Scovai Scovai
Hiring 2026-11-07 1 min read

They're Not Rejecting the AI. They're Rejecting the Empty Chair: 38% of US Candidates Have Already Quit a Hiring Process Over an AI Interview

DSL

Dr. Sarah Liu

They're Not Rejecting the AI. They're Rejecting the Empty Chair: 38% of US Candidates Have Already Quit a Hiring Process Over an AI Interview

Thirty-eight percent of US job seekers have abandoned a hiring process at least once specifically because it included an AI interview. Another 12% say they would if required. Half your applicant pool, on the report's own reading, is either already walking away or prepared to (Greenhouse 2026 Candidate AI Interview Report).

The obvious inference is that candidates have turned against automated screening. The same dataset rules it out. Asked what they want going forward, only 19% of US candidates said they wanted less AI involvement than they already have. That one figure changes what the withdrawal number means โ€” and what it costs you to fix.

The withdrawal is not a referendum on AI

Greenhouse surveyed 2,950 active job seekers across the US, UK, Ireland, Germany and Australia. Exposure is now the norm rather than the experiment: 63% of US seekers have been interviewed by AI in the past twelve months, a 12-percentage-point rise in six months.

If withdrawal tracked AI exposure, the markets with the most AI would show the most resistance. They do not, cleanly. Germany has the highest international exposure at 57% and the highest withdrawal rate of any market at 42%. Ireland has the lowest exposure at 36% โ€” and non-disclosure is its top abandonment trigger, consistent with only 9% of Irish candidates being clearly told in advance that AI would be used.

There is a second test that separates distaste for machines from distaste for your process. Greenhouse asked candidates whether they had felt evaluated differently because of a personal characteristic, under AI and under human interviewers, and the answers are close to identical: age was flagged by 36% under AI and 36% under human interviewers, race or ethnicity by 27% under both, gender by 24% versus 26%. On appearance, candidates rated AI as less biased than people โ€” 21% against 28%. Candidates are not attributing a unique prejudice to the algorithm.

Look at what US candidates say they want instead of today's arrangement and the picture resolves further: 22% want more AI with stronger human oversight, 21% want the same level with more transparency, 19% want less AI, and only 12% want more AI on the grounds that it makes hiring faster and fairer. Across all five markets that last option never exceeds 14%. The plurality position is not retreat. It is the same amount of machine with a person attached to it.

What candidates walk away from is the unattended AI interview, not the automation

The abandonment triggers in the Greenhouse data are specific, and they are not "the company used AI."

The single largest is a pre-recorded video interview scored by AI with no human present โ€” 33%. Then: companies failing to disclose how AI would be used (27%), AI monitoring (26%), a required AI-led interview (26%), and not being able to tell whether they were talking to a human or a machine at all (20%).

Three of those five are about the absence of a person or the absence of an explanation. Only one โ€” the required AI-led interview โ€” is about automation per se.

The remedies candidates name line up exactly. The top ask in the US is the option to request a human interview instead (46%), followed by being told upfront that AI is involved (44%), a clear explanation of what the AI is measuring (39%), and knowing that a human reviews the AI's evaluation before a decision (38%). The human-interview option is the number-one comfort measure in every market surveyed without exception, from 41% in Germany to 49% in Ireland. And 59% of US candidates want a live conversation with a human recruiter as their first contact with a company; 12% prefer an AI-conducted live interview (Greenhouse, 2026).

The empty chair is the product defect. The automation is not.

The counter-argument: stated preference is cheap

A reasonable Head of Operations should push back here. Survey respondents say they value human review because it is free to say so. Revealed behaviour is a different matter, and "I would have withdrawn" is not the same as withdrawing.

That objection has been tested directly, and it does not survive. A preregistered conjoint experiment put 1,919 US job seekers through eight forced choices each between AI screening systems โ€” 30,704 profiles in total โ€” independently randomising six attributes: decision authority, error rate, explanation, opt-out, appeal, and independent bias audit (Wang, Sturgis & de Kadt, 2026).

Human involvement moved stated choice by +0.272 in probability of selecting the profile. Cutting the system's wrongful rejection rate from 30% to 10% moved it by +0.285. The gap between those two is statistically indistinguishable from zero. Put plainly: putting a person in the loop bought roughly as much candidate willingness as tripling the system's accuracy.

For comparison, the procedural features most vendors lead with came in well below that: appeal +0.156, opt-out +0.129, explanation +0.128, independent bias audit +0.068.

The authors' sharper finding is the one worth sitting with. The value candidates placed on appeal, opt-out and bias audit did not rise as wrongful rejections became more common โ€” each stayed inside a preregistered equivalence bound. Procedure is not being valued as error insurance. It is being valued for its own sake, which means improving your model's accuracy does not substitute for it.

