I'm Becoming a Progress Junkie (and AI is the Dealer)

In July 2025, METR ran a randomized controlled trial on sixteen experienced developers, 246 real tasks, on the developers' own codebases. The devs estimated AI made them 20% faster. The data said they were 19% SLOWER ( Becker et al., METR, 2025 ). That number rocketed around the internet. But it's worth knowing what happened next. In February 2026, METR tried to run a clean follow-up and couldn't — too many developers refused to participate in the AI-forbidden arm of the study because they didn't want to work without AI ( METR, Feb 2026 ). Their newer 800+ task dataset showed a much smaller -4% slowdown, with a confidence interval running from -15% to +9%. METR's own read: AI tools are probably better now than they were a year ago, and the gap between perception and reality is likely narrower. So the picture is moving. But the perception/reality gap is still there. Which leaves the question that won't leave me alone: how can we be even partially that wrong about something we do every day? I think it's because we're not measuring speed. We're measuring something else. And whatever that something is, it FEELS like speed, and it feels really, really good. What I actually feel when I'm "fast" It started small. I'd ask my coding agent to fix one thing, and while it was working, I'd line up the next. Then while THAT was working, I'd line up another. By the time I looked up, three hours had vanished and I had closed fifteen "small" tasks instead of the one I actually meant to work on. Each "task complete" gave me a tiny bump. A commit lands — bump. A test passes — bump. The agent writes a doc I didn't even ask for but it's exactly what I needed — bigger bump. By 4pm I'm wired, slightly twitchy, looking for the next hit. Sound familiar? It should. It's the same shape Netflix figured out years ago. The same shape TikTok perfected. And now? Now I'm running it on myself, voluntarily, for "work." A careful word about the brain stuff I have to be a little honest here, because the internet is FLOODED with "dopamine = X" pieces that oversimplify the science badly. I don't want to write another one of those. The short version: in the late 90s, Wolfram Schultz and colleagues showed that dopamine neurons fire for prediction error — the gap between what you expected and what you got — not for pleasure itself ( Schultz et al., Science , 1997 ). Later work showed dopamine ALSO ramps up gradually as animals approach a goal in spatial tasks ( Howe et al., Nature , 2013 ). Sounds neat. But it's not the only story. Kent Berridge and colleagues have spent decades arguing for a different mechanism — incentive salience, the distinction between "wanting" and "liking" — and the field hasn't settled which framework wins where ( Berridge, 2007 ). Even the ramp finding gets interpreted differently by different labs. So I'm NOT going to tell you dopamine "causes" addiction to AI. That would be the lazy pop version. What I AM going to say is more careful: the workflow I just described — uncertain-but-frequent successes, short distances between "start" and "done," many small closures per hour — happens to produce the kind of stimulus pattern that reward circuits respond strongly to, regardless of which theoretical framework you prefer. That's not a strong claim. It's just enough to take seriously. The TikTok parallel (with a footnote) TikTok was DESIGNED to exploit this. Engineers and product folks at content platforms know exactly what they're doing — variable ratio reinforcement, infinite scroll, autoplay. Behavioral addiction researchers have been documenting it for years ( Flayelle et al., 2018 ). AI coding tools, writing tools, agents — they weren't designed for this. They were designed to produce useful output. The intent is completely different. I want to be clear about that. BUT — and this is the part I keep tripping over — the structural shape of the workflow ends up looking similar. Most prompts produce meh results. Some are brilliant. Some are absolute garbage. I never know which I'm getting. So I keep prompting. The whole "vibe coding" phenomenon that exploded across dev Twitter in 2026 is the natural endpoint of this loop: lean all the way into the structure, let the model drive, ride the hits. Same shape. Different intent. An accident is still an accident with consequences. And here's where it gets quietly uncomfortable So far we have: a workflow whose structure happens to light up engagement circuits. That alone isn't a scandal. But there's another layer that the empirical literature is starting to nail down, and once I saw it I couldn't unsee it. Microsoft researchers led by Hao-Ping (Hank) Lee published a CHI 2025 study based on 