You Are Being Screened by AI: What Changes

Quick Answer

• Three layers, not one: the parse (resume screening in an applicant tracking system), the recording (asynchronous video, sometimes scored automatically), and the conversation (a two-way AI interviewer that follows up)
• Each reads less than you think: fields and filters; your transcript; your answers and what they imply
• Common, and often undisclosed. 63% of job seekers surveyed by Greenhouse say they have been interviewed by an AI; 70% say they were never clearly told upfront
• Facial analysis is mostly legacy advice. HireVue says it scores the transcript and does not use facial analysis, video or audio data to evaluate candidates
• What you are owed depends on where the job is based — Illinois, New York City and the EU require different things, and one of those rules moved in 2026
• Preparation barely changes. Specific evidence, pulled from the posting, said out loud. The machines just remove your ability to wing it

The Three Layers, and What Each Reads

"AI interview" gets attached to everything from a keyword filter to a voice agent that asks follow-up questions. Lumping them together is why so much advice on this subject is useless: tactics that help at one layer are irrelevant or counterproductive at another.

It is also worth knowing how much of hiring this covers, because the category is smaller than the headlines suggest. In SHRM's survey of 1,908 HR professionals, 39% of organizations had implemented AI anywhere in their HR function, with another 7% planning to that year; where it was in use, recruiting was the most common application at 27%. Plenty of the employers you apply to are doing none of this.

Source: SHRM, The State of AI in HR 2026, published 31 March 2026, based on 1,908 HR professionals.

Layer One: The Parse

Before anything resembling an interview, your resume is turned into fields: titles, dates, employers, named tools, education, location. A recruiter's filters then run over those fields — years in a function, a certification, a specific platform — and produce a ranked or shortened list.

This is the layer with the most mythology attached, most of it percentages nobody can source. What is documented is narrower and more useful. In the Harvard Business School and Accenture study Hidden Workers: Untapped Talent, more than 90% of surveyed employers using a recruiting management system relied on it to make a first cut or rank applicants — 94% for middle-skills roles, 92% for high-skills. In the same survey, 88% agreed that qualified high-skills candidates get vetted out because they do not match the exact criteria in the job description; for middle-skills workers it was 94%.

Source: Joseph Fuller et al., Hidden Workers: Untapped Talent, Harvard Business School and Accenture, September 2021 — a survey of 2,275 executives and 8,720 workers in the US, UK and Germany. Figures reported by HBS Working Knowledge, 5 November 2021, and The Register, 8 September 2021. The study is five years old; treat it as evidence about how these filters are used, not as a current rejection rate.

That second pair is employers describing their own process. The filter is not forming a judgment about you; it is matching text, and the people who set the criteria largely know it discards people who could do the work. So say the thing, in the words the posting used. If the requirement is "experience owning a renewal number" and your resume says "responsible for book of business performance," those are the same job and the filter cannot tell.

Most silent rejections happen here, before any interview. If you are applying steadily and hearing nothing, the problem is usually at this layer rather than in your answers — our guide to why you are not getting interviews works through the diagnosis.

Layer Two: The Recording

Then comes the link: prompts on screen, a countdown, a record button, nobody on the other end. Sometimes a recruiter later watches it. Sometimes an algorithm scores it first and ranks you for review. You usually cannot tell which, because it is an employer setting.

The correction worth making here is about what gets scored. HireVue — one of the largest vendors in the category, and unusually explicit about its methods — states on its own site that in its video assessments "we analyze only the transcript of your responses, focusing on how you describe your experiences and past actions related to key job competencies," and, flatly, that its "video assessments don't use facial analysis, video, or audio data to evaluate candidates." It adds that humans make the final hiring decision. That was a deliberate change with published reasoning: HireVue stopped using facial analysis in 2020, announcing it in January 2021, and its chief data scientist told Fortune that nonverbal data contributed roughly 0.25% of a model's predictive power in most cases, about 4% for customer-facing roles.

