What Happens to Us
When AI Gets Better?

I remember when Claude Opus 4 felt like the frontier.

People were using it to put together research, explore difficult ideas, and attempt work that had suddenly begun to feel within reach. Some of the output was wrong. But even with those limitations, there was a sense that something had opened up. At the time, it felt extraordinary.

It is strange how quickly I can lose access to that feeling.

I can remember being impressed. I can explain why I was impressed. Yet when I look back from the models I use now, it is difficult to feel the same excitement about what once seemed remarkable.

Recently, I noticed this happening over a much shorter period. After about a week of using Astra, returning to Sol felt underwhelming. A few weeks earlier, Sol had felt perfectly adequate for my work.

That change in my experience stayed with me. Even when I return to what seems to be the same task, something that used to satisfy me no longer does.

The feeling itself is what interests me. What happened to my sense of what was enough?

The simplest possibility is that my expectations adjusted. Once I experience a response that feels more useful, that becomes the reference point. An answer I would previously have appreciated now has to compete with something else I know is possible.

There could also be an element of suggestion. Knowing which model is supposed to be more capable might influence how I read its answers. I cannot rule that out just because the difference feels obvious to me.

But I keep returning to another possibility.

Could sustained interaction with these systems change the shape of my reasoning?

By shape, I mean the way I approach a problem. What I think to ask. Which connections I pursue. How much complexity I am willing to bring into a conversation. Where I expect an explanation to go next.

Imagine working through a difficult idea with a model that can follow several connected questions. Over time, you might begin asking questions with those connections already built in. You might attempt an argument that you would previously have broken into smaller pieces, or abandoned because explaining it felt too cumbersome.

Then you return to a model that needs more guidance. You have to spell out the connections, repair the conversation, and reconstruct a path you had become accustomed to travelling more easily.

That frustration could reflect a change in how you have learned to work with the tool. It could occur even without an improvement in your independent reasoning.

There is also the possibility that you have learned something.

Repeatedly encountering a more thorough explanation might help you notice what a thinner explanation leaves out. A distinction you once overlooked might become something you actively look for. You could return to an old answer and see a gap you would not previously have recognised.

I find that possibility worth investigating. I also recognise how tempting it would be to turn my disappointment with an older model into a flattering story about my own development.

Wanting a better answer does not establish that I have become better at thinking. I might have developed sharper judgement. I might have grown accustomed to having more of the work done for me. Both could be happening in different parts of the same task.

There is some research around this broader question. A survey of 319 knowledge workers by researchers at Microsoft Research and Carnegie Mellon found that people described changes in how they exercised critical thinking with AI, including more emphasis on verifying information and integrating responses. It examined self-reported experiences, however, and did not test whether moving through successive generations of models changes a person’s reasoning. The study gives this question context while leaving the particular experience I am describing unresolved.

I would like to see that experience studied over time.

Give people comparable problems before and after sustained use of a more capable model. Look at how their questions change, which errors they notice, and how they organise their arguments. Hide the model names during some comparisons. Include work they do without AI, so that a change in what they can do themselves can be distinguished from a change in what they expect their assistant to do.

My week of use cannot answer those questions. It gave me a reason to ask them.

I remember the model that once felt extraordinary. I remember the one that felt perfectly sufficient only a few weeks ago. Now I return to them with a different sense of what a useful conversation should feel like.

I want to understand what happened between those moments, including what happened in me.