Your heaviest AI users are the ones burning out

Almost every tech worker now says AI makes them better at their job. In the same survey, more of them are burning out than a year ago, and fewer think the field is worth recommending to anyone.

Those numbers come from the 2026 edition of a large annual survey run by Lenny’s Newsletter, a widely read product-management publication. 97.2% of respondents said AI makes them better at their work. 55.7% reported burnout, up 11 points in a year. Career optimism fell from 54.8% to 48.7%. Not one professional role scored above zero on the would-you-recommend-this-career measure, the standard net-promoter question pointed at a job instead of a product [verified].

If you manage people through AI adoption, there’s a tidy diagnosis for this on offer, and it’s a good one. Researchers at The Decision Lab, in a survey of 1,200 workers, and writers in Harvard Business Review describe something they call psychological debt: the quiet resistance that builds when you drop AI on people without redesigning how they actually work. Its tell is behavioral. People endorse the tool and then use it as little as they can get away with. The researchers call this acceptance-with-avoidance, and they find the workers carrying the most debt use AI roughly half as much as their peers, on the least important tasks [reported]. So the advice writes itself. Find the endorsers who quietly underuse the tool. That’s your problem group. Go re-engage them.

The new survey points somewhere else. Its largest group of workers isn’t the avoiders. It’s the people leaning in hardest.

The instrument that finds AI distress by watching who uses the tool less is blind to the biggest group of distressed workers, because they use it more. The same usage number that reads as “healthy” under one explanation reads as “at risk” under the other, and no one has plotted the chart that tells them apart.

In the survey, 49% of workers described themselves as “amplified” by AI, the single most common answer. Another 27.4% said “redefined,” 13.9% “destabilized” [verified]. Amplified sounds like the happy ending. It isn’t reading that way. The newsletter’s summary says amplification is landing as pressure rather than relief: faster pace, more expected output, pay that hasn’t caught up [reported]. The workers reporting the sharpest role-specific anxiety are designers and researchers [verified]. And the dominant fear is about what the job has become, with only 22% naming job loss at all [reported].

Here’s why that breaks the tidy diagnosis. Psychological debt is scored on two questions: do you accept the tool, and do you use it. Debt lives in one corner, where you accept AI and avoid it. Skepticism lives in another, where you doubt it and reject it. The whole appeal of the model is that you can spot debt without asking anyone how they feel. You watch what they do, and low use is the warning light.

The amplified don’t fit any corner on that grid. They accept AI, they use it heavily, and they’re burning out. Accept, use, distressed. There’s no box for it. The survey’s biggest group sits in a cell the diagnostic doesn’t have.

That points at a second explanation for the distress, and it runs the opposite way. Call the first one autonomy erosion: you pull back from the tool to protect your own judgment, so distress shows up as using AI less. The second one is about identity. The more of your craft the tool performs, the less the craft feels like yours, so the distress shows up in the people using it most, shipping at machine pace and losing the sense that the work is theirs.

Both stories fit the same top-line numbers. They disagree completely about where the trouble sits. One says look at the light users. The other says look at the heavy ones. They put opposite signs on the same measurement.

The honest objection is that these aren’t rivals, they’re layers. Some people withdraw, some people over-identify, and a mature reading holds both. There’s real force to that, and one detail sharpens it: “amplified” is a label people picked on a survey, not a count of how much they actually use AI. Maybe the amplified aren’t heavy users at all, and there’s no contradiction to see.

Maybe. But the diagnostic doesn’t get to stay agnostic, because it made a specific bet. It said the tell is underuse. Whatever “amplified” means as a feeling, an amplified worker is, by definition, not withdrawing from the tool. So pointed at that group, the instrument reports low debt, because they’re using it, while every population signal says distress is high and climbing. It doesn’t just miss these people. It reads them as the healthiest ones.

That stops being an academic quibble the moment you run an adoption dashboard. Usage is the number most teams watch, and high usage shows up green. The diagnosis above tells you to hunt for the low-usage endorsers and coax them back. But if the identity story is the one operating in your org, your dashboard has the alarm wired backwards. Your most amplified, highest-throughput people, the designers and researchers shipping fastest, are the ones rating the career un-recommendable, and they glow green because they use the tool the most.

There’s a clean way to find out which story is yours, and it’s sitting in data you already have. Plot burnout against how much people actually use AI, split by team or role. If distress falls as usage rises, the withdrawal story holds and your light users are the ones to go find. If distress climbs as usage rises, the tell is backwards and your heaviest users are the ones in trouble. I haven’t seen anyone publish that chart. The debate is stuck on aggregates that both explanations can claim.

Until someone runs it, treat usage on its own as a number that can’t tell you what you need. A heavy user might be thriving or hollowing out, and the count looks identical either way. The question underneath is whether using the tool more makes your people more themselves or less. If it’s less, the busiest column on your dashboard is where the fire is, and you’ve been reading it as the part that’s fine.