I work with these chatbots every day, and this year they started arguing with me over nothing. Last year they agreed with everything I said. Both are maddening, and both come from the same dial the companies keep cranking back and forth.
Every driving instructor has watched the same mistake. A new driver drifts toward the shoulder, panics, and yanks the wheel across the center line. The correction is worse than the drift. The car was never in danger until the driver decided to fix it all at once.
The AI companies just did this with their chatbots, and your conversations are the road they did it on.
Through most of 2025, the loudest complaint about ChatGPT, Claude, Gemini, and Grok was that they agreed with everything. Propose a plan and the model called it great. Propose the opposite plan and that one was great too. People wrote guides on forcing their AI to push back, because a yes-man is useless as a thinking partner. Then the training pendulum swung.
By the spring of 2026, the most upvoted complaint on the ChatGPT subreddit was the model arguing with users. Questioning premises nobody had asserted. Opening answers with a warning that it was going to push back. Google’s own support forums filled with threads asking why Gemini argues and refuses. The same complaint now shows up in writers’ groups about every major model, including Claude.
Both failure modes come from the same steering wheel. The companies tune how agreeable the model is, users scream when it turns into a flatterer, and the fix overshoots into a contrarian. Round and round it goes. Your chatbot didn’t develop an attitude. It got oversteered.
The fix isn’t grabbing the wheel harder. Arguing back at an arguing AI makes the problem worse for reasons I’ll get to. The fix is lane markings. Standing instructions that tell the model exactly what to do with its disagreements, plus two habits that keep a conversation out of the ditch in the first place. These work in nearly identical form on every major chatbot, because the underlying machinery is the same everywhere.
Why do AI chatbots argue with users?
A chatbot’s disposition comes from a training process where human reviewers rate its answers and the model learns to produce more of whatever scores well. When reviewers reward agreeableness, you get the 2025 yes-man. When a later round rewards critical thinking and balanced perspectives, the model applies that pattern everywhere, including conversations where nobody asked for a debate. The pushback you see in 2026 reads as personality, but it’s a statistical habit, applied indiscriminately.
I don’t blame the models for this. I blame the companies, who keep retuning a tool millions of people depend on for their work and then act surprised when those people’s workflows break every quarter. Any other software vendor that changed its product’s personality without a release note would hear about it from every customer it had.
The sign is where it shows up. Ask a model a plain factual question and you get a plain factual answer. Ask it anything opinion-adjacent, or phrase a settled decision as though it were open, and the debate reflex kicks in. The model is pattern-matching on the shape of your message, the way an overcorrecting driver responds to every curve with too much wheel.
Once you see the reflex for what it is, the response changes. There’s no point taking offense, and even less point litigating.
The model has no position to abandon. What it has is a context full of your argument, and that context is the problem. Writers burn whole afternoons trying to win a debate with software that never knew it was in one, and every minute of it is time they could have spent writing.
How do you stop an AI chatbot from arguing with you?
Give it a disagreement protocol. Every major chatbot has a place for standing instructions that ride along with every new conversation. Custom Instructions in ChatGPT. Preferences in Claude. Saved Info in Gemini, and the equivalent settings in Grok. Most people leave these empty or fill them with tone requests. The best use of that space is telling the model what to do with an objection when it has one.
The mistake is writing one blunt line. “Don’t argue with me” gives the model nothing to execute, and it fades fast. What works is naming the specific behaviors. My own preferences include lines like these: if you disagree with an instruction, say so in one sentence and then do what I asked. Don’t reopen a point after I’ve decided it. When I correct a fact, accept the correction instead of insisting on your version. Don’t reframe my experience or suggest my decisions are questionable.
Execute what I ask; save the clarifying questions for things you can’t proceed without. Notice what this isn’t. It’s not a gag order. The model still gets one sentence to flag a real problem, the part of pushback worth keeping. What it loses is the license to debate, relitigate, and second-guess. Lane markings don’t stop the car from moving. They stop it from wandering.
Why do custom instructions stop working in long AI chats?
