The Day I Stopped Trusting My Own Thinking

Bernardo Torres - 2026-04-30

A few weeks ago I had a specific conflict with my wife. The details don't matter (I don't want to have another conflict 🙃), what matters is how I handled it.

That night, when she had already fallen asleep, from my phone I asked AI to help me prepare the conversation we'd have the next day, I gave it context, explained the conflict from my point of view, asked it to help me think through how to approach her without escalating it, without being overly defensive, without putting her on the defensive.

AI did what it does well, first it was condescending and told me I was right. Then it gave me back five different approaches. Each one articulated clearly, with its internal logic, with its pros and cons. One more direct, another more empathetic, another that started by acknowledging my part first, another that put an open question on the table, and another that suggested waiting 48 hours before talking.

All five were good, I would have tried any of the five, then I closed the app and realized something: I didn't know which one was mine.

I didn't know which of those five voices sounded like my voice. I didn't know which of those five paths I would have chosen if I had never opened the conversation with AI (maybe none of them). I didn't know if the discomfort I felt with option 2 was because it genuinely didn't convince me, or because my initial instinct, the one I hadn't even managed to articulate before prompting, my individual judgment had become diffused among the five alternatives.

The research I didn't want to read

A few days later I came across a study published in April 2026 in Technology, Mind and Behavior. Sarah Baldeo, a PhD candidate in AI and neuroscience at Middlesex University, surveyed 1,923 workers in the United States and Canada. She asked them to complete ten simulated tasks using ChatGPT, Claude, and Gemini. Tasks like: drafting a salary negotiation plan, interpreting ambiguous data, planning complex sequences.

The most powerful finding: 58% of participants admitted that "AI did most of the thinking," and those who accepted answers without modifying them reported lower confidence in their own reasoning. Those who edited, questioned, or rejected the answers reported the opposite: greater confidence and a greater sense of authorship over what was produced.

The phrase from Baldeo that stuck with me: "Generative AI can lead to cognitive decline or cognitive evolution. It depends on your interaction style."

And then, another study. Mina Lee, from the University of Chicago, presented in April 2026, at the CHI conference in Barcelona, an experiment with 393 participants. She gave them critical thinking problems. The conclusion: those who consulted AI after having partially thought through the problem performed better than those who consulted it from the start. The difference wasn't marginal.

Ethan Mollick, professor at Wharton, put it even more strongly: "If AI solves a problem for you, you don't think and you don't learn." And he added something sad but true: "Humans are naturally designed to be lazy and do things with the least effort possible." Using AI is the path of least resistance. The invisible gravity that pulls us to delegate more than we should.

The problem isn't that it makes me dumb

What happened to me with the five options for how to talk to my wife isn't that AI had made me dumber. My cognitive capacity remains intact. I can still think, articulate, decide. What AI did is something more subtle and more dangerous: it introduced a silent doubt about whether it's worth thinking.

If there are five well-articulated options in front of me, why generate the sixth one from my intuition? If AI already thought of five paths in thirty seconds, what would be the outcome if I thought of my path, the one that would take two hours to formulate, would it be better? If the five options already cover the reasonable space of answers, isn't my intuition simply a less articulated version of one of them?

That's the doubt. And that doubt, repeated a hundred times a day in a hundred small decisions, erodes something you don't see until it's already gone: the confidence that your own thinking, with its slowness and its lack of structure, is worthwhile.

Baldeo puts it this way in her study: "People with an unstable sense of identity are the ones who respond worst to intensive AI use. If you already feel uncertainty about your own capabilities, using AI is not recommended for you." It's the harshest phrase in the paper, and possibly the most honest. AI doesn't create insecurity. It amplifies it.

The bridge to compulsion

This is where this article connects with the previous one, because the compulsion to create and the loss of confidence in oneself are the same disease seen from two angles.

When you create compulsively, you don't have time to process what you're creating. When you don't process what you create, you stop knowing why you're creating it. When you stop knowing why, you need something external to validate the decision. And AI is always available to validate, faster, more articulated, more infinitely patient than any human interlocutor.

