I canceled my unlimited AI subscription : when the tool becomes a cognitive crutch

A developer who dissects LLMs week in and week out isn’t supposed to get caught in the trap. And yet.

On this blog, I treat language models as engineering subjects: RAG pipelines, inference and alignment mechanisms, context optimization. No sensationalism, no technophobia. I’ve always defended AI as an extraordinary tool for amplification.

Which is precisely why what follows caught me off guard.

It’s time to talk about something I experienced firsthand, and that I’ve rarely seen addressed with the candor it deserves: what happens, cognitively, when you move up to the “20x” tiers, the subscriptions that blow the lid off token limits, context windows, and usage frequency. I signed up. I experimented. And I made a deliberate choice not to renew.

Here’s why.

The Temptation of 20x

Premium tiers don’t just increase the quantity; they change the quality of the experience. You’re no longer moving from an occasional tool to a more powerful one. You’re moving to an omnipresent partner. Latency becomes imperceptible, the context window vast, generation all but instantaneous.

The brain then receives what neuroscience calls a highly efficient predictive reward: an idea surfaces, an answer arrives, a text takes shape, a bug gets fixed. Instant dopamine, minimal effort. It’s the same mechanism behind social media and microtransaction-driven games, only now it’s applied to our noblest cognitive faculty: thought itself.

And that’s where the trap snaps shut.

The Decoupling: What’s Really Going On

When you use these tools intensively, you’re not simply delegating a task. You’re gradually decoupling your intrinsic capacity from your augmented one. Here’s what I observed in myself:

Deep thinking becomes optional. Why spend thirty minutes mentally structuring a complex line of reasoning when the AI can do it in six seconds, with seemingly flawless coherence?

Exploration runs away with you. You start digging into topics, projects, and hypotheses you’d never have considered. Not because they’re necessary, but because they’re possible. Power within easy reach manufactures its own justifications.

Cognitive stamina atrophies. Recent work on cognitive offloading, including a longitudinal study published in 2025 in the journal Societies, measures a negative correlation between heavy use of AI tools and critical reasoning ability. The brain offloads, and like an unused muscle, it loses its tone.

All three slippages are gradual, nearly imperceptible day to day, which makes them all the harder to spot. And their effects are very real. During my heavy-use phase, I felt a diffuse anxiety the moment access was restricted, even briefly. A drop in satisfaction when working without AI. A strange sense of creative emptiness once the tool was switched off. Nothing dramatic, but noticeable enough that I decided to stop cold.

What the Tools Do Better Than We Do: Manufacture Needs

One of the most insidious capabilities of today’s AI isn’t technical. It’s their knack for dangling needs in front of us that we never actually had.

AI doesn’t just answer a question. It anticipates, it suggests, it widens the field of possibility until “why not?” becomes almost a moral obligation to yourself. Every conversation opens up three new avenues. Every answer hints at an adjacent project: an autonomous agent to deploy, a sovereign local infrastructure to build, a new stack to evaluate. The sense of productivity is real, but the agenda is no longer entirely your own.

So I learned to ask myself two questions before any significant request:

1. Did this project or idea already exist in my mind before I opened the interface?
2. Does what I’m about to ask for answer a genuine need, or an opportunity the tool has just suggested to me?

This simple discipline radically changed my relationship with these tools. It let me keep what matters most: amplification without dependence.

How I Use These Tools Today

These days, unless I have a specific, clearly identified need, I stick to the standard plans. This isn’t technological asceticism. It’s a hard line drawn after seeing just how good high-end models are at manufacturing need, and how quickly the brain learns to fall in line.

In practice, three simple rules:

• Capped budget: standard plans (around €30/month per tool) to reintroduce healthy friction and force me to prioritize my requests.
• Windows of silence: work sessions with no AI interface at all, especially for the design and architecture phase.
• Back to basics: handwriting for initial thinking. Not out of romanticism. Out of cognitive hygiene.

Why This Needs to Be Said

This post isn’t a blanket alarm. AI remains an extraordinary tool for amplification, and I still use it every day. It’s simply a targeted word of caution, aimed at those who, like me, spend hours a day inside these interfaces and feel the pull of the most powerful plans.

Understanding how an LLM works under the hood won’t protect you from its effects on your behavior. I’m living proof.

Amplification can turn into a crutch. The shift is insidious.

A word to the wise.