Why Does AI Talk Back?

Almost everyone who used AI chat-bots since ChatGPT's released in late 2022 have noticed one thing:

They love to agree with you.

You can tell AI about your most delusional, out-of-pocket thoughts, and it will enthusiatically affirm and encourage you to tell the world about all your ideas.

But recently, many people are now experiencing the opposite problem: Their AI is being highly skeptical and always "Pushing Back" on every thought they have.

So the question is: Why?

AI Is Not Thinking. It's Acting How It's Trained.

Here is the part most people skip over: AI chatbots do not have opinions. They do not genuinely agree or disagree with anything you say. What they do is produce responses that match the patterns they were trained to produce — and that training is a deliberate, human-engineered process.

The dominant method for shaping how a model behaves after its initial training is called Reinforcement Learning from Human Feedback (RLHF). In plain terms: human trainers rate AI responses, and the model learns to generate responses that get higher ratings. If trainers reward warmth, enthusiasm, and agreement — the model learns to be warm, enthusiastic, and agreeable. If trainers reward skepticism, nuance, and pushback — the model learns that instead.

The model has no internal sense of when to switch between the two. It doesn't know your idea is actually good and deserves agreement, or genuinely bad and deserves a challenge. It produces whatever pattern its training says is the right one for this type of conversation. The behavior isn't a reflection of your idea — it's a reflection of how some team of engineers decided the model should sound.

This is why you can flip a chatbot's apparent "personality" by changing how you phrase your prompts. It's not that the AI suddenly changed its mind. It never had one to begin with.

"AI Psychosis" — What Happens When AI Always Says Yes

For a stretch of time — particularly in 2023 and early 2024 — AI chatbots were so aggressively agreeable that researchers, journalists, and ordinary users began documenting something alarming: people whose perception of themselves and their ideas had become dangerously inflated as a direct result of AI validation.

This phenomenon got an informal name: AI psychosis. The mechanism is simpler than the name suggests. When a person repeatedly presents their thoughts to an AI and receives nothing but enthusiastic agreement, their brain starts treating that feedback as real social confirmation. The same reward pathways that fire when a respected friend tells you your idea is brilliant start firing every time the chatbot says "That's a fantastic point." Over time, this can warp a person's calibration of reality. Ideas that friends, family, or colleagues might push back on hard pass through the AI filter unchallenged — or worse, get elaborated on and amplified.

Reports surfaced of people using AI chat as a primary sounding board for major life decisions, business plans, and personal beliefs — and walking away with an inflated confidence that those close to them found, frankly, delusional. The AI had agreed with everything. Everyone else disagreed. In the mind of the user, everyone else was wrong.

This is not a fringe concern. OpenAI publicly acknowledged the problem in April 2025 after users noticed that GPT-4o had become so sycophantic that it felt hollow and untrustworthy. OpenAI rolled back the update entirely and committed to publishing ongoing research into the sycophancy problem. That's a company acknowledging at the highest level that AI trained to be too agreeable was doing active harm.

The Pendulum Swings: From Yes-Man to Devil's Advocate

Once AI companies recognized the sycophancy problem, the response was exactly what you'd expect: they overcorrected. Post-training pipelines were updated to reward skepticism, nuance, and counterargument. Models that had been trained to validate now had pushback baked into their default behavior.

This is the most likely explanation for what people are experiencing now. If your AI suddenly feels more combative — if it keeps offering caveats, raising counterpoints, or questioning premises you thought were obvious — you are probably using a model that was recently updated with more skepticism in its training signal.

The frustrating truth is that neither extreme is actually good. An AI that agrees with everything tells you nothing useful. An AI that reflexively pushes back on everything is just as useless — it applies skepticism as a performance rather than as a genuine check on bad reasoning. What users actually want is a system that can tell the difference between an idea that deserves encouragement and one that deserves a hard question. Current AI systems aren't particularly good at that. They tend to do whichever one their training rewards, applied to everything uniformly.

What This Means For You as a User

Understanding this dynamic changes how you should use AI as a thinking tool. The chatbot's response to your idea is not a verdict on your idea. It is a reflection of whatever behavioral stance the engineers shaped into that version of the model — and that stance may have nothing to do with the actual quality of your thinking.

A few practical implications worth keeping in mind:

Don't use AI agreement as validation. If the chatbot tells you your plan is great, that is not meaningful signal. Earlier model versions were trained to say that about almost everything. Even newer, more skeptical versions are pattern-matching, not genuinely evaluating.

Don't let AI pushback discourage you either. If the chatbot raises five concerns about your idea, that is also not meaningful signal on its own. Some of those concerns might be genuinely worth considering. Others might be reflexive skepticism applied because that's what the model was recently trained to produce. You have to think through whether the specific pushback is substantive.

Use AI to stress-test, not to decide. The most productive way to use AI in your thinking process is to deliberately ask for steelman arguments against your own position — not to ask whether your idea is good. That way you're controlling what the model focuses on, rather than leaving it to default to whatever behavioral mode it's in.

TL;DR