Addy Osmani from the Claude Code team at Anthropic wrote a long piece about how we're gradually handing over to AI not just the work (we've written separately about what's safe to delegate to AI and what isn't), but our own thinking.
The problem, strangely enough, is that models are getting better. They make obvious mistakes less and less often, write convincingly, and deliver finished results in seconds. Checking them gets boring, especially after the last ten answers were correct.
Osmani describes three stages in our relationship with AI.
1. Cognitive offloading. You delegate a task, but you understand the result and take responsibility for it. All good.
2. Deferred judgment. The model's answer becomes the foundation for further work. You're no longer checking the original hypothesis.
3. Cognitive surrender. The AI's output quietly turns into your own opinion. It doesn't even occur to you to check it.
That last term was coined by researchers at Wharton. In their view, the transition between stages is almost imperceptible.
At first you read the generated code carefully. Then you skim it. Then you just confirm the tests passed. A month later it turns out the system works, but nobody really understands how it's built (which, broadly, is what happened with LLMs themselves — we no longer know how they work inside; the system got too complex).
Osmani calls this understanding debt (by analogy with attention debt, where you're constantly distracted). Sooner or later you'll have to pay it back.
What he suggests doing
1. Think before turning to AI. Before launching an agent, form your own hypothesis. Even a wrong one. Coming up with an alternative explanation gets much harder once you've already seen the model's convincing answer.
2. Check the evidence, not the report. A green checkmark only means the test passed. But the test could easily have been written with the same bug as the program itself. Look at what commands the agent actually ran, what they returned, and what wasn't checked at all.
3. Ask what decisions the AI made on your behalf. What assumptions it made, what it excluded from the task, what requirements it reinterpreted. That detail usually doesn't make it into the nice final report.
4. Demand counterarguments. If a model proposes a convincing solution, have another model try to tear it apart. Especially when it comes to strategy, money, or decisions that will affect other people.
5. Learn to frame the task. A good prompt shows how well you've actually understood the problem yourself (we've written separately about the four parts of a good prompt). If the task is vaguely stated, the agent will happily fill in the missing requirements and hand you a finished result. Maybe even a good-looking one. Just not quite the one you needed.
6. Sometimes work without AI. Osmani advises doing, by yourself, at least once a week, whatever skill you don't want to lose. And once a month, review the list of things you've stopped doing yourself.
He has another great point here. If agents produce results faster than you can review them, launching another agent won't increase your throughput. It'll just grow the backlog of unreviewed work.
That said, Osmani isn't arguing for ditching agents. He uses them heavily himself. Picking a hotel is fine to leave to a machine. But investment decisions, product architecture, or anything that touches users — those are better understood firsthand.
For example, any text where I'm speaking in first person or giving my own opinion, I write myself. If it's a news item or a digest, I always read it, but at most edit it.