87% of companies now use AI somewhere in hiring. Only 26% of applicants trust AI to evaluate them fairly. That gap is usually presented as a problem of perception — candidates will come round once the technology proves itself.
The data says otherwise. The gap is not really about AI at all. It is about being told, and the fix is already being written into law.
Both sides of the gap are real
The employer case is genuine and worth stating before criticising anything. Adoption went from 26% of companies in 2024 to 87% in 2026 because the gains are measurable.
- Talent teams using AI report saving around 20% of their working week — roughly a full day per person.
- 72% of recruiters say AI sourcing tools find qualified candidates faster than traditional methods.
- AI-enabled talent acquisition is projected to cut costs by 20–30% and reduce time-to-hire by 30–40%.
The candidate side is equally well evidenced, and less comfortable.
| Finding | Figure |
|---|---|
| Applicants who trust AI to evaluate them fairly | 26% |
| Job seekers uneasy about AI-led hiring | 67% |
| Candidates who dropped out after discovering an AI-led interview | 30% (Greenhouse, 2026) |
| Candidates who withdrew from a process specifically because of an AI interview | 38% (separate survey) |
| Would walk if surprised by one in future | A further 12% |
Between three and four candidates in ten leave the process. That is not a sentiment score. It is attrition, and it lands on the same funnel the efficiency gains are measured in.
The cause is disclosure, not automation
Two numbers reframe everything above.
- 82% of candidates interviewed by AI were not told beforehand.
- 24% only realised AI was involved once the interview had already begun.
Read the dropout figures again with that context. Candidates are described as dropping out “after discovering” an AI interview. Discovery implies concealment. The measured behaviour is not a reaction to AI — it is a reaction to finding out mid-process that nobody mentioned it.
That distinction matters because the two problems have opposite solutions. If candidates object to AI, your options are to stop using it or wait for attitudes to shift. If candidates object to being surprised, the fix is a sentence in the invitation email.
What the dropout actually costs
Put the two halves of the business case in the same calculation, because they are almost never presented together.
| Gain | Cost |
|---|---|
| 20% of a recruiter’s week returned | 30–38% of candidates leaving the process |
| Time-to-hire down 30–40% | Shortlist quality falls with the pool |
| Cost per hire down 20–30% | Self-selection is not random |
The third row is the one that should worry a hiring manager. Candidates who walk away are not a random sample. People with options leave first — those with competing offers, in-demand skills, or the confidence that another process will have them. Undisclosed AI does not shrink your pipeline evenly. It filters out the people you were competing hardest for.
A tool that saves a recruiter a day a week while removing a third of the strongest candidates has not improved hiring. It has improved a metric.
Regulation is about to mandate the fix anyway
The disclosure that would close the trust gap is largely the same disclosure the law is moving toward requiring.
| Rule | Status | What it requires |
|---|---|---|
| NYC Local Law 144 | Enforceable since July 2023 | Annual independent bias audit, public disclosure of results, and at least 10 business days’ notice to candidates before an automated employment decision tool is used |
| EU AI Act | High-risk employment obligations apply from 2 December 2027 | Risk assessments, technical documentation, bias testing, human oversight, transparency disclosures, continuous monitoring |
| Colorado SB 205 | In force | Proactive steps to prevent discriminatory outcomes, with documented bias mitigation |
| California, Illinois, Connecticut | In force | Discrimination and transparency liability rather than a named audit requirement |
The EU AI Act classifies AI used for recruitment, candidate selection, performance evaluation, task allocation, monitoring and promotion decisions as high-risk. That is the second-highest tier in the Act, below only prohibited uses.
December 2027 sounds distant. It is about fifteen months away, and the obligations — documented risk assessments, bias testing, demonstrable human oversight — are not things you implement in a quarter. Outside the EU, only NYC mandates a bias audit by name, but in jurisdictions creating discrimination liability instead, an independent audit is the strongest available evidence that you took reasonable care.
