Interviewing
Smiling for the Camera Is Aimed at a Scorer HireVue Says It Dropped
You don't have to fake a smile for an AI interview. HireVue, a large recorded-interview vendor, says it removed visual analysis from new assessments early in 2020, and its 2025 statement says the AI does not assess facial expressions, body language or tone of voice. That is the vendor's own account, not an audit. On HireVue, a performed smile is aimed at a scorer the vendor says is gone, and other vendors have made other choices. Whether anyone watches the recording is a separate question the statement leaves open.
The takeThe advice to perform is the cheapest thing a prep site can sell, because nobody can check whether it worked. It costs the most from candidates whose natural affect is flat, quiet or unusual, who end up rehearsing a face instead of an answer. Recorded rounds are an odd format and deserve the criticism they get. Rehearsing warmth is not the fix, and it thins out the answer underneath.
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 or look enters the findings: there is no camera in the product, you work a real assignment with an AI assistant, and a person writes or edits the findings in words. The candidate is granted the same report the employer reads, free.
Assess your shortlistWhere Does the 29 Percent Figure Come From?
From a 2019 complaint, quoting the vendor. On November 6, 2019 the Electronic Privacy Information Center asked the Federal Trade Commission to investigate HireVue's video assessments 1. The filing quotes the company's own chief industrial-organizational psychologist saying that Facial Action Units could make up 29 percent of a candidate's score, and its chief technology officer putting facial expressions at 10 to 30 percent 1.
A complaint is a set of allegations, not a finding by the agency 1.
The figure travels without either of its dates.
The statements EPIC quoted were made in 2019, and they describe a scoring design the company later said it had stopped using early in 2020 12.
On January 12, 2021 HireVue's chief data scientist published the decision itself: no visual analysis in the company's pre-hire algorithms going forward 2. The stated reason was that, for the significant majority of jobs, visual analysis correlated far less with job performance than the other parts of the assessment, and that adding non-verbal data to language data produced no significant additional predictive power 2. A press release the same day dates the removal to early in 2020, for all new assessments 2.
No data was shown 2.
Those are the company's characterisations of internal research nobody outside it has examined, and they cover HireVue's own assessments rather than anybody else's 2.
Prep advice absorbed the first half of that story and never the second.
What Does the Vendor Say It Scores Now?
Only the words. HireVue's 2025 Explainability Statement describes a three-stage pipeline: speech-to-text transcription, then a natural-language model reading the transcript, then a score benchmarked against expert human raters 3. The same document says the AI uses no video analysis and no other audio characteristics, and that it does not assess facial expressions, body language, background and surroundings, or tone of voice 3.
The statement is HireVue's account of HireVue, not an independent audit 3.
It covers HireVue's AI-scored interviews and nothing else, and the statement calls itself a living document, so it describes the 2025 configuration rather than whatever a candidate met three years ago 3.
None of that settles who watches the video. The statement describes what the model reads, and who opens the recording afterwards is a different question. Whether a human ever reviews an AI-scored interview works through what is and is not known about that step.
When a person does watch, the guidance pointed at hiring teams is not what prep sites assume. The article on a candidate whose eyes keep moving off-camera tells interviewers that gaze is not evidence of deception, and that acting on it risks screening on a disability.
So the smile is aimed at two readers, one that says it stopped looking and one whose habits no vendor statement describes.
Which Systems Still Read How You Look?
At least one did as recently as 2021, and the interface will not tell you which. Retorio, a German vendor, told the public broadcaster Bayerischer Rundfunk in 2021 that its algorithm is trained to measure the impression a person makes, meaning how people judge someone's personality from observation 4. Asked why appearance changed its scores, the company said those factors are taken into account for the assessment just as in a normal job interview 4.
By the company's account, the model was trained on videos of more than 12,000 people, with a further 2,500 people rating how they perceived each one on the Big Five dimensions 4. Those ratings are the target the model learns to reproduce.
A score built that way predicts how you came across to a panel of strangers, including whatever the panel reacted to.
Those are the company's statements about itself, relayed by a broadcaster rather than established by an audit, and they describe one vendor as it stood in 2021 4. Other tools train on other targets, and the screen you record into will not say which.
In the EU there is now a rule pointed at this. The AI Act prohibits inferring emotions in the workplace, and the European Commission's guidelines of 29 July 2025 read 'workplace' as reaching candidates during selection and hiring, stating that using emotion recognition during the recruitment process is prohibited 5.
Guidelines interpret the law rather than settle it, and only the Court of Justice can give an authoritative reading 5.
The prohibition bites on inferring emotions, so a vendor whose model scores transcribed words will argue it sits outside 5. It is an EU rule only.
You do not have to work this out from the interface. What an AI interview scores and what it does not covers two US disclosure laws and what each one actually requires.
