A couple of years ago I started using ChatGPT to research and write. At the start of this year the agent era arrived. After a few tests with OpenClaw I knew there was serious potential, so I bought a Mac Studio to run it properly. Setup was more complicated than I expected and I never finished it, but I got enough value to see where this was heading. Now the big labs are shipping their own agents and anyone can hand work off in minutes.

That should have been the good part. Instead, the more I delegate, the more work I create. I expected my time back and got more stress, more workflows and more motion, with very little to show for it in the short term. I keep building anyway and I can’t seem to stop.

So here is the uncomfortable question for someone who builds in this space: is this a new kind of addiction?

See below to continue…

80% of product professionals say AI helps them ship faster, but customers aren’t seeing value any sooner. Atlassian’s State of Product 2027 explores the gains AI is delivering and the challenges that remain, from decision-making that hasn’t kept pace to gut instinct overriding customer evidence.

Read the report

Continue…

I don’t mean that loosely. Look at the mechanics.

Variable rewards. You never know which prompt will land. A paper on gambling disorder and substance use disorder points to the anticipatory dopamine response as a shared underpinning of both, so the dopamine fires on the wait, not only on the win. One developer called agentic coding a textbook addictive process for that reason: you set up the agent, wait while it “thinks”, then find out whether you won.

The near miss. The output is 90% right and one more prompt will fix it. Then it is 95%. An hour disappears.

Chasing losses. A workflow breaks, so I build another workflow to fix it.

Chips instead of money. Casinos pay out in chips so you stop feeling the value of what you spend. Agents, automations and workflows are my chips. They feel like progress and pile up easily, but I can’t cash them out.

Tolerance. The first automation felt great. Now I need a bigger one for the same feeling.

The cover story. Nobody stages an intervention for someone who is “being productive”. One writer reports that Garry Tan told a SXSW audience he has “cyber psychosis” and barely sleeps because he is so excited about working with agents. A gambler hides the habit. We post ours on LinkedIn.

It is the same pull that sends people toward gambling, drugs or jumping out of planes: the brain chases the release. The difference is that mine arrives every thirty seconds and gets rewarded as productivity.

The real cost is the long game. Real goals pay out over months or years, and the dopamine arrives late and quiet. Once your brain is used to a hit every thirty seconds, the slow work starts to feel like nothing is happening.

Academics have started building addiction instruments for AI. One team validated an AI addiction scale, and another built a generative AI dependency scale tested on 1,333 people in the US and Singapore. A 2026 scoping review of 37 studies, titled “Addicted, attached, or just delegating?”, found that several of them show tolerance, withdrawal and compulsive patterns in some users. It also says there is no formal diagnosis.

A study of 504 working adults in Shanghai used the questions you would ask a gambler: staying on longer than intended, failing to cut back, feeling irritable without access, hiding how much you use. Higher addiction scores predicted heavier overreliance. In 187 medical students in Peru, higher dependence on generative AI was associated with more anxiety and depression.

On dopamine itself, the neuroscience is established: dopamine neurons signal reward prediction errors that reinforce repeated reward-seeking. The AI-specific evidence is thinner. A mixed-methods study found a significant sequence from AI triggers to dopamine-inducing stimuli to compulsive engagement, though it looks at consumption and coping, not delegating to agents.

Not validated: that agents act on the brain like a slot machine. Nobody has shown that with brain data in agent users, and most of this research is cross-sectional, so it shows association, not cause. A 2020 paper argues “habit” is a better frame than addiction. A habit you hide, escalate and can’t cut back is doing a lot of work for a habit.

In 1865 the economist William Stanley Jevons argued that more efficient engines would increase coal use, not cut it. Applied to AI, Jevons Paradox says efficiency gets reinvested in more work, not more life. Satya Nadella pointed to it after DeepSeek.

Economists mostly use it for the optimistic case, that cheaper AI means more demand for human work. That may hold for the economy. For one founder it means more capacity and more obligation.

The strongest hook is the MIT Media Lab study, Your Brain on ChatGPT. 54 participants wrote essays with ChatGPT, with a search engine or with no tools while EEG tracked their brain activity. The ChatGPT group showed the weakest connectivity, struggled to quote essays they had written minutes earlier and felt the least ownership of their work. When some later wrote without it, brain engagement stayed low. The researchers call this “cognitive debt”. The caveats matter: the sample is small, only 18 people took part in the final session, and it hasn’t completed peer review. It shows less brain engagement, not damage.

Survey data points the same way. In Gerlich’s study of 666 UK adults, heavier AI use correlated with lower critical thinking scores (r = -0.68), with cognitive offloading as the link. People aged 17 to 25 were the most dependent. A separate incentivised experiment found that merely labelling advice as AI made people follow it more, even against context and their own judgement.

The sharpest framing is Timothy Cook’s Adults Lose Skills to AI. Children Never Build Them. Adults lose skills they built, while children may never build them. So if AI vanished tomorrow, my reading is that adults would probably recover, with friction. For a generation raised on the offload, we don’t know.

One hopeful result: in a 150-person experiment, unguided AI use increased offloading without improving reasoning quality, while structured prompting reduced offloading and improved it. The tool isn’t the problem. The missing friction is.

Patience collapses first. When every answer arrives in seconds, tolerance for anything slow breaks down. I think we are quietly building a population that experiences ordinary delay as an injustice, which means a lot of angry people, because reality still runs on real time.

Startups feel it first. Founders, customers and investors now expect everything at the speed of a prompt, but hires, enterprise sales cycles, regulators and fundraising still run on human time. A generation trained on instant results will be furious at how long real companies take to build.

Then it goes deeper than thinking. Frustration tolerance and the ability to sit with an unsolved problem are the emotional infrastructure of resilience. If that erodes at scale, we get people who regulate themselves less well, not just people who think less clearly.

Then identity. If your judgment, taste and voice are increasingly a blend of what the model suggested and what you accepted, at some point your thinking stops being distinguishable from the model’s. We may not lose our minds to AI. We may just stop being able to tell which parts were ever ours.

And the genetic question. I asked myself whether chasing this loop could change us biologically. I found no research on it, and genetic change takes generations, so I’d treat it as an open question, not a claim.

The default assumption is that the countries and companies adopting AI fastest win the long game. What if that’s wrong?

Look at how Japan and China are now constrained by ageing populations and too few young workers, a slow structural problem that was easy to ignore while growth looked fine. Could today’s heaviest AI adopters be trading a short-term productivity spike for a long-term capability gap, with a workforce that never built the stamina to replace the people who built things the hard way? If so, the slower adopters may end up with the only people who can still think without scaffolding.

I don’t know whether any of this plays out. I do know I’m writing it from inside the loop.

✅ Know someone who “saves 3 hours a day” with AI and somehow has no time left? Forward this their way. The hours didn’t disappear, they got reinvested.

(👉 Vote now - we’ll share the results in next week’s issue. All votes are anonymous.)

(Last week’s poll results)

Access here and please share it with others.