Sourcing and screening are not the same risk
This distinction gets lost in most coverage, and it decides how much of this applies to you.
| Sourcing | Screening and selection | |
|---|---|---|
| What it does | Finds candidates who might fit | Evaluates, ranks or rejects candidates |
| Effect on the individual | Adds them to a pool | Determines an outcome |
| Regulatory exposure | Generally lower | The high-risk category |
| Candidate trust impact | Minimal — usually invisible | Where the 26% figure comes from |
Using AI to build a longlist from a job description, or to find people whose profiles resemble a strong past hire, does not decide anyone’s outcome. Nobody is rejected by a search query. Using AI to rank, score or auto-reject applicants does decide outcomes, and that is where both the regulation and the trust problem concentrate.
Many teams worried about AI hiring risk are running sourcing tools and applying screening-grade anxiety to them. Others are auto-rejecting at scale and assuming the rules are about chatbots. Work out which side of that line each of your tools sits on before doing anything else.
Closing the gap
- Disclose before, never during. One sentence in the invitation stating that an AI-assisted stage is involved, what it assesses, and that a human reviews the outcome. This is the highest-return change available and it costs nothing.
- Offer an alternative. A stated route to a human-led equivalent converts a dealbreaker into a preference. Most candidates will not take it; the option is what does the work.
- Make human oversight visible, not nominal. “Reviewed by our hiring team” only helps if it is true, because candidates compare notes and the claim is checkable.
- Never auto-reject silently. Roughly half of job seekers report being rejected without any communication at all. Silence after an AI stage is where a compliance question becomes a public complaint.
- Audit anything that scores or ranks. Required in NYC, sound evidence of care everywhere else, and mandatory across the EU from December 2027.
Measure the effect where it shows up: completion rate by stage. If disclosure is working, the drop between invitation and completed assessment narrows.
Frequently asked questions
Will disclosing AI use scare candidates off?
Some, but fewer than concealment does. The measured dropout happens on discovery mid-process, and 82% of candidates were never told in advance. A disclosed AI stage lets people opt out early rather than abandon a process they had already invested in — and it removes the sense of having been deceived, which is what turns a withdrawal into a complaint.
Does the EU AI Act apply to us outside the EU?
If you recruit candidates located in the EU, generally yes, regardless of where your company sits. The high-risk employment obligations apply from 2 December 2027. Given that documented risk assessments and bias testing take time to establish, treat that as a 2027 programme rather than a 2027 deadline.
Is AI sourcing subject to the same rules as AI screening?
Generally no. The high-risk category centres on tools that evaluate, rank or reject candidates. Sourcing tools that surface people into a pool do not determine an individual’s outcome. Check what each tool in your stack actually does rather than what category the vendor markets it under.
Do we need a bias audit?
Legally required by name only in New York City, where it must be annual, independent and published. Elsewhere — Colorado, California, Illinois, Connecticut, the EU — the liability is for discriminatory outcomes, and an independent audit is the strongest evidence that you exercised reasonable care.
What is the single highest-return change?
Adding one disclosure sentence to your interview invitations. It addresses the mechanism behind a 30–38% dropout rate, costs nothing, and moves you toward compliance with rules arriving in 2027.
The takeaway
The distance between 87% adoption and 26% trust is not a gap that time will close. But it is also not evidence that candidates reject automation. It is evidence that they reject finding out about it halfway through.
The useful conclusion is that the trust fix and the compliance fix are the same work, done once. Disclose the AI stage, keep a human genuinely in the loop, explain the outcome, and audit anything that scores people. That satisfies NYC today, prepares for the EU in 2027, and stops a third of your strongest candidates leaving before the shortlist.
For the sales-side counterpart to this question, see whether AI is really replacing SDRs. On using hiring data as a signal rather than a screen, see B2B buying signals in 2026.
Request a demo and we will run a live search against your best-fit candidate profile. Or explore the Recruitment Bundle, and read more in our AI for Sales & Recruiting topic hub.