There is a version of this article that says employees are frightened of AI, and a version that says they have quietly embraced it. Both are supported by survey data. Neither is much use to you.
The more useful finding is that employee acceptance is not a single number at all. It is a gradient, and it slides in a direction so consistent that you can design around it.
The gradient
The cleanest UK evidence comes from YouGov polling commissioned by the TUC in April 2024, fielded to 1,451 working adults. Opposition to AI rises with the consequence of the decision:
- 71% oppose AI being used in performance management and bonus decisions
- 77% oppose AI making hiring decisions
- 86% oppose AI making firing decisions
And separately, 69% think employers should have to consult staff before introducing AI at all.
It is worth naming the sponsor — the TUC commissioned this, and the TUC has a position. But the fieldwork is YouGov, the sample and dates are published, and the internal pattern is what matters: the numbers move in one direction, monotonically, as the stakes rise.
American data from Pew Research Center shows the same shape from a different angle. 71% of Americans oppose AI making a final hiring decision. But only 41% oppose AI reviewing applications. Same technology, same process, different point in the chain — and a thirty-point swing in acceptance.
The CIPD found something similar in January 2025: just 1% of respondents would trust AI to make important decisions at work. That figure came from a self-selecting LinkedIn poll rather than a representative sample, so treat the study’s headline claims carefully. But 1% is striking precisely because the audience was a self-selected professional one, predisposed to be interested in the technology. Even they wouldn’t hand over the decision.
What they actually like
Here is the part that gets left out of the anxious version of this article.
Pew found that 47% of Americans think AI would do a better job than humans at treating all applicants the same way — against 15% who think it would do worse. Among the large majority who see racial bias in hiring as a real problem, 53% believe more AI use would improve it.
That is not naivety. It is a reasonable read of the alternative. People who have experienced inconsistent treatment from human decision-makers are not sentimental about human judgement.
The CIPD’s Good Work Index 2025 found that the 16% of employees whose tasks had been automated using AI — typically repetitive ones — generally welcomed it, reporting better performance, job satisfaction and mental health.
So the picture is not “employees fear AI.” It is closer to: employees welcome AI where it removes friction, and refuse it where it removes judgement.
The disclosure paradox
There is one area where the evidence genuinely conflicts, and it happens to be the one most relevant to employee relations: will people tell a machine things they won’t tell a person?
The case for yes is strong and old. A 2014 study put participants in front of the same virtual interviewer, but told half it was operated by a human and half that it was fully automated. Only the belief changed. Those who thought they were talking to a computer showed lower fear of self-disclosure, less impression management, and were rated as more willing to disclose — including on substance abuse and PTSD symptoms. Remove the social evaluative threat, and people say more.
That mechanism is exactly what suppresses reporting of harassment and whistleblowing concerns at work.
The case for no is newer and larger. Nine studies with 6,282 participants, published in Nature Human Behaviour in 2025, took AI-generated empathic responses and varied only the attribution. Responses labelled as human were rated more empathic and more supportive than identical text labelled as AI. Merely suspecting AI involvement reduced the perceived support. In companion work, participants chose to wait days or weeks for human feedback rather than take immediate AI support.
These findings look contradictory. They aren’t — they measure different things, and the reconciliation is the practical insight:
- The 2014 work measures willingness to say the thing.
- The 2025 work measures the value of the response received.
An AI intake channel may well raise the volume and candour of first disclosure while simultaneously failing at the support function and reducing how cared-for someone feels. Which points at an obvious design: AI for intake, human for response.
One honest caveat, because it matters: neither study was conducted in a workplace. Neither included the variable that dominates real disclosure decisions — that what you say is logged by your employer, and might end up in a case file. We could find no robust independent study comparing actual disclosure rates of harassment or whistleblowing to AI versus human channels at work. Given that this is the entire commercial premise of AI intake tools, that gap should make you sceptical of anyone quoting a confident number at you.
What to do about it
Put AI where the gradient is friendly. Capture, summarisation, triage, drafting, deadline tracking, pattern-spotting across cases. This is the work employees are happy to see automated, and it is also where the efficiency gains are real.
Keep it visibly away from outcomes. Not because AI is incapable, but because the acceptance data is unambiguous and the trust cost of getting it wrong is asymmetric. A machine that drafts a letter a human signs is a tool. A machine that decides an outcome is a grievance waiting to be filed.
Tell people. The single most useful figure in Workday’s 2024 trust research — a study of 5,375 leaders and employees, and worth noting it comes from a vendor with a stake — is that only 22% of employees said their employer had shared any guidelines on responsible AI use. Prospect’s 2026 members’ survey found one in five members unclear on when or how they were supposed to use AI at all. Most of the trust deficit is not a technology problem. It is a communication vacuum.
Consult before, not after. Sixty-nine per cent think you should have to. You almost certainly don’t have to, legally. Doing it anyway is cheap, and it converts a thing that happens to people into a thing they had a say in.
Don’t hide the AI. The instinct is to make it invisible so nobody objects. The Nature Human Behaviour work suggests that backfires — suspicion of undisclosed AI depressed perceived empathy on its own. Concealment is a bet that nobody finds out, and it is a bad bet.
The tension in this category is usually framed as speed versus warmth, as though you must trade one for the other. The evidence says otherwise. People want the speed. They also want a person at the moment it matters. Those are compatible requirements — they are just not the same job, and the mistake is asking one system to do both.
For the wider picture of where AI is genuinely working in employee relations, and where it isn’t, see our state-of-the-nation guide to AI in employee relations.