AI interviews in 2026: what changed, and what did not

For most of the last decade, "AI interview" meant a one-way video assessment: a question on screen, a countdown, and a recording sent off to be scored. HireVue, the most widely used tool of that kind, still describes that shape in its candidate FAQ, where "most video-based assessments include between five and eight questions".

That is no longer the only option, and increasingly not the default. Two other shapes are now common:

• Structured chat. Sapia.ai runs a text-based conversational interview rather than video, and publishes an independent bias audit of it.
• Spoken conversation. A real-time voice model conducts the interview: it listens, decides what to ask next, and follows up when an answer is thin. Career1's AI interviewer works this way, in the browser, for about eight minutes per candidate.

The reason is mundane rather than visionary. Speech models became fast and cheap enough to hold a two-way conversation at hiring volumes. Once that was true, the one-way format lost its only advantage, because a fixed question list cannot ask the question that actually matters: "what exactly did you do?"

The practical consequence for a hiring team is that the two formats produce different evidence. A recording gives you the same answer from everyone. A conversation gives you each candidate tested at the point where their answer got interesting, which is harder to compare side by side and much harder to fake.

The honest boundary has moved less than the marketing suggests.

What it does well. Coverage and consistency. Every candidate gets asked about the skills the role needs, in the same structure, with the same criteria applied, and nothing is skipped because the interviewer was on their fourth call of the day. Where a report must quote the transcript as its evidence, it also produces something a hiring manager can argue with, which an unstructured gut call does not.

What it cannot do. It cannot tell you whether someone will still be good in two years, whether they will get on with your team, or what they would be like in a crisis. It cannot judge the thing a candidate did not mention. And it cannot recover a badly specified role: if the job description is vague, the interview is vague, whoever is conducting it.

The part that quietly retreated. The claim that AI can read character from a face or a voice has been pulled back by the industry itself. HireVue announced on 12 January 2021 that it had "made the decision to not use any visual analysis in our pre-hire algorithms going forward", because visual analysis "has far less correlation to job performance than other elements of our algorithmic assessment". Its candidate FAQ also states that it "never uses the video recording of your interview in order to identify you". If a vendor in 2026 sells you inferred emotion, confidence or personality from video, that is the claim to interrogate hardest.

Career1 is not affiliated with HireVue or Sapia.ai, and everything above is taken from their own published pages. On Career1, the report is written from the transcript, the resume and the job description. The video is recorded so people can watch it; it is not analysed for faces or expressions.

Four things are exactly as they were, and they are the ones that decide whether any of this works.

• A structured interview beats an unstructured one, and that is the actual mechanism. The value is asking every candidate the same well-chosen questions and judging against stated criteria. AI makes that cheap to do at volume. It does not make it true where the questions are poor.
• A person still decides. Every serious vendor positions its output as a first step. HireVue's FAQ says those "in the top group move on to person-to-person interviews before hiring decisions are made by people". Career1 ranks and recommends; it does not reject anyone.
• Candidate experience is still the constraint. A tedious, opaque or broken assessment loses you the people with options first. That has been true of every screening stage ever invented.
• Nothing fixes a bad funnel. If the role is mispriced or the description attracts the wrong people, faster screening just sorts the wrong pile more efficiently.

This is the part that genuinely changed, and it is now specific enough to plan around. None of the following is legal advice; ask counsel what applies to you.

New York City. Local Law 144 has been enforced by the Department of Consumer and Worker Protection since 5 July 2023. Its own guidance requires that an automated employment decision tool has had "a bias audit within one year of the use of the tool", that a summary of the results is published, and that candidates get notice 10 business days before the tool is used on them. It applies to jobs based in New York City, including remote jobs attached to a New York City address.

Illinois. The Artificial Intelligence Video Interview Act covers AI analysis of video interviews. Before the interview the employer must tell the applicant that AI may be used, explain how it works and what general types of characteristics it evaluates, and obtain consent; an employer may not evaluate an applicant who has not consented. Videos may not be shared except with people whose expertise or technology is necessary to evaluate the applicant, and on request they must be deleted within 30 days, including backups. An employer that relies solely on AI analysis to decide who gets an in-person interview must also report applicant race and ethnicity data annually to the Illinois Department of Commerce and Economic Opportunity.

The European Union. The AI Act treats hiring as high risk. Annex III, point 4(a), covers "AI systems intended to be used for the recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates". Point 4(b) covers promotion, termination, task allocation and performance monitoring. The timing moved: the Commission's own page on the AI Act now says the rules for high-risk uses in those sensitive areas "have been extended to 2 December 2027" following the political agreement on the AI Omnibus proposal. If you were planning against August 2026, that is the date to correct.

The questions that survive all three regimes are the same short list: what is scored, is every candidate treated the same way, can a person see the evidence behind a score, does anything reject automatically, and what happens to the recording. Career1's guide to evaluating an AI interview platform goes through them as a buying checklist.

Worth stating plainly, because a hiring team that knows this writes better invitations and gets fewer dropouts.

• To be told. That an interview is conducted or analysed by AI, and roughly what is being assessed. In some jurisdictions this is a legal requirement rather than a courtesy.
• A conversation, often, rather than a countdown. And therefore follow-up questions, which is where candidates who prepared only scripted answers come unstuck.
• That a person reads the result. No reputable process ends at the model.
• To be able to recover from a technical failure. Connections drop. HireVue's FAQ, for example, tells candidates they "will be able to pick up where you left off with the same link". Any process without a recovery path is quietly screening for good broadband.
• Not to be charged. On Career1, candidates never pay; getting vetted is free, one interview, no retakes, and the resulting profile can be hidden by the candidate at any time.

For the candidate's side of all of this, see how to pass an AI interview.

The gain is real but narrow: the first round stops being a scheduling problem, and you read evidence instead of resumes. What it does not buy you is a decision.

Before you sign anything, get written answers to these:

Swipe the table sideways to see every column.

You still need a scorecard, a human loop and a decision rule, and you still need to tell candidates what is happening. The tooling changed; the obligations did not.

Three predictions are worth discounting heavily.

"AI can assess personality or culture fit from an interview." This is the claim with the weakest track record and the one the industry has already walked back once. Treat any version of it as unproven.

"AI will replace the hiring decision." It will not, and the regulatory direction in New York City, Illinois and the EU is explicitly toward more human accountability rather than less.

"Resumes are finished." They still start almost every process, and an interview built from a resume needs a good one.

What will actually keep improving is unglamorous: better follow-up questions, evidence a reviewer can check, and less friction for the candidate. If you want to see the output rather than read about it, Career1 publishes one complete example report at sample interview report, for a fictional applicant, and the free plan vets two candidates a month without a card.