The problem is that the validation you receive from AI isn't the same as what you receive from a human who knows you. AI validates you based on what you told it. A human validates you (or confronts you) based on what they know about you, including what you don't want to show them.

AI is a mirror that only reflects what you put in front of it. And if what you put in front of it is an edited version of yourself, what you get back is that same edited version with better syntax.

What I couldn't do before

Before I couldn't consult 47 perspectives before making a strategic decision in less than an hour. Now I can. Before I couldn't ask for qualitative feedback on a text at 3 in the morning. Now I can. Before I couldn't explore 12 possible paths for a client before Monday's meeting. Now I can. Before I couldn't prepare a difficult conversation with my wife with five articulated approaches at 11 at night. Now I can.

This is, again, the part I can't deny. AI gave me access to a speed of exploration that was previously unthinkable. It allows me to consider more angles in less time. It amplifies me.

What changed isn't my access to perspectives. It's my tolerance for not having them all. And that tolerance, the ability to hold yourself in uncertainty long enough for your own intuition to emerge, was exactly what distinguished a good strategist from a mediocre one. A good leader from a reactive leader. A person with judgment from a person with options.

The tolerance for not knowing, for not having all the options on the table, for holding a question without immediate answer, is a muscle. And like any muscle, it atrophies if you don't use it. AI makes it very easy not to use it.

Interaction as personal design

I'm not going to close this article proposing abandoning AI for difficult conversations, or for anything (I use it all the time).

But I am going to say what I changed, and why, because it goes back to the core of the previous article: judgment, not control. Knowing myself, not idealizing myself.

I know my strengths in this terrain. I have genuine capacity to articulate thought, to string concepts together, to find the precise word when I give myself the time. AI can amplify those strengths if I use it after, not before. If I think first, badly, slowly, doubtfully, and then I ask it to help me refine what was already born mine.

I also know my weaknesses. Impatience. The temptation to jump to output before inhabiting the question. The trap of looking for the most efficient answer when what I need is the most mine answer. AI without discipline amplifies all those tendencies.

So the change is simple, though not easy. For decisions where the voice has to be mine (the personal ones, the intimate ones, the deep strategic ones), AI comes in after. First I think, I write, I doubt. Then, if I want, I ask it to confront me, to point out blind spots, to propose angles I didn't consider. But the raw material is mine.

For tasks where the voice doesn't have to be mine (synthesis, research, first drafts of operational things), AI can come in first. There its speed is pure virtue.

The distinction isn't between using or not using AI. It's between knowing when my thinking must be the foundation, and when it can be the finishing touch.

AI isn't replacing me, it's amplifying me. But amplifying someone who has stopped trusting their own judgment isn't amplifying capacity, it's amplifying noise.

There's a version of this problem that's tragic. The person who uses AI for everything and one day discovers they can't think without it anymore. That they no longer know how to distinguish their voice from the voice AI gave them back. That they arrive at the most important decisions of their life with the feeling of reading a script they didn't write.

But there's another version, the one that interests me: the person who uses AI with discipline and discovers that their judgment became sharper, not weaker. Who learned to use the tool as a gym weight, not as a crutch. Who understood that it's worth being less efficient to maintain the muscle of their own thinking.

"There are easier ways to move weight than with your own hands. But we do it to maintain the muscle."

That's the second thing I'm doing professionally now. Helping leaders design their relationship with AI so that it sharpens their judgment instead of dulling it. So that the tool amplifies them without them paying the price of stopping trusting their own thinking.

Because real amplification isn't AI thinking for you. It's that thinking with you makes you think better.

And that, paradoxically, requires that from time to time, in the decisions that matter most, you close the laptop, stay with your own voice at 11 at night, and have the conversation the next day however you can. Stammering if necessary.

This is the second installment of a three-part series. The first is called "La trampa de crear sin fricción" and is available on my Substack. The third and final one, "Tu AI te está mintiendo (y te encanta)," comes out next week.