Candidates Don't Hate AI Interviews. They Hate Not Being Told About Them.

Two numbers published six weeks apart, from two different sides of the hiring table, describe the same gap without either source naming it directly.

On October 28, 2025, Korn Ferry released its twelfth annual Talent Acquisition Trends report, based on a survey of more than 1,670 external talent leaders plus 230 of its own recruiting consultants. The headline figure: 52% of talent leaders plan to add autonomous AI agents to their recruiting teams in 2026, and 84% plan to use AI in some form next year (Korn Ferry). Several outlets covering the report noted that some companies are already creating internal employee-style records for their AI agents, treating them as a new category of team member rather than a tool (Fast Company Middle East).

On May 1, 2026, Greenhouse published the candidate-side mirror of that same shift. Surveying 2,950 active U.S. job seekers, it found that 63% have now been interviewed by an AI — a 13-point jump in just six months. But 38% of candidates say they have walked away from a hiring process specifically because it included an AI interview, and another 12% say they would. The reason isn't a blanket objection to the technology: only 19% of respondents said they want less AI in hiring overall. The reason, in Greenhouse's own data, is that 70% of candidates say they were never clearly told upfront that AI would be evaluating them, and 21% only found out once the interview had already started (Greenhouse).

Put the two reports side by side and the pattern is not "employers are automating hiring and candidates don't like it." It's narrower and more fixable than that: employers are automating hiring faster than they are disclosing that they're doing it, and candidates are punishing the second thing, not the first.

The industry is solving the wrong half of the trust problem

Most of the public conversation about AI and hiring trust in 2026 has pointed at candidates, not employers — specifically, at whether candidates are using AI to cheat. That framing is worth checking against the actual measured data, because the gap between what gets reported and what gets detected is large.

A peer-reviewed study presented at USENIX Security 2026, analyzing roughly 196,000 de-identified resumes from the hireEZ platform, found hidden prompt-injection text — text rendered invisible to a human reader through tiny font sizes, matching background colors, or off-page placement — in about 1% of resumes, with the rate rising roughly sevenfold between July 2024 and November 2025 (Let's Data Science; reported separately via Duke-affiliated research in The Korea Herald). The study also found that more than 90% of the hidden text it detected wasn't an explicit instruction to an AI screener at all — it was ordinary keyword stuffing, just made invisible instead of visible. Staffing firm ManpowerGroup separately told the New York Times it detects hidden text in about 10% of the resumes it AI-scans, using its own scanning methodology — a higher number than the academic study, but still a small single-digit-to-low-double-digit share, not a crisis of scale (cited via The Interview Guys).

Compare that to self-reported survey data: Greenhouse's 2025 AI in Hiring report found 41% of U.S. job seekers say they have tried hiding instructional text in a resume to influence an AI screener. That's a roughly 40-point gap between what people claim to have done and what shows up when anyone actually scans for it. Some of that gap is surely people trying it unsuccessfully, or rounding "I used AI to write my resume" up to "I tried to manipulate the AI." But a gap that large, reported consistently across multiple measurement methods, suggests the industry's anxiety about candidate-side manipulation is disproportionate to its measured prevalence — while the employer-side disclosure failure, at 70%, is both larger and far easier to verify, because it doesn't require scanning anyone's resume. It just requires reading your own interview script.

Disclosure is a design decision, not a policy statement

The useful thing about the disclosure half of this problem is that, unlike detecting manipulated resumes, it isn't adversarial or probabilistic. An employer either tells a candidate up front that they're talking to an AI, or they don't. There's no cat-and-mouse dynamic, no arms race, no false-positive rate to tune. It's a single binary decision made once, at the point a screening flow is built — which also means it's one of the few parts of this whole AI-hiring-trust mess that a company can simply go check today.

This is the one piece of that design decision we can describe about our own product with precision rather than marketing language: every NiceHire AI screening interview discloses, in the greeting itself, that the candidate is speaking with an AI — in every language the product ships (English, Japanese, Traditional and Simplified Chinese), on a candidate's first attempt and on a reconnect after a dropped call alike. A customer can customize the wording and tone of that greeting, but not remove the disclosure sentence itself; the mechanism that assembles every greeting enforces it rather than leaving it to whoever wrote that particular script to remember. That's a narrow, specific claim about one control in one product, not a claim that disclosure makes an AI interview good, fair, or something candidates will enjoy — Greenhouse's own data shows candidates walking away from AI interviews even when they know going in, just at a lower rate than when they find out mid-interview. Disclosure removes one specific failure mode. It doesn't re-write the other 30 points of the problem.

What this means for a team building or buying an AI screening flow this quarter

Three things follow directly from the data above, not from generic advice:

1. Audit your own greeting script, not your vendor's marketing page. The Greenhouse number that should worry a hiring team isn't "63% of candidates now face an AI interview" — that trend line is set by the market, not by any one employer. It's "70% say they weren't told." That second number is entirely inside your own control and costs nothing to check: read the literal first sentence your AI screening tool says to a candidate, in every language you offer it in, including whatever happens on a reconnect after a dropped call. If a custom greeting can override the disclosure instead of simply changing its wording, that's a configuration gap waiting to be hit the first time someone writes a greeting without thinking about it.
2. Don't treat the 2026 Korn Ferry agent push as a reason to defer disclosure work. 52% of talent leaders adding autonomous agents to recruiting teams next year means more candidate-facing interactions will be AI-mediated, not fewer. Scaling the automation before fixing the disclosure gap scales the gap along with it.
3. Calibrate resume-screening suspicion to the actual measured rate, not the self-reported one. If your screening process is built around catching hidden prompt injections, the published detection rates — roughly 1% in the peer-reviewed study, up to around 10% by ManpowerGroup's own scanning — are the numbers to design against, not the 41% self-report figure. Over-indexing on the larger, softer number risks treating ordinary AI-assisted resume writing as adversarial behavior it usually isn't.

None of this is a claim that disclosure solves candidate trust in AI hiring, and we're not making that claim about our own product either. The Greenhouse data is explicit that even well-disclosed AI interviews lose some candidates — people are allowed to simply not want one, and telling them in advance doesn't change that preference, it just lets them make it on correct information instead of finding out mid-sentence. What the data does support is narrower and, we think, more actionable: the industry is spending a lot of attention on detecting candidates who might be gaming AI screening, and measurably less on the simpler, larger, fully-controllable question of whether employers are telling candidates what they're walking into. One of those problems requires better detection technology. The other just requires reading your own script.

Sources consulted directly for this piece: Korn Ferry's press release and Fast Company Middle East's coverage of the 2026 Talent Acquisition Trends report; Greenhouse's May 2026 newsroom post on its AI-in-hiring candidate survey; the USENIX Security 2026 prompt-injection study as summarized by Let's Data Science and by Korea Herald's coverage of the underlying Duke-affiliated research; and The Interview Guys' analysis of Greenhouse's 2025 self-report figure against ManpowerGroup's independently reported detection rate.