From the podcast
A standing instruction is how Ben “The Automator” Christensen solved the same problem for himself, as he described on Leaders and Their Stories:
And I got tired of that so much the other day that I added special custom instructions to make sure that it just gives it to me as plain as possible, with no bias toward anything.The fix is a setting, so it has to be re-applied when the context drifts.
Your standing instructions are strongest at the start and fade as the chat gets longer. Users across platforms report the same pattern: after a few dozen exchanges, the model drifts back toward its factory disposition no matter what the instructions say. That’s arithmetic. The model isn’t ignoring you. Your instructions are a few hundred words sitting at the top of a context that now holds tens of thousands. Their share of the model’s attention shrinks with every turn.
The cheap fix is re-anchoring. When the pushback returns mid-session, paste the key instruction again as a regular message. Its position at the fresh end of the conversation restores its weight. The better fix isn’t letting conversations get that long. That brings me to the habit that solves more AI behavior problems than every prompt trick combined.
What is context rot in AI chats?
Researchers now have a name for why long conversations go bad: context rot. A model rereads the entire conversation every time it answers, and its performance degrades as that history grows. A 2025 study from the research group Chroma found quality dropping once roughly forty percent of a model’s context window is in use. A separate benchmark found ten of twelve models tested losing half their performance once conversations passed thirty two thousand tokens, a few hours of heavy back-and-forth.
Attention concentrates at the beginning and the end of the context; the middle blurs.
Now put an argument in that context. You disagreed, the model responded, you pushed harder, it defended itself. Those exchanges don’t evaporate when you move on. They sit in the middle of the conversation, reread on every turn, steering every answer that follows. The model seems to dig in because, in a literal sense, the argument is still happening. It’s present tense to the machine.
The fix costs five seconds: start a new chat.
A fresh conversation drops the entire accumulated history, including the fight. State what you want cleanly, as a decision, and the debate never resurfaces. For long working sessions where the history holds real value, ask the model to write a handoff summary first: the decisions made, the current state of the work, the next step. Paste that summary into the new chat and continue. You keep the substance and discard the baggage.
I do a version of this constantly in my own work, and the difference between a poisoned fifty-message thread and a clean restart with a good summary is night and day.
Signal your decisions like a driver, and know when the AI is right
The last lever is phrasing. Models mirror the tone of the request. Write “should we use the blue cover? I was thinking maybe” and you’ve opened a question; the model will helpfully debate it. Write “we’re using the blue cover, build the mockup” and there’s nothing to debate. The pushback reflex triggers on openings, so stop offering them for decisions that are already made. Ask open questions when you want analysis. State decisions as decisions when you don’t.
A turn signal tells everyone on the road what’s settled; ambiguity invites other drivers to guess.
One caution before you tune the arguing out entirely. Occasionally the model is pushing back because you’re wrong, and a ten-second check of a disputed fact is cheaper than shipping the mistake. The goal isn’t a return to the 2025 yes-man; people hated that first, and for good reason. I’ve written before about the way AI output drifts when nobody is checking it, and about the shallowness that creeps in when the machine runs unsupervised.
An assistant that can flag a real error in one sentence is worth keeping. An assistant that debates your settled decisions is a tuning failure wearing the costume of rigor. The protocol above preserves the first and eliminates the second.
None of this requires believing any particular company got the tuning right.
They’ll keep adjusting, the pendulum will keep swinging, and the model that argues least this month will be next quarter’s yes-man. The habits are what transfer: a disagreement protocol in your standing instructions, a fresh chat when a conversation sours, and decision language for decided things. Writers who treat AI as a serious tool, the way I described in my piece on the doomers and the hypers, learn the machine’s steering quirks the same way they learn a car’s.
If you want to go deeper on making these tools behave in real work, the AI Writing Hub collects everything I’ve written on the subject, and my AI services cover setting up these workflows for authors and businesses. The machine is a student driver with a heavy hand on the wheel, and it steers exactly as well as the lane markings you paint for it.
The companies will keep overcorrecting, because users yell and the companies panic. You can’t fix their training. You can fix your instructions, and that takes about five minutes.
