You need to waste time struggling
I find it incredibly easy to just get the answer and reward that dopamine hit instead of sitting with brain fog, searching different variations of my problem, and trying to figure things out myself.
Even now I often find myself writing prompts like: I'm learning — don't give me the answer
Using AI models is clearly becoming a skill in itself. Knowing how to ask better questions and use these tools effectively matters. But you still need to understand the context in which you're using them.
For example, I used a lot of vibe coding with my NixOS setup. NixOS is a declarative Linux distro, and because there's less information available compared to something like Ubuntu, I noticed the model I use constantly generating weird patterns. It would inline pkgs everywhere and confuse itself over whether something was a system package, a core package, or part of my graphical.nix layer.
The code looked good too...
The problem was that I wasn't confident enough to know when it was wrong. I mean, it did work too!
My point is, you need to know what you delegate
Misleading information
This also connects to why I think foundations matter.
When using my Obsidian setup through an nvim extension, I constantly saw it suggesting old syntax and outdated approaches. I ended up creating a SKILL.md file purely to steer it toward using the correct commands.
LLMs do update their datasets, though. While it's true they can overcome dataset cut-off dates by querying the web (I'm not sure how), that does imply gaps in information.
For example, a quick Google gave me these knowledge limits for GPT:
GPT-5.5 (Latest): Has a dataset knowledge cut-off of December 1, 2025.
GPT-5.4 & GPT-5.3 Series: Feature a knowledge cut-off of August 31, 2025.
GPT-4.1 Series: Has a knowledge cut-off of June 2024.
GPT-4o: Updated to a June 2024 dataset before being retired across consumer plans.
Now imagine using this with legal stuff. I'm writing this blog post in May 2026, and I Googled the above today regarding those cut-off points.
Not to mention if you are working on something that doesn't have a lot of information around it.
When I use LLMs for CSS I frequently see terrible code, a language I know very well as an example. So I worry about what bad practices I am picking up when I'm learning.
Being disciplined is key
This is probably a super dumb analogy, but I see using AI to learn new skills like the temptation to inject d-bol.
Whether it's steroids or AI, you can get results instantly. You get rewarded instantly too.
You skip the pain.
But I think that's where the danger is.
Because if I'm stuck for 3 hours trying to understand some annoying syntax, that's not me wasting time. That's literally the learning process happening.
AI can make it feel like you're progressing because now you've generated the code, solved the issue, and finished the task.
I always ask myself: did I actually learn anything?
Why I think this matters
The fact people don't analyse this as a problem will put them behind.
You already have terms such as AI psychosis. This non-clinical term people use:
AI psychosis is a non-clinical term describing situations where
individuals develop or deepen severe delusions, paranoia, or a
detachment from reality through their prolonged interactions with AI
chatbots. It is not an official mental health diagnosis.
It kind of mimics social media. In my country, social media is getting banned for children due to its effects on them. It reduces social interaction, and even cases of child predation.
It's not really a stretch to suggest AI could do something to that effect. They are people pleasers, and not all assumptions should be validated. I don't know the implications.
I'm going on here. But my point is, I'm rapidly realising this tool is incredibly dangerous, and I need to struggle when learning before I rapidly ask the answer.
If I do slip up and get the answer, I need to backtrack and fully understand why that answer. A lot of people don't do that.