline. If a specific tool does claim to score expressions, that’s a question worth asking the recruiter about.
  • Nervousness is mostly safe. In transcript-based systems, a bouncing leg or stiff delivery doesn’t register as a trait, and where a vendor scores paraverbal signals, the largest vendor’s own numbers show they added almost nothing.
  • Retakes are the employer’s setting, not yours. Many on-demand platforms allow one or two takes per question; don’t burn one on a stumble that doesn’t change your meaning.
  • Use your legal levers. In Illinois you can decline AI analysis; in NYC you can request an alternative process or accommodation; everywhere, asking “what gets scored?” is a question disclosure laws increasingly force employers to answer, not a favor they’re doing you.

If the interview itself starts asking for things it shouldn’t (family plans, health details, your religion) that’s a separate problem, and our guide to interview questions that cross the line covers how to respond.

The verdict

AI video interview tools judge your personality about as well as a stranger judges your handwriting: some signal in the words, none in the face, and no proven link to how you’d actually perform. The 2019 emotion-science review, HireVue’s own 0.25% finding, and Hickman’s mixed validity results all point the same direction.

The defensible version of this technology is narrow. Score how well answers match the competencies the job needs, audit it for bias, disclose it to candidates, and let a human make the call; that’s a tool. Everything beyond that is unproven, illegal in major markets, or both, so spend your prep time where the evidence says the signal lives (the content of your answers) and your questions on any employer who claims to read your soul from a webcam.

Frequently asked questions

Can an AI video interview really detect your personality?

Only weakly, and mostly through your words. Face- and voice-based claims failed scientific review, and the largest vendor’s own audit found nonverbal cues added about 0.25% of predictive power before it dropped facial analysis in 2020.

Do AI interview tools still watch your face?

The biggest US vendor doesn’t score your face anymore. HireVue stopped facial analysis in March 2020 and says its models work by “analyzing only the content of a response”, though video identity checks are still common, and smaller tools’ practices vary.

Does nervousness hurt your score in an AI video interview?

In transcript-based scoring, nerves mostly don’t register. Tools that claim to score tone or facial expressions are a different story, and worth asking about, since the evidence for those claims is thin to nonexistent.

Can you fool an AI personality assessment?

You can game the words more easily than the face. Verbal cues carry most of the measured signal, so structured, specific answers move scores, which also means the “personality” being measured is partly just how well you organize an answer.

In the US, yes, with growing disclosure and audit duties: Illinois requires notice, explanation, and consent, and New York City requires an independent bias audit. In the EU, inferring emotions from video in workplace contexts has been outright banned since February 2, 2025.

Are these tools fair to non-native English speakers?

Current evidence says not reliably enough. A University of Washington study published November 4, 2026, found leading AI models scored identical interview answers lower when delivered in Chinese- and Nigerian-accented speech, even though they transcribed it accurately.

Can you ask for a human instead of an AI review?

Often yes, and sometimes it’s your legal right. Illinois’ video interview law lets you decline AI analysis, and New York City’s pre-use notice must explain how to request an alternative process or accommodation.

Should you rehearse your answers word for word?

No. Over-rehearsed takes tend to sound flat, and humans often review what the software flags anyway. Prepare structured examples and real numbers instead of a script; that’s where the scoring signal actually lives.

Sources

Written by ResumeFlex. Questions or corrections? Contact ResumeFlex. Read our source and correction standards.

Filed under

Back to guidesBack to top

Keep reading

More in Interviews