Say the Specific Thing and Let Your Face Do What It Does
Spend the preparation on content, because a transcript scorer and a person watching both read it. Pull the competencies out of the posting, write one real example for each, and name what you actually did: the decision, the constraint, the thing that went wrong. A transcript has something to grade there. A rehearsed smile leaves the same generic answer underneath it.
The format is doing more work than the performance advice admits. A 2025 meta-analysis of 21 articles found that non-autistic observers formed significantly less favourable first impressions of autistic people 6. The size of the penalty tracked the medium: largest when observers both saw and heard the person, and on a transcript alone small enough that it could not be told apart from zero 6.
Measured as an effect size, the penalty was 0.63 on audio and video and 0.22 on a transcript alone, with a confidence interval crossing zero 6.
Those were short lab clips and thin-slice rating tasks 6.
Not one included study measured a real hiring decision, the transcript subgroup is the smallest with a wide interval, and a null result is not evidence that text is fair 6.
What it supports is narrower and still useful: the medium decides how much of the judgment lands on presentation.
Hiring teams are being told something adjacent about the format itself. The article on whether the async video round is worth keeping when answers are scripted tells employers the round separates people on delivery, which matters for a job spent talking to strangers and hardly at all otherwise.
If your worry is about accent or disability rather than expression, what the record shows about AI interviews, accents and disability collects the evidence on that ground.
Ask what the tool evaluates before you record, then answer the question that was asked.
Common questions
Does the AI score my facial expression?
For HireVue, the vendor says no. Its 2025 Explainability Statement says the AI scores only the transcribed words and does not assess facial expressions, body language, background or tone of voice. That is the company describing its own product rather than an independent audit, and it covers HireVue's AI-scored interviews only. Other vendors have made other choices, and the screen you record into will not tell you which one you are using.
Where does the claim that 29 percent of the score is your face come from?
A 2019 complaint the Electronic Privacy Information Center filed with the US Federal Trade Commission, quoting HireVue's own chief industrial-organizational psychologist. A complaint is a set of allegations, not an agency finding. The statements it quotes were made in 2019 and describe a scoring design HireVue later said it had stopped using early in 2020, when it removed visual analysis from new assessments. The figure keeps circulating with neither date attached.
Will a person watch my recording?
Scoring and watching are two separate questions, and a vendor statement about scoring does not answer the second. The vendor's published scoring description says what the model reads, not how often a recruiter opens the recording. The recording still exists and somebody can open it. Ask the employer who reviews recorded answers and at which stage, which is a reasonable thing to ask before you record one.
Is it worth smiling anyway?
Smile if that is what your face does. The case against forcing it is not that warmth gets punished, it is that holding an expression for half an hour takes attention from the answer, and the answer is the part a transcript scorer and a person watching both read. If a natural smile arrives, nothing is lost. If it never arrives, HireVue's published description says its scorer does not look at faces at all.
What if my natural expression is flat or unusual?
First impressions formed from video may weigh on you more than a transcript would, and performing is not the repair. A 2025 meta-analysis found first impressions of autistic people were markedly less favourable on audio and video than on a transcript, where the difference could not be told apart from zero. Those were lab studies rather than hiring decisions. If the recorded round is a barrier for you because of a disability or a medical condition, asking for a written or untimed alternative is a request an employer can act on.
References
- 1. In the Matter of HireVue, Inc.: Complaint and Request for Investigation, Injunction, and Other Relief epic.org Supports the provenance and date of the 29 percent figure: 2019 HireVue statements quoted in an advocacy complaint, which is a set of allegations rather than an FTC finding.
- 2. Industry leadership: New audit results and decision on visual analysis hirevue.com Supports the dated claim that HireVue said it would use no visual analysis in its pre-hire algorithms, its stated reason, and the early 2020 removal date, all as vendor statements with no data shown.
- 3. 2025 Explainability Statement hirevue.com Supports the claim that the AI scores transcribed words through a three-stage pipeline and does not assess facial expressions, body language, surroundings or tone of voice, and that the document describes itself as living.
- 4. Objective or Biased: On the questionable use of Artificial Intelligence for job applications interaktiv.br.de Supports the claim that at least one vendor said its model is trained to reproduce how human raters perceived people on video, with the training numbers as the company's own account.
- 5. Commission Guidelines on prohibited artificial intelligence practices established by Regulation (EU) 2024/1689 (AI Act), C(2025) 5052 final ai-act-service-desk.ec.europa.eu Supports the claim that the Commission reads the workplace emotion-recognition ban as covering candidates during recruitment, as interpretive guidance rather than settled law.
- 6. First Impressions Towards Autistic People: A Systematic Review and Meta-Analysis pmc.ncbi.nlm.nih.gov Supports the claim that the presentation medium changes how much a candidate is judged on presentation, with the audio-video penalty largest and the transcript effect indistinguishable from zero.
6 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.