319 knowledge workers and 936 first-hand examples of GenAI use at work ( Lee et al., CHI 2025 ). The finding: AI usage SHIFTS the nature of cognitive effort. Specifically, it moves people "from problem-solving to incorporating the AI's output, and shifts other types of higher-level thinking to merely stewarding the AI." Workers reported "much less effort" on knowledge and comprehension tasks in roughly three quarters of examples. And — the part that really got me — workers with HIGHER confidence in the AI tended to invest LESS critical thinking effort. The more you trust it, the less you think. A separate 2025 cross-disciplinary survey of 666 participants found a strong negative correlation (r = −0.68) between AI tool usage and critical thinking scores, with cognitive offloading as the proposed mechanism ( Gerlich, 2025, MDPI Societies ). These line up with a theoretical framing from Eduardo Junqueira at the Universidade Federal do Ceará: what he calls precarious agency ( Junqueira ). When you spend your day correcting, rephrasing, or accepting AI output, you THINK you're driving. You're the human. You decide. But you might actually be exercising a much smaller, more algorithmically-shaped form of decision-making than you realize. The reward circuits can't tell the difference between "I built this from scratch and learned something" and "I accepted what the model gave me, hit run, and it worked." Both fire. Both give you the hit. So I'm not just hooked on progress. I'm hooked on a flavor of progress where I'm not entirely sure how much of the thinking was actually mine. And it's costing us — real, measurable cost This is where "pace yourself" stops being soft advice and starts having evidence under it. In March 2026, BCG and UC Berkeley researchers published a Harvard Business Review study based on 1,488 full-time US workers across multiple industries. Their finding: 14% of workers reported a new pattern of mental fog, headaches, and slower decision-making the authors named "AI brain fry" ( HBR, March 2026 ). The drivers, specifically: supervising multiple AI agents in parallel, constantly verifying outputs, juggling several tools at once. Workers in marketing (25.9%), HR (19.3%), operations (17.9%), and software engineering (17.8%) reported it at the highest rates. Workers experiencing it reported more mistakes, slower decisions, and higher fatigue. The same study found a paradox worth sitting with: AI offloading repetitive tasks actually REDUCED stress. AI piling on supervision-of-AI work INCREASED it. So this isn't "AI bad." It's "the agentic supervision workflow is harder on us than people realize." Stack it all up: you FEEL fast (METR says you might not be, especially a year ago — and the gap may be closing but is probably still there), you FEEL like an author (Lee et al. and Gerlich say you're closer to a steward, and the more you trust the AI, the less you think), and the agentic workflow itself produces measurable mental fatigue that didn't exist a few years ago. That's not vibes. That's converging research streams pointing at the same place. So… what now? I'm not writing a 7-step plan to "beat AI addiction" because that's the wrong frame. The point is honesty about what's happening: I noticed it. The brain is harder to play once you know what's being played. I'm pacing. Some days I work AI hard. Some days I close the agent at 2pm and actually think — with a notebook, like an animal. The hit-free hours are where the original ideas come from. I check my agency at the end of each session. Did I actually decide anything today? Or did I just edit what the model decided? If I can't answer that, I worked too fast. I'm watching what I supervise. Three agents at once feels productive and IS measurably worse for me. Two felt fine. The HBR study put a number on what I was already noticing in my own week. I'm comparing this to physical health, not output. A runner who sprints every day blows out their knees. A progress junkie who chases hits every day blows out their judgment. Pacing isn't laziness — it's the only way this is sustainable. The hardest part is admitting that the FEELING of velocity is its own reward, separate from whether velocity is what I actually need. Sometimes I need slow. Sometimes I need to sit with a problem for a day. The reward circuits don't care. They want the next ramp. Knowing that is the whole game. Pace yourself for health. Not for output. So am I the only one on this ride? Or are you also closing your laptop at 6pm feeling vaguely electric and slightly wrong? Engage me on X or LinkedIn. This article was lovingly written by a living, breathing, warm-blooded human. Yes, AI helped fix typos and grammar. No, it did not write the thinking. \ \