Why People Resist AI at Work, and When They're Right

Picture the dashboard a few months after a company rolls out its AI assistant. A handful of people use it daily. Many tried it twice and stopped. A few have said openly that they won't. The usage line is flat, and the conclusion writes itself: people resist change, so run more training.

Now put a second number next to that dashboard. In a 2024 survey of about 31,000 knowledge workers in 31 markets, three in four said they already used generative AI at work. And 78 percent of those users were bringing their own tools. More than half were reluctant to admit using it for their most important tasks.

So in many companies the flat line doesn't mean people aren't using AI. It means they aren't using yours, or they're using it and not saying so.

That's the first misreading. The second is bigger. Resistance to AI at work is usually treated as an emotion to manage. More often it's information: a specific complaint about the work, the risk or the bargain, delivered as reluctance. Treated as noise, it gets suppressed. Treated as data, it's some of the cheapest information an AI program can get.

First, separate resistance from poor fit

Not every non-user is resisting. A large Pew survey of US workers in late 2024 found that most rarely or never used AI chatbots at work, and among those who didn't use AI much, fewer than a third thought any of their work could be done with it. For many of them, the tool simply doesn't fit the job as they see it. Some are wrong about that. But showing someone where a tool helps is a different task from listening to resistance.

Resistance, in the sense that matters here, is narrower: people whose work the tool clearly could touch, who avoid it, use it reluctantly or hide that they use it. That group carries the signal.

Five things resistance can mean

Resistance is a behavior with several possible causes, and each calls for a different response. The five below are my synthesis of the evidence, not a measured taxonomy; nobody has counted how often each occurs.

1. The work got worse for them

When AI speeds up one person's step, the output often lands on someone else to check, fix or decide (Where the Saved Time Goes). The person who saved time reports a gain. The person who absorbed the cost rarely reports anything. A reviewer whose queue doubled, a support lead who now rewrites AI drafts before they go out: from their seat, the tool made the job harder.

There's a name for one version of this. Workslop is AI-generated work that looks finished but doesn't have the substance to move a task forward. In a 2025 survey of US desk workers, 40 percent said they'd received some in the previous month.

This kind of resistance is often right about the facts. The fix is in the flow: budget review capacity, make output cheaper to check, or move the AI to where the work actually waits.

2. Accountability without control

People are asked to use a tool and stay responsible for what it produces, with no say in how it was configured, what it may do or how its errors get caught. A claims handler who signs off on an AI-drafted decision carries the consequences of a mistake she didn't make and couldn't easily see.

Also often right. The response is to make accountability explicit and fair: who owns the rule, who checks what, what happens when the system is wrong (Human Review Where It Matters).

3. Trust broken by a visible error

People who see an algorithm make a mistake lose confidence in it faster than they would in a person who made the same mistake. Researchers named this algorithm aversion in 2015, after forecasting experiments in which people who had watched an algorithm err were less likely to bet on it than on a human forecaster, even when they had seen it do better.

It's partly a bias and partly good sense. AI tools are reliable on some tasks and unreliable on neighbors that look similar, so a person who saw it fail on their task has real information. Telling them the tool is "95 percent accurate" doesn't repair trust broken by a specific failure they watched. Showing them where it's reliable on their work, and letting them keep checking where it isn't, does.

4. Identity and craft

Work is part of who people are. A writer whose job becomes editing machine drafts, an analyst whose model-building now takes seconds, a programmer who enjoyed the code and now reviews it: each loses something real, even when the output is better.

Some of this protects status. Some of it protects something worth protecting. In a field experiment with nearly 1,000 high school math students, those who practiced with an unrestricted GPT-4 tutor did better while they had it, then scored about 17 percent lower than peers once it was taken away. A tutor designed to give hints instead of answers largely avoided the drop. Students aren't employees, but the mechanism carries over: a tool that does the part of a task that builds skill also stops the skill being built (Skills That Fade When the Machine Does Them).

The response is to decide deliberately which skills the organization needs people to keep, and design the work so they keep practicing them.