The honest limits

The conjoint measures stated choice in a forced-choice task, not applications actually submitted. The authors themselves note that intention to apply (3.24 on a five-point scale) ran well above perceived legitimacy (2.85), and that 7.8% of respondents combined below-midpoint legitimacy with above-midpoint intention to apply โ€” people apply to processes they consider illegitimate. Application volume is therefore a poor instrument for measuring this. Greenhouse is vendor research with a commercial interest in the answer, and its US sub-sample is not broken out. Both sources point the same way; neither is a clean causal estimate of funnel loss.

Transparency repairs the candidate side. It does not repair the decision.

Here is where most responses to this data go wrong. The intuitive fix โ€” disclose what the AI evaluates โ€” works, and it works on precisely one side of the transaction.

Eight studies published in Information Systems Research, including two main experiments and a quasi-field replication, found that algorithmic evaluation of one-way video interviews raised candidate stress and deceptive impression management relative to human evaluation, and that telling candidates how the algorithm evaluates them pulled both back into line with a human-run interview (Lakhiwal, Liu, Bala & Suen, 2026).

The employer-side result is the uncomfortable one. In AI-augmented configurations โ€” human decision-makers shown the AI score โ€” evaluators anchored on that score and discounted their own judgment, even though the score failed to distinguish honest from deceptive answers. Transparency's corrective effect on the evaluator side was limited.

So disclosure buys back candidate behaviour without buying back decision quality. Those are two separate repairs requiring two separate changes, and announcing the first does nothing for the second. (The full text is paywalled and the published abstract reports directions rather than magnitudes โ€” treat this as a direction of effect, not a quantity.)

What the loss actually looks like downstream

The withdrawal rate is the visible cost. The quieter one is what happens to candidates who stay.

Of those who completed an AI interview, 51% had no outcome at all: 38% were ghosted entirely, 13% are still waiting. Only 28% moved forward (Greenhouse, 2026). That is not an AI problem โ€” it is a process that was never resourced to close the loop, now running at machine throughput. Independent research puts the same trend on a longer baseline: the share of job seekers rating their experience as excellent fell to 11% in 2026 from 18% in 2023, during a period of rising hiring-technology investment, with fewer than one in ten candidates never ghosted (PeopleScout, 2026).

Trust is priced accordingly. Only 21% of US candidates believe most employers use AI responsibly and transparently, and across all five markets between 63% and 70% do not believe it. Fifty-seven percent think companies should be legally required to disclose AI use in candidate evaluation โ€” a constituency that tends to get what it asks for eventually (Greenhouse, 2026).

The brand effect is closer than the headlines suggest, which is the strategically interesting part: in the US, 38% of candidates viewed the employer more positively after an AI interview against 34% more negatively, a net of +4. In the UK it is net โˆ’20. Same technology, different configuration, opposite outcome. The variance is in the implementation, not the tool.

What to change this quarter

If your AI interview stage was justified on screening hours saved, the business case needs re-running against a funnel that is self-selecting out at the top. Four moves, in cost order:

  1. Disclose before the stage, not during it. Say plainly, in the invitation, that AI is involved and what it looks at. This is the 44% ask, it is a template change, and it also happens to be the lever the ISR work shows reduces deceptive answers โ€” so it improves your signal quality at the same time.
  2. Offer a human-interview path on request. The 46% ask, and the top-ranked remedy in all five markets. You are not offering it to everyone; you are offering it to the fraction who will otherwise leave. Price it against the cost of replacing those candidates in the pipeline.
  3. Have reviewers score the recording before they see the AI score. The anchoring effect only operates if the number arrives first. This is a sequencing change in your workflow, not a procurement decision, and it is the only one of these four that protects decision quality rather than candidate experience.
  4. Keep the disclosure and the human decision on the record. If you are going to claim a person reviewed the evaluation, the claim should be inspectable rather than asserted โ€” what the system discloses, which provider processed the data, and the human decision logged against each hire are things a platform either publishes or does not. The 57% who want statutory disclosure are describing the direction of travel.

Note what is absent from that list: changing vendors, reducing automation, or going back to manual screening. Only 19% of your candidates are asking for that.

The AI interview did not cost you those applicants. The missing person did. Before you approve another screening-automation line item this quarter, find out what your own withdrawal rate is at that stage โ€” because right now it is almost certainly being recorded as disinterest.

Ready to go beyond the CV?

Scovai's AI-powered Talent Passport reveals what resumes can't: personality, potential, and true job fit.