dscout研究：AI抵触源于身份认同而非能力不足，96名在职用户揭示“保留区”不随技术提升而移动

A flight attendant wasn’t someone I expected to speak to about where AI might threaten their job, but this participant’s point of view wasn’t unique among the 96 workers who we spoke to about where their job ends and AI’s begins.

The flight attendant didn’t cite a skill gap or even necessarily a doubt in AI’s ability to do certain aspects of her job, she described a loss. 

We heard some version of this across 96 workers. All of them were already using AI tools, because that was the point of the study. We didn't want people speculating about a hypothetical future. We wanted people living inside the reality of AI-assisted work, telling us what it actually feels like from the inside.

What they told us reframed everything we thought we knew about the "AI and work" conversation.

Nobody said “because AI can't.”

The standard assumption about worker resistance to AI goes like this: people protect their jobs because they don't trust the technology. As AI gets better, the resistance erodes and eventually, workers accept it.

It's a reasonable hypothesis and it's also mostly wrong.

When we asked people what they'd protect from AI; even if AI could technically handle it, almost nobody reached for a capability argument. They didn't say the tools weren't good enough, or that the output would be worse. They said something quieter and more personal.

“That's mine.”

A salesperson said reading a client in the room: the hesitation, the moment trust lands, the look that tells you the pitch is working, is something they'd keep for themselves. This wasn’t because AI would fumble it, but rather because handing it off would mean something was lost that wasn't replaceable by a better result.

A manager said the judgment calls, especially the hard ones, where you're weighing people's livelihoods and trying to be fair, belong only to them. It's not just a task, it's a responsibility they've claimed.

What unites these isn't skepticism about AI's capabilities. It's something more like ownership. The sense that certain work is theirs; not because they're the only ones who can do it, but because doing it is part of who they are.

What work means vs. what work produces

There's a useful distinction lurking in this data that doesn't get enough attention.

When researchers and product teams think about AI adoption, the conversation tends to center on output. Will the work get done? Will it get done well? Will it be faster, cheaper, better? These are reasonable questions, but they miss something the workers in this study kept naming without quite naming it.

Work doesn't just produce things. It makes people who they are.

The salesperson who reads a room isn't just generating a successful pitch. They get to clock in as “the person who knows how to read rooms.” The manager making a hard call isn't just resolving a situation, they sign on as “someone whose judgment can be trusted.” The flight attendant isn't just transporting passengers, they show up as someone “who makes people feel cared for.”

When AI takes over a task, it doesn't just change how the work gets done. It changes who gets to do that becoming.

That's what the "mine" language is pointing at, an issue of identity rather than about the quality of AI outputs. And it turns out that claim is remarkably stable, even as the tools get dramatically and demonstrably better.

There are two fears in this conversation, and we keep covering only one.

The AI-at-work anxiety that gets the most attention is economic: AI might take my seat. This is the job displacement story, which is real, it matters, and workers feel it.

But a quieter fear surfaced in this study: AI might take my edge.

That “edge” is not the job, it is the sharpness, the practiced judgment, the thing that makes someone genuinely good at what they do, rather than just technically capable of doing it. Several workers described worrying that over-reliance on AI would erode skills they'd spent years building. Not because the AI would take those skills, but because you only keep skills by using them, and if you stop using them, eventually they diminish.

This is a more sophisticated anxiety than job displacement. It isn’t about losing a seat at the table, it's about arriving at the table and discovering that you no longer belong there.

And it changes the nature of what workers are protecting when they draw the “hands off” line. For some, the protected zone is about meaning and identity. For others, it's about not letting a capability atrophy that they might need when the AI isn't there, or isn't right, or gets it wrong in a way only they would notice.

Both of these: the meaning protection and the skill protection, are doing real work in how people relate to AI. Neither of them shows up when you just measure “task acceptance” or simple adoption.

The line that didn't move

Here's the finding that should surprise any researcher who studies adoption.

Despite AI getting dramatically better over the years leading up to this study, workers' sense of what they'd keep for themselves hasn't shifted much. When we probed on this (“has your thinking changed as AI has improved?”) most folks said “no.” The zone was roughly the same.

If the “hands off” boundary were capability-based, you'd expect it to erode as AI improves, right? That's how it works with most tools: once the technology is good enough, resistance fades, but that's not what happened here.

Which tells us the line was never about capability in the first place.

The work people protect isn't protected because AI can't do it. It's protected because it matters to them and because it's part of how they understand themselves, how they stay sharp, how they make meaning out of what they spend their time doing. That kind of protection doesn't yield to a better tool. It yields, if it at all, to something more fundamental changing in how people think about work itself.

For researchers studying AI adoption, this is the crux of it. If you're modeling adoption as a capability problem: something that improves as tools improve, you may be fundamentally misreading the resistance you're seeing.

What this means if you're doing AI research

A few things to take with you:

Stop measuring adoption at the task level

Acceptance of AI for a specific task doesn't tell you much about how workers actually feel about AI in their work. The "mine" zone operates above the task level. It's about categories of work that carry identity weight and those don't get unlocked by a smoother interface or a more accurate model.

Ask people what their work means before you ask what AI should do with it 

This is a sequencing question as much as a research design question. Workers have a rich understanding of which parts of their job are about output and which parts are about who they are. If you jump straight to "how could AI help," you'll miss the terrain that actually governs their behavior.

Treat skill erosion anxiety as its own phenomenon 

It's not the same as job displacement fear, even though they often travel together. A worker who's afraid of losing their seat and a worker who's afraid of losing their edge will respond very differently to AI. That means they will need very different things from the tools, the organizations deploying those tools, and the researchers trying to understand what's happening. Right now, most product strategy conflates the two.

The question that keeps me up at night

We've spent a lot of the last few years building AI tools on a particular assumption: that the goal is maximum delegation. That people want to offload as much as possible, and the job of the technology is to let them do so.

The workers in this study push back on that, gently but clearly. They want to offload some things: the repetitive stuff, the administrative drag, the tasks that drain them without building them. But they want to keep other things and The Keeping is part of the point, it’s very intentional.

So here's the question we're sitting with: what happens to work, and to workers, when the tools are built for maximum delegation, but the people using them never actually wanted to delegate everything?

We don't have a clean answer. But we think it's the right question to be asking.