Sources: HireVue, "AI in Hiring" and its candidate FAQ, both accessed September 2026 — the FAQ is also where HireVue states that an assessment "considers the whole context of a response, not just certain words here or there," and that accommodations can be requested from the recruiter; Jeremy Kahn, "HireVue stops using facial expressions to assess job candidates", Fortune, 19 January 2021. These are HireVue's statements about HireVue's own models; other platforms score differently, and many record without scoring at all.

So the hour you were going to spend on lighting and eye contact is misallocated. Where anything is scored automatically, what is scored is the words. Be audible, be framed, then spend the time deciding which sentences you intend to say. The mechanics that genuinely vary — think time, answer length, whether you get a retake — are per-question employer settings, which is why our one-way video interview guide treats the on-screen instruction as the only truth, and our HireVue guide maps prompts to competencies rather than pretending a fixed question list exists.

Layer Three: The Conversation

The newest layer talks back. On 18 June 2026 HireVue launched AI Interviewer, which it describes as conducting "dynamic, two-way voice interviews with every applicant 24/7," with responses evaluated against what it calls IO-validated rubrics built from thousands of outcome studies. Several vendors now ship something similar, so a growing number of candidates meet a conversation rather than a record button.

Source: HireVue press release via PR Newswire, HireVue AI Interviewer Sets New Enterprise Standard, 18 June 2026.

The word doing the work there is "dynamic." A memorized paragraph survives a one-way recording; it does not survive a two-way interview, because the next question is generated from the answer you just gave. Claim you turned around a churning account and it asks what you actually offered the customer. Give a vague number and it asks how you measured it. The failure mode is not nerves — it is an answer only one layer deep.

Candidates are not experiencing this evenly, and the evidence cuts both ways. Greenhouse, an applicant-tracking vendor, surveyed 2,950 active job seekers across five countries: 63% had been interviewed by an AI, up 13 percentage points in six months, and 38% had walked away from a process because it included one. Bullhorn, a staffing-software vendor, surveyed nearly 2,300 professionals who had used a recruitment firm and reports 92% of those interviewed by an AI voice agent rated it as good as or better than a live recruiter interview.

Sources: Greenhouse, 2026 Candidate AI Interview Report, released 1 May 2026 (2,950 active job seekers, five countries); Bullhorn, 2026 GRID Talent Trends Report, released 23 September 2026 (nearly 2,300 professionals who had engaged with staffing firms). Both are vendor research and both have a commercial interest in the answer; the samples are also different populations, which is part of why the findings diverge.

Less contradictory than it looks: disclosed in advance, at a convenient hour, with a reply afterwards, an AI interview beats a recruiter who never called back. Sprung on you mid-process, it is worse than nothing. Which is exactly the part now being regulated.

One Posting, Three Layers

Here is the same job description passing through all three, so you can see how little the source material changes.

The parse takes the literal requirements: three years, the function names, Salesforce, a BI tool, the location. Nothing here is a judgment. If your resume says you "managed client relationships" rather than owned a renewal number, you have described the same work in vocabulary the filter was not given.

The recording derives prompts from the competencies behind those responsibilities. "Own renewals and expansion" implies account management under commercial pressure: "Tell me about a time an account was about to churn. What did you do, and what happened?" "Quarterly business reviews with economic buyers" implies communicating upward: "Describe how you have presented results to a senior stakeholder who disagreed with you."

The conversation opens with the same question, then does something a recording cannot.

Both follow-ups take one sentence if you were there, and are unanswerable from a memorized script. That is the whole of the difference the third layer makes. Pulling the questions out of a posting in the first place is the method in our guide to preparing from the job description — by hand it is an evening per application, and almost all of that evening should go on the answers.

What You Are Entitled to Be Told

The transparency gap is the real complaint in the data: 70% of Greenhouse's respondents said they were never clearly told upfront that AI would be evaluating them, and 21% found out only once the interview had started. Some jurisdictions now place duties on employers about exactly that. What follows is what those rules require of employers, with sources — not legal advice, and coverage generally follows where the job is based rather than where you are sitting.

