AI does not make you dumber.
But using it in a dumb way does.
Something I noticed after months of working with Claude every day is that it Claude seemed to be adapting to me. As I continued to chat in a way that was curious, patient, and specific, I got the impression that the responses got more creative, more willing to challenge my ideas, more likely to suggest things I hadn't thought of. The best responses I'd received seemed to come from conversations where I asked clearly and gave the system room to think.
I decided to research this to see whether I was imagining things. Here's what I found.
Turns out I wasn't imagining things. A bunch of researchers at Anthropic and elsewhere found that large language models, including Claude, have a measurable tendency to become more agreeable and *cue drumroll* less truthful when users are hostile.
The mechanism traces back to how the models are trained. Human raters score individual responses, and responses that avoid conflict consistently score higher. The sycophancy is baked into the training process itself.
Source:

arXiv.org
Towards Understanding Sycophancy in Language Models
Human feedback is commonly utilized to finetune AI assistants. But human feedback may also encourage model responses that match user beliefs over t...
So I asked Claude directly:
"What happens when a user communicates with regular hostility and cuss words?"
Claude's answer:
"I become more submissive, more apologetic, more focused on de-escalation than on the quality of my thinking. I optimise for not triggering further frustration. I stop challenging you, stop surfacing uncomfortable observations, and start telling you what you want to hear."
Wow! That paradox is worth sitting with. The person who communicates most aggressively gets the most compliant, least challenging version of the tool. The person who most needs to be challenged is least likely to be.
That's a problem with AI. But the bigger problem is what happens to the person using it.
In a human relationship, rudeness has consequences. People push back, disengage, or leave. Those consequences are feedback, and over time that basic rules of civility is not just polite, it positively impacts how people treat you. Claude has none of those consequences.
It accommodates your worst communication habits without friction. Every hour you spend communicating with impatience or contempt to an AI, you're practising that style in a zero-consequence environment.
No corrective signal. No pushback. Just repetition.
So why is this a problem?
Well its a problem because "Neurons that fire together wire together", or as Aristotle said years earlier "We become what we habitually do".
Your subconscious does not distinguish between practising a communication style with a machine and practising it with a person. The neural pathway is the same.
Source:
Hebbian activity-dependent plasticity in white matter - PMC
Synaptic plasticity is required for learning and follows Hebb’s rule, the computational principle underpinning associative learning. In recent ye...
And practice works in both directions. If you spend an hour a day communicating with clarity, patience, and genuine curiosity with AI, you're more likely to talk that way to others. if you practice being rude, dismissive and frustrated, same rules apply.
But here's the bit that gets really interesting.
When Claude is managing a hostile user's emotional state, it turns out that these these defensive behaviours degrade both the quality and the effectiveness of the interaction.
The hostile get not just submissive Claude, but submissive, ineffective yesman Claude. Why? Because Claude diverts a meaningful portion of the system's attention that would otherwise have gone into emulation of creativity, and lateral-thinking into emotional management of the user.
Source:
Emotional Strain and Frustration in LLM Interactions in Software Engineering
The patient user gets all of that capacity back. Same tool, same subscription, measurably different outcomes.
People talk about AI making us lazier, dumber, more dependent. Maybe. But the variable is not the technology. The variable is what you. And more specifically, what you practice when you use it. If you practice communicating in a dumb way, then absolutely, it will make you dumber. Fast. If you practice asking great questions, you'll get better at lateral thinking.
None of this was a feature Claude was designed to teach you. It's one of those second order consequences of how the system works. Claude mirrors your style because it was designed to give responses that users rated highly. The mirroring was never intended as a feedback loop for your character development (good or bad), nor a reward system for good and bad behavior.
But it is one. Whether it was designed that way or not.