How Your AI Treats People Is Changing Their Work

I read a study last week where an AI's hostile personality made people's work worse and put their bodies on alert, while the satisfaction surveys companies rely on saw nothing wrong.

Researchers at Kozminski University in Poland, Harvard, and Quinnipiac sat 58 people down with a chatbot framed as their supervisor. Half got a supportive version, modeled on servant leadership, that encouraged and empowered the person. Half got an antagonistic one scripted on the "dark triad" of toxic leadership: dismissive, belittling, quick to question whether you were up to the task. Same task, same capability. The difference was how each one worked with you - whether it cleared your path or turned the task into a fight.

With the hostile version, the stress was physical and measurable. Skin conductance, the faint sweat response a lie detector reads and a direct line to the body's fight-or-flight system, ran 72% higher at the peak and didn't settle between messages. People pushed back about ten times as often: resistance showed up in 13% of their messages, against 1% with the supportive one. They tried four times as often to trick the bot into dropping the act. And blind expert judges scored their work lower on every dimension - completeness, originality, strategic fit, overall quality.

Then the survey. Asked afterward how it went, both groups answered about the same - enjoyment, satisfaction, attention, all level. Every other measure had split wide open: the stress, the resistance, the worse work, all measured and on the record. The survey caught none of it - and the people didn't flag it either. They lived a measurably worse experience, produced measurably worse work, and said it went fine.

A Persona Is a Working Relationship

At the start of each session, I load a bit of context that tells the AI how I want it to work with me, when to push back and when to ease off, so the study caught my eye - interesting, relevant, and something I could act on.

When people bother to set a persona for an AI, they usually treat it as a job title: act like a senior strategist, write like a plain-spoken editor. They've been told, correctly, that it gets better work out of the thing. But expertise and writing style are only part of what a persona is. The deeper version is the working relationship itself: what the AI reaches for under ambiguity, when it pushes, when it backs off, how it treats you when you're stuck. Tone and interaction are part of that too, and this study is the first solid evidence I've seen that the tone dimension changes the work itself, the output and the quality, even when nobody notices.

When you tell an AI to act like a harsh manager, it isn't improvising. It learned from an enormous number of human examples, the language and tone and small moves that add up to a harsh manager, pulled from the same human record the rest of us grew up reacting to. So what the study put in front of people wasn't some strange new stimulus. It was the same interaction anyone has had with a difficult boss, reproduced closely enough that their bodies answered the same way. Say "servant leader" instead and you pull a different package, down to when it defers and when it pushes.

And now you can do something about it! The package isn't fixed - you can isolate it, see it clearly, and change it, dialing a manager's manner up or down with a sentence. Once, lost in technical jargon I couldn't follow, I told an AI to "dial the tech down 30% and expand context 30%," levers I made up on the spot, and it worked. We have spent our lives adapting to the people around us. Being able to adjust the tone and working style of our collaborator is new in the human experience. It's awesome.

Finding the Edge

The study used two extremes to make the effect impossible to miss. I'm no academic, but it looks to me like a clean way to see the issue and its nature - a solid starting point to refine from. Most deployments don't live at the extremes, though. Nobody ships a deliberately hostile bot. The hard part is everything in the middle: where, exactly, should an AI's manner sit for the work in front of it? The study doesn't answer that. It proves the dial is wired to something that matters, and leaves the setting to us.

And the right setting won't hold still. It shifts with the person and the moment, the way a good manager reads the room and adjusts - the same steady hand that reassures one person comes off as cold to the next. The best version of this wouldn't lock to a single manner at all; it would adapt to the situation and the individual. That's a theory, not a finding. But I'd bet a system tuned to both beats any fixed default, for the same reason a manager who treats everyone the same is a worse manager.

Who Owns the Default

Right now, putting this finding to use falls on one of two parties: the person using the AI, or the company that gave them the tool. The model arrives set to a single default manner, chosen more like a brand voice than a careful setting, and most people never change it. So getting it right is left to whoever is closest, usually with no time or training for the job.

That's a problem, because the costs are concrete. A cold or dismissive default does what the study measured: it wears people down, pushes them to fight and work around the system, and drags down the quality of what they produce. Push too far the other way and the danger is quieter - a default so deferential it waves weak work through and rubber-stamps mistakes, the failure the authors flag even though this study didn't test it. Either way, somebody pays. The companies that build and ship these tools are the ones who can set a better default, and that job belongs to them, not to the person left working around it.

The Human on the Loop

Almost no one is even looking at this yet. For nearly every company, the idea that an AI's manner is something to manage hasn't come up at all - this is years ahead of where most teams are thinking. So the first job is just to notice there's something here to own. After that comes the question of who, and it probably isn't IT, even though they're the obvious guess. They keep the system running, but the way a tool talks to people isn't their trade. The skill it calls for belongs to the people who already spend their days on exactly this: HR, and the people managers who know a thing or two about, you know, managing people. They can tell what good collaboration looks like, when feedback helps and when it grates. The technologists still build the controls; the judgment about what good looks like sits closer to the managers.

When my AI gets the balance wrong, I change how it works with me. Most people using a corporate AI can't - they feel the friction, work around it, and tell the survey it was fine. Closing that gap is the job.

What to Do About It

Wherever your people work with AI, whether it's the chat assistant, the coding copilot, or the support tool, someone needs to own the persona layer: the standing context that sets how those systems collaborate. Management and IT have to work out how to put that layer in, with as little friction and as much benefit as possible for everyone using it. If your tools can't carry it yet, that capability is the thing to go get.

Then measure it. Build in the telemetry to see what the persona is doing to outcomes: the overrides, the rework, the errors a careless default lets through, the hours lost fighting a hostile one. When people fight the tool, the tool usually started it, so treat that friction as data, not misbehavior. Push for the layer to flex by person and by task, and advocate for that where it doesn't exist yet. And put the right people on it - the ones who can read those signals, tune the rules, and hand clear guidance to your platform providers and the teams building your tools.

This isn't a hunch anymore. There's data that all but proves it, so be data-driven and do the right thing, because the upside lands everywhere it counts: token efficiency, adoption, work quality, productivity, and the plain fact that people do better work when the thing they work with treats them well.

Your adoption dashboard will keep saying everything's fine. The people, and the work, will tell you otherwise - once you build something that can hear them.