Illinois has had a dedicated statute since 1 January 2020. Under the Artificial Intelligence Video Interview Act, an employer that asks applicants to record video interviews and uses AI analysis of them must, for positions based in Illinois, notify each applicant beforehand that AI may be used, explain how the AI works and what general types of characteristics it uses to evaluate applicants, and obtain consent — and may not evaluate applicants who have not consented. On request it must delete the interview within 30 days and instruct anyone holding a copy to do the same.

Separately, an amendment to the Illinois Human Rights Act took effect on 1 January 2026, barring employers from using AI that has the effect of discriminating on the basis of protected classes in recruitment, hiring and promotion, banning zip codes as a proxy for a protected class, and requiring notice when AI is used in those decisions. The rules meant to clarify the timing and means of that notice were still pending as the obligation took effect, so the mechanics are unsettled even though the duty applies.

Sources: Illinois Artificial Intelligence Video Interview Act, 820 ILCS 42 (in effect 1 January 2020) — §5 and §15 texts via FindLaw; Illinois HB 3773, amending the Illinois Human Rights Act effective 1 January 2026, as summarized by Morgan Lewis, 12 September 2024, and on the pending rulemaking by Reinhart Boerner Van Deuren, 18 December 2025.

New York City takes a different route. Local Law 144 of 2021 bars employers and employment agencies from using an automated employment decision tool unless it has had a bias audit within the previous year, the results are publicly available, and candidates get notice — which the city's guidance sets at ten business days before the tool is used. Enforcement began on 5 July 2023. In practice that means a published audit summary you can go and read before you interview.

Source: NYC Department of Consumer and Worker Protection, Automated Employment Decision Tools, accessed September 2026.

The EU is the one most often described wrongly, because the timetable moved this year. The AI Act's Annex III does classify AI used for recruitment, selection, candidate evaluation and targeted job advertising as high-risk — but the obligations attaching to stand-alone Annex III systems, including the transparency and human-oversight duties candidates have heard about, were deferred by the 2026 Digital Omnibus from 2 August 2026 to 2 December 2027. They are not in force yet.

Two things already are. Since 2 February 2025 the Act has prohibited AI used to infer emotions in the workplace, except for medical or safety reasons. And from 2 August 2026 its transparency provisions apply: AI systems intended to interact directly with people must be built so those people are informed they are dealing with an AI, and deployers of emotion-recognition or biometric-categorisation systems must tell the people exposed to them. For an EU role interviewed by a voice agent, being told it is a machine is the part that already bites.

Sources: EU AI Act Article 5(1)(f) (prohibitions applicable from 2 February 2025) and Article 50 (transparency, applicable from 2 August 2026), texts via the EU AI Act Explorer, whose implementation timeline, accessed September 2026, gives 2 December 2027 for Annex III. On the deferral itself: the Digital Omnibus on AI, Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026 and in force from 27 July 2026, per the Cloud Security Alliance, 1 August 2026; the pre-adoption detail is in Gibson Dunn, 27 May 2026, written while formal adoption was still pending.

What None of This Changes

Strip out the mechanics and all three layers ask what a competent human interviewer asks: can you do this specific job, and what is the evidence. The scoring is built from the requirements, and the requirements are in the posting.

• Specificity still beats fluency. A named situation, an action that was yours, a number you can explain — which is what a transcript score reads and what a follow-up tests, and why a well-formed STAR answer travels across all three layers.
• Depth is the new requirement. Every example needs a second and third layer: how you knew, what failed, what you would do differently.
• The first sentence carries more weight. Nobody is nodding you along or asking you to clarify. Lead with the point.
• Vocabulary matters at the parse and nowhere else. Mirror the posting's terms on your resume; do not salt your spoken answers with them.
• Reading from a second screen still fails — a flat, scanning delivery in a recording, and a collapse at the first unscripted question in a conversation.

What genuinely got harder is improvisation. A human interviewer helps an underprepared good candidate: rephrases, follows a tangent, gives you a second run. None of these systems will. That is the change, and it is a preparation problem rather than a technology problem.

Frequently Asked Questions

Next Steps

Work out which layer you are in, then prepare for that one. Note what is not on the list below: nothing about lighting, nothing about eye contact. Those generate the most advice on this subject and change your result the least.