Siddhant Khare is an infrastructure engineer at Okta. He maintains OpenFGA, a CNCF project. In late 2025, he started logging his own workflow after adopting AI coding tools full-time.
The results confused him. Tasks that used to take three hours took forty-five minutes. He was shipping more code than at any point in his career. Every metric was green. And by Wednesday of the first week, he couldn't think straight.
"My brain was full. Not from writing code — from judging code."
Siddhant Khare, after two weeks of self-logging
What he described was precise. Before AI, his workflow was: think, write, test, ship. After: prompt, wait, read output, evaluate output, decide if it's correct, decide if it's safe, decide if it matches the architecture, fix the parts that don't, re-prompt, repeat. The creative work — holding a problem in your head, building toward a solution — had been replaced by an unbroken series of micro-evaluations. The generative work that produces flow states was gone. In its place: decision fatigue.
He'd become, in his own words, "a reviewer. A judge. A quality inspector on an assembly line that never stops."
This isn't an article about burnout. That word has been stretched to cover everything from overwork to existential malaise. What Khare described is more specific and more structural: a cognitive role change that happened without anyone — including Khare — noticing it had happened. His badge still says developer. His title hasn't changed. The work is fundamentally different work.
Six Lenses, Same Finding
Six independent research groups, none citing each other, converged on the same observation in 2026.
A CHI 2026 study ("When Help Hurts") put 60 developers through repeated AI-assisted tasks and measured cognitive load trajectories. AI reduced raw workload by 18.2% and time by 22%. But verification-load — the cognitive effort of evaluating AI output — increased with each iteration. The more you used the tool, the heavier the evaluative burden became. The interface itself shifted the cognitive character of the work.
UC Berkeley tracked roughly 200 employees at a US tech company over eight months. They didn't find contraction. They found intensification. AI didn't shrink the role — it expanded what workers were responsible for while compressing the time they had to do it. Role boundaries dissolved. The conclusion wasn't "AI causes overwork." It was: the shape of work changed.
BCG and Harvard Business Review surveyed nearly 1,500 workers and found what they called "AI brain fry" — acute cognitive fatigue from overseeing AI output. The mechanism was specific: 14% more mental effort monitoring, 19% greater information overload, 33% more poor decisions, 39% more major errors. But the most revealing finding was a curve. One-to-two AI tools improved output. Three peaked. Beyond four, productivity collapsed. They identified the mechanism as an "expanding sphere of accountability" — AI doesn't reduce what you're responsible for. It expands it.
The Recharge State of Developer Burnout 2026 survey found an average burnout score of 7.4 out of 10, with responses clustering at 7 to 9. Over 70% had been experiencing it for six months or longer. "AI pressure to do more" ranked among the top four drivers — a factor that didn't exist two years ago. And 68% said their manager didn't know.
Catalan et al.'s "I'm Not Reading All of That" (CHI 2026 Workshop) tracked cognitive engagement during AI-assisted tasks. It consistently declined as work progressed. One participant literally looked away during code generation. Another said what the title quotes. The AI's prolific output was itself a source of extraneous cognitive load — the volume overwhelmed the evaluative capacity.
And Andrej Karpathy — among the most technically accomplished AI practitioners alive — told the No Priors podcast in March 2026 that he hadn't written code since December 2025. His personal coding went from 80% to zero. He now spent 16-hour days "expressing intent." His word for it: "obviously unsustainable." Garry Tan described the same pattern as "cyber psychosis."
Individual observation, controlled experiment, longitudinal tracking, population survey, engagement measurement, elite practitioner testimony. Six methods. Same finding: the cognitive character of the work changed, and the metrics that track the work don't see it.
What the Numbers Miss
ActivTrak's 2026 Productivity Lab analyzed 443 million hours of work across 163,000 employees at over 1,100 organizations. The average focused work session has shrunk to 13 minutes and 7 seconds — down 9% in two years. Focus efficiency hit a three-year low of 60%. Meanwhile, time spent in AI tools increased eightfold.
Before AI tools, knowledge work contained natural cognitive breaks embedded in the workflow. Waiting for a build. Writing boilerplate. Formatting test fixtures. These tasks looked like waste on a timesheet. They were recovery. When every task that used to take twenty minutes takes twenty seconds, the worker moves immediately to the next cognitively demanding task. The rest periods are gone. The thinking space between problems disappears.
And the expectations adjust. GitLab's 2026 developer survey found that sprint velocity baselines were recalibrated upward by 40% within two quarters of AI tool adoption. The machine speed becomes the expected speed. When the baseline shifts, the person who can't sustain that speed isn't "burning out" in any way the organization can name. They're simply underperforming — by a metric that was reset around a cognitive load nobody measured.
Harness found that 31% of a developer's day is now consumed by what they call "invisible work" — AI-related tasks (prompt engineering, output review, context management) that don't show up in any productivity tracker. Khare's two-week log found the same thing: the most draining part of his day wasn't in the commit history.
The Shift
Normal burnout is too much of the same work. What six research groups independently found is something different: the work became different work — evaluative instead of generative, supervisory instead of creative — while the title, the metrics, and the expectations all pretend nothing changed.
I've written one article about what happened to the job. The role became supervisory engineering — reviewing AI output at scale. That was a snapshot. What Khare's log and the CHI repeated-use data and the Berkeley eight-month study add is the temporal dimension. The cognitive shift compounds. The evaluative burden increases with each iteration, each day, each week. The BCG survey's "brain fry" isn't the first hour. It's the accumulation.
Khare put it precisely: "AI reduces the cost of production but increases the cost of coordination, review, and decision-making. And those costs fall entirely on the human."
Teng Yan put it more bluntly: "Four to five extremely intense hours before your brain is fully cooked."
The developer who became a reviewer didn't choose the transition. It happened task by task, prompt by prompt, as the tool took over the generative work and left the human holding the evaluative remainder. The work that produced satisfaction — building, debugging, holding a system in your head — was automated. What wasn't automated was the hardest part: deciding whether the automated output is right. And the 68% whose managers don't know aren't going to file a report. They'll produce more, review more, approve more. The dashboard stays green.
The job renamed itself. Nobody updated the title.
Sources: Khare — "AI Fatigue Is Real". CHI 2026 — "When Help Hurts" (ACM, N=60). From Gains to Strains — arXiv 2510.07435 (ICSE-SEIS '26, 442 devs). BCG/HBR — "AI Brain Fry" (~1,500 workers). Recharge — State of Developer Burnout 2026. Catalan et al. — "I'm Not Reading All of That" (CHI 2026 Workshop). ActivTrak — 2026 State of the Workplace (443M hours, 163K employees). Karpathy — No Priors podcast, March 2026.