5. The bargain is unclear

AI changes how much work a job takes. Nobody has said who gets the difference: more output for the same pay, the same output in less time, fewer positions, a new role. Without an answer, many people assume the worst for themselves, and they have some basis. Employment of 22-to-25-year-olds in the occupations most exposed to AI has fallen behind that of less-exposed peers, though nobody has shown AI caused it (A Job Is a Bundle of Tasks).

Resistance here is negotiation, sometimes explicit. The Writers Guild of America's 2023 contract, reached after a long strike, set terms for AI in screenwriting: AI-written material can't count as literary material, a company can't require a writer to use AI, and it must disclose when material it hands a writer was AI-generated. Most workplaces won't have a union contract. The question underneath is the same everywhere.

Resistance is often a specific complaint about the work, the risk or the bargain, delivered as reluctance.

Hidden use is the mirror image

The people who use AI and don't say so deserve as much attention as the ones who refuse. They've often done the hard part already: figured out where AI helps in their actual work. The organization learns nothing from them, because the incentives punish disclosure. Using AI might signal that the job is easier than it looks, or that the person is replaceable, or that they broke a policy written before anyone used these tools. In the knowledge-worker survey, 53 percent of users worried that using AI on important tasks made them look replaceable.

The cost is easy to name. Individual gains stay individual. Sensitive data flows into tools nobody vetted. And the company plans its AI program around the official tool while the real adoption happens somewhere else.

Notice that the refusers and the hiders are probably responding to the same thing: uncertainty about what using AI will mean for them. The surveys don't test that link, but it's the most plausible common cause, and it suggests a common remedy. Make the consequences clear and safe, and both groups come into the open.

Reading the signal

A simple map helps sort what you hear. One axis asks whether the complaint is specific or general. The other asks whether it's about the work itself or about status, security and fairness.

Scroll sideways to see the whole figure

Specific complaints about the work ("it doubles my checking", "it gets our refund rules wrong") are signals to act on: change the flow, the configuration or the checks. General complaints about the work ("it's unreliable") call for calibration: find out where it's unreliable for them, then fix it or narrow the use. Specific complaints about status or fairness ("who gets the time it saves?") call for an answer, which means a decision by someone with the authority to make it. General fear calls for clarity: what will change, for whom, when, and what won't.

The common mistake is to have one response for all four, more communication. It fits exactly one quadrant.

What to do

My advice, and these are judgments from how the five causes and the incentives work, not results from controlled trials:

Don't set usage targets. Requiring people to use AI a certain number of times produces usage, not value. It invites prompts written to satisfy the metric (Measuring AI Impact Without Fooling Yourself). Worse, it tells people the organization cares more about the appearance of adoption than about their work.

Don't monitor AI use to judge individuals. It pushes use further underground. Monitoring for security risks is different and necessary; mixing the two makes people distrust both.

Start where AI removes work people dislike. Data entry, chasing documents, reformatting, summarizing long threads. Early wins that make people's days better build the trust later changes need.

Bring in the people who do the work while the design can still change. Ask them early what the tool would break and what it would fix. Treat them as the first critics of the design, not its last obstacle. They know the exceptions and the failure modes, and they'll be the first to notice when it goes wrong in production.

Answer the bargain question, even partially. "Time saved goes to these three backlogs, not to headcount reduction this year" is an answer. "We're exploring the opportunities" isn't.

Make sharing safe. Ask who already uses AI, for what, and what they learned. Reward the answers. It's the cheapest way to find out what your real adoption has already discovered.

Count the losers of the change. Every redesign makes some jobs better and some worse. Programs that only measure the winners discover the losers later, as attrition, quiet non-use or quality problems.

The people pushing back are often the ones who know where the design will fail. Ask them first.

Check yourself

Why might a flat usage line be misleading?

Many people use their own tools instead of the official one, or use AI and don't say so. The flat line can mean they aren't using yours.

Why does "more communication" fit only one kind of resistance?

Specific complaints about the work need changes to the flow, general complaints need calibration, and fairness questions need a decision. Only general fear is answered by clarity about what will change.

Why do refusers and hidden users deserve the same attention?

Both are probably responding to the same uncertainty about what using AI will mean for them. Hidden users have also often found where AI helps in their real work, and the organization learns nothing from them.