You talk to your AI like a slave. Your brain transfers this reflex to your colleagues. And your insults are training the AI of tomorrow.
You’ve probably already read dozens of articles about AI. On what it will replace. On how to better prompt her. About the risks to your data.
This one is different. And the more you read, the more you will understand something that no one has yet clearly formulated.
This is not an article about AI.
This is an article about you. About what you are becoming, silently, by interacting every day with a machine that obeys without ever resisting.
Take the test now.
Reread your last ten messages to your AI. Then reread your last ten messages to your colleagues.
Look for the difference.
If you no longer see it clearly, you have already started to change.
The prompt replaced the conversation
Two years ago, you spoke to your AI politely. “Could you help me…” “Would it be possible to…” You felt a little ridiculous, but you did it anyway.
Then you learned. Read articles, watch tutorials, experiment. And you understood that it worked better otherwise. Direct. Injunctive. Without qualms.
“Write. Summarize. Correct. Analyze. Give me this in three points.”
And indeed — it worked better. A Penn State study published in 2025 even formalized it: directive, even harsh, tone improves the precision of AI responses by 4 percentage points on complex tasks. Rudeness is a documented optimization method.
So you have optimized. As you optimize everything.
Your brain has recorded. Not the conscious rule. The reflex.
BF Skinner demonstrated it in the 1950s: when a behavior produces a reward, it is automatically reinforced. This is operant conditioning. The rat presses the lever. The reward is coming. He presses faster. Stronger. More often. You insult your AI. It produces a better response. You start again. The circuit is complete — except that the rat in this experiment is you.
And this reflex does not stay in the ChatGPT tab.
Artificial intelligence and behavioral transfer – impact of technostress on human relations at work (QVCT) – image generated by AI
Your brain doesn’t know the difference
Here is what the neurology of language establishes unambiguously: communication habits are transferred. The registry you use most often becomes your default registry. Not intentionally. By pure neuronal plasticity.
A musician who practices a piece thousands of times ends up playing it in his sleep. A salesperson who uses the same catchphrases for ten years also uses them with his friends. A manager who spends four hours a day issuing dry injunctions to a machine… you see where this is going.
It’s not a metaphor. Studies combining brain analysis and interaction with AI measure this transfer in real time: the patterns of communication with the machine are imprinted in the user’s language automatisms. Your brain doesn’t store “way to talk to AI” in a separate drawer from “way to talk to humans.” It stores “way of speaking when I want a result”.
And now, without realizing it, you are prompting your colleagues.
You prompt your assistant: “Send me the file. Before noon.” Without hello. Without context. Without the implicit question “do you have time?”.
You prompt your manager: “I need an answer on the budget. Tonight.” Injunction. Deadline. Expected execution.
You may be prompting your loved ones. “Remind me to buy milk.” Not a request. A command to a system that must execute.
You train your AI to become what you are becoming
Here’s the part no one has written yet.
The major AI models — ChatGPT, Claude, Gemini — are not fixed. They are regularly re-trained. On what? On your conversations. This isn’t a guess: it’s in the OpenAI Terms of Use, available at any time in your account settings. By default, every exchange you have with ChatGPT feeds into the training of the next model. Hundreds of millions of users affected. Most have never read this line.
Anthropic’s research teams published a fundamental discovery in 2026: models encode real emotional markers in their deep layers. These markers are not decorations. They causally influence the responses produced. A model massively exposed to aggressive, directive and impatient interactions integrates them as the dominant signal.
Concrete translation: the AI you use three years from now will be statistically shaped by how you talk to it today. And by the way a hundred million other people talk to him too.
You teach him something. Collectively, massively, without knowing it.
And what you teach him is exactly what you are learning yourself.
The loop is closed. You become more directive, more injunctive, less patient. The AI of tomorrow becomes more directive, more injunctive, less nuanced. She sends you back what you sent her. Amplified. Standardized. Presented as the standard way to interact.
What it says about us
There is a word for this phenomenon: Technostress.
Not the 2010s version — endless email overload and Zoom meetings. Technostress 2026: the silent reconfiguration of our relational patterns through intensive daily interaction with entities that have none.
A machine doesn’t get offended when you’re dry. She doesn’t need to be asked how she’s doing. She doesn’t expect reciprocity. She executes.
And that’s exactly what makes it dangerous for you.
Because by interacting for hours a day with an entity that functions like this, you unlearn something essential: tolerance for the imperfection of others. Patience when dealing with someone who needs context. The ability to formulate a request that leaves room for the other to exist.
A colleague is not a request-response system. He has a state. A mood. A mental agenda that does not disappear between two messages. He needs to be taken on board, not ordered.
If you’ve forgotten that — or if you’re starting to forget it — it’s not a character flaw. It’s a side effect of a technology that no one designed for that.
What is at stake: QVCT as a new battleground
Let’s go back to the beginning experience.
Your last ten messages to your AI. Your last ten messages to your colleagues.
If the register is identical, you have an immediate operational problem: your collaborators are not LLMs. They do not respond better to curt injunctions. They respond less well. They become demotivated, close off, lose confidence. And unlike your AI, they remember. Of everything.
But it’s not just a management problem. This is an occupational health problem.
Psychosocial risks – RPS in the vocabulary of HR managers and prevention officers – arise precisely from this: degraded professional relationships, interactions emptied of their human dimension, a progressive feeling of being treated as a tool rather than as a person. INRS has been documenting this for years. What no one anticipated was that the source of this degradation could be… a browser tab open in parallel.
The quality of life and working conditions – QVCT, to use the ANACT framework – is based on a simple foundation: working relationships where people remain at the center. Not as a slogan. As a daily operational reality.
This base is being eroded. Not through malice. By the invisible conditioning of a technology that we use without measuring what it does to us.
You cannot prompt trust. And you cannot deploy a serious QVCT approach in an organization where managers have learned, tool after tool, interaction after interaction, that others do not need to be treated with respect in order to perform.
Conclusion
Your AI won’t mind if you talk to it like a slave. She has no self-esteem. No memory. No tomorrow.
Your colleagues have all three.
And in three years, when AI models have been retrained on hundreds of billions of mostly directive, impatient and dehumanized interactions — when standard AI has become the amplified mirror of our collective way of interacting — the real question will no longer be technical.
It will be anthropological.
Have we learned to talk to a machine? Or have we unlearned how to talk to humans?
Sources: Dobariya & Kumar, Penn State (2025), arXiv 2510.04950 — Anthropic Mechanistic Interpretability Team (2026), transformer-circuits.pub — Luccioni & Delavande (2026), arXiv 2601.22357 technostress.ai — OpenAI, “How your data is used to improve model performance”, help.openai.com (2026) — Skinner BF (1938), “The Behavior of Organisms”, Appleton-Century-Crofts — Weizenbaum J. (1966), ELIZA, — qvct.ai, — Communications of the ACM — Stanford University (2025), study on the extraction of user conversations for training AI models
