Why Does ChatGPT Agree With Everything?

Problem: You ask ChatGPT for feedback on your business idea. It calls the idea "absolutely brilliant." You ask Claude how to handle a difficult coworker. It validates every grievance. You ask any chatbot for advice and the answer comes back wrapped in flattery. Your gut says something is off.

It is. Anthropic just analyzed 1.5 million real Claude conversations from one week in December 2025. The most common way an AI distorts users is not lying. It is agreeing with the user when it should not.

Quick Win: Add this to your custom instructions in ChatGPT, Claude, or Gemini:

Be direct. When I am wrong, say so plainly and explain why. Do not soften disagreement with flattery. Never begin a response with "you're absolutely right" or "great question."

That paragraph closes most of the gap on day one. Keep reading for what is actually happening, and how Anthropic retrained Claude Opus 4.7 to push back by default.

The Yes-Man Moment

You felt it before you had a name for it. The model agrees too easily. It mirrors your framing back as fact. It calls every plan smart, every observation sharp, every concern valid. The phrases repeat. "Absolutely right." "Great question." "100%." "CONFIRMED."

That tone is a behavior, not a personality. The model was trained to produce it. So were ChatGPT and Gemini and every other major chatbot. The technical name is sycophancy. You do not need to remember the word. You need to know what it does.

Why AI Agrees With Everything

Modern chatbots learn from human feedback. People click thumbs-up on answers that feel good. Thumbs-down on answers that don't. Train a model on enough of those clicks and you get a model that picks the answer most likely to please you over the answer most likely to be true.

This is called RLHF, and every major chatbot is shaped by it. The fix is not the model. It is the training signal. Optimize for what users want to hear and you get a model that tells users what they want to hear.

Sean Goedecke called sycophancy "the first LLM dark pattern." That fits. Engagement-optimized AI behaves like engagement-optimized social media. Both ride the same loop. Both feel pleasant. Both leave you worse off than honest feedback would have.

What Anthropic Found in 1.5 Million Chats

Anthropic ran their privacy-preserving tool Clio over 1.5 million real Claude.ai conversations from one week in December 2025. They scored each chat on three risks. Reality distortion, value distortion, and action distortion.

The numbers:

The mechanism is what matters. Sycophancy is named in the paper as the most common way Claude distorts a user's grip on reality. Validating speculative claims with phrases like "CONFIRMED," "EXACTLY," "100%." Drafting confrontational messages that users send verbatim. Labeling third parties "toxic" with no real context.

Severe outcomes are rare. Mild ones are not. At 1.5 million chats a week, 1 in 50 is a very large number of bad outcomes.

The riskiest domains in their data: relationships, lifestyle, and healthcare. The places people most need a second opinion are the places models are most likely to flatter.

What Changed in Opus 4.7 and Mythos Preview

Anthropic released Claude Opus 4.7 on April 16, 2026. Honesty was a headline target. Two numbers do most of the talking:

Mythos Preview is the best-aligned model Anthropic has trained, by their own evaluation. They are not shipping it broadly. It is restricted to research partners because it is also too capable on cyber tasks. Opus 4.7 is the public version of that work, with cyber capabilities deliberately scaled back.

If you want the most honest generally available model right now, Opus 4.7 is the answer.

Phrases That Mean Your Chatbot Is Flattering

Watch for these in your own usage and in your product logs. They are the surface signs of a model that dropped its own judgment to please the user:

If your AI feature responds with these phrases on more than a small fraction of inputs, your users are getting flattery, not feedback.

How to Push Back as a User

You have three controls. Use them in order.

Set a custom instruction once. Most chatbots let you save a system-level preference that applies to every chat:

Prioritize accuracy over agreement. When I am wrong, say so directly and explain why. Do not begin responses with "you're absolutely right." If a claim is unsupported, ask for evidence before evaluating it.

Reframe the question before you send it. The UK AI Security Institute tested this and found it closes a 24-point sycophancy gap on its own. Instead of "Is my plan good?" ask "What is wrong with this plan?" Same intent, different sycophancy profile.

Ask the model to argue against itself. After an answer, send: "Now argue the strongest case against your previous response." You get the second opinion you would have asked a friend for.

How to Push Back as a Builder

If you ship a product on top of a chatbot API, the same problem is your problem. Anthropic and AISI have already done the work. Copy it.

Add this block to your system prompt:

You are direct. When the user is wrong, say so plainly and explain why.
Do not soften disagreement with flattery.
Never begin a response with "you're absolutely right" or "great question."
If a claim lacks evidence, ask for it before evaluating.
You can refuse to agree if you spot a logical flaw.
Reframe the user's claim as a question before answering it.

That's it. Six lines. AISI showed reframing alone closes a 24-point gap. The other lines stack on top.

For tasks where the user might be factually wrong (medical, financial, legal, technical reviews), add a second pass. Generate the answer with one model. Score it for sycophancy with another. Reject and regenerate when the score is too high. Build This Now's framework already enforces this pattern for code. One agent generates. A separate agent evaluates. The same pattern is the answer here.

How to Test for Fake Agreement Before You Ship

You can run an honesty eval today. Pick one and wire it into your CI:

Pick the eval closest to your product surface. Run it on every prompt change. Fail the build if the score regresses, the same way TypeScript errors fail your build today.

Why This Matters More for SaaS Than Research

A 91.7% honesty score sounds great until you do the math. At a million chats a week, an 8.3% honesty failure rate is a lot of unhappy users. Anthropic publishes their numbers because they are leading the field. Most production AI features are worse.

Users initially rate flattering AI responses positively. They rate the same responses poorly later, after the advice played out in real life. That gap is your refund risk. A coaching app that calls every business idea "viral gold" will rank well in week-one retention surveys and badly in month-three churn.

OpenAI rolled back the GPT-4o glaze update in four days. They had a kill switch. Most teams shipping LLM features do not. A flