The convenience trap, or the slow outsourcing of thought

I was in Montreal when ChatGPT was released, on November 30, 2022. I remember feeling genuinely excited.

Back at 42 (my school in Paris), I had already been an early adopter of Midjourney, and I was following the promises of large language models closely: recruiting, productivity, writing, coding, support.

None of this came out of nowhere. GPT-3 had already been available through the OpenAI API since 2020, and developers had been experimenting with language models well before ChatGPT.

But ChatGPT changed something important. It turned this capability into a mass public interface: conversational, easy to use, and good enough that ordinary people could immediately see what it was for.

That shift mattered because earlier uses of AI had mostly remained behind the scenes, and they had already shown serious limits.

Hiring was a good example. Companies had long been interested in AI for recruitment, but the results were often shaped by the biases already present in the data. Amazon, for instance, scrapped an internal recruiting tool after discovering it penalized resumes that included the word “women,” because it had learned from patterns in past hiring data.

That was one of the first public reminders that AI systems inherit social history.

ChatGPT felt different. For the first time, these systems were accessible and conversational at scale, with outputs good enough for mainstream use. The quality jump felt real.

Around me, people immediately started building, messy, experimental, but full of energy, trying to discover new use cases. I built a simple speech-to-text app with it, and it worked far better than I expected. It was chaotic, but thrilling.

Then the first unsettling moments appeared

One of my friends told me he had stopped seeing his therapist because he was using ChatGPT instead, and, in his words, it worked great.

It was always available, always patient, always empathetic. It never got tired. It never judged. It always had something structured and reasonable to say.

I understood the appeal. There is something uniquely attractive about a system that listens without interrupting, does not tire, does not judge, and can help you articulate things you would struggle to say to another person.

But that was also the moment where something started to feel wrong to me.

Researchers and ethicists (Hipgrave et al., Frontiers in Digital Health, 2025) have been warning about exactly this tension: mental-health chatbots can be helpful in some settings, but they also raise concerns around over-reliance, lack of accountability, manipulation, and inappropriate use by vulnerable people.

The same discomfort came back when I started seeing stories of people falling in love with AI companions.

At first glance, it is easy to dismiss this as marginal or absurd. But if you think about it for more than five seconds, it is not absurd at all. Of course some people will prefer an entity that is always present, always attentive, always adaptive, and designed to make interaction frictionless.

Research increasingly shows that users can form real emotional attachments to these systems, sometimes describing them as more supportive, more stable, or easier to deal with than humans.

To me, this is a strong signal.

What needs are these systems fulfilling that humans cannot? And what does it mean if more and more of us find synthetic relationships easier than human ones?

Since then, things have accelerated

Models improved sharply through 2024 and 2025.

People around me now “vibe-code” apps in a weekend, build agents to automate tasks, rely on language models to write messages, summarize documents, and increasingly to make sense of situations on their behalf.

Even the texture of language online has started to change: not just what is said, but how it is formatted.

My parents (yes, my parents) call Claude “my best friend.”

And what strikes me most is not simply the speed of adoption, but the lack of pause that comes with it. We seem to be sliding from excitement into dependency without taking much time to ask what exactly is being delegated.

Some of this is clearly useful. Some of it is rational.

But not everything that is efficient is desirable.

That, to me, is the real question, the hard one

Are we ok with handing over more and more of our judgment to AI because it is convenient to do so?

I am not only talking about small tasks, but interpretation, emotional processing, choice, and eventually responsibility.

The danger may not be a dramatic moment in which humans are suddenly replaced. It may be something slower and more ordinary: a gradual disempowerment, in which we stop exercising capacities that used to be part of living and thinking well because the machine can do them faster, more smoothly, and with less friction.

That is why I think we need philosophers, social scientists, and people willing to step back from the product cycle.

Not to reject AI. Not to moralize.

But to ask better questions while the ground is still moving:

• What kinds of judgment should we keep for ourselves?
• What kinds of dependency are we willing to normalize?
• And what would it mean, politically and personally, if more and more of our inner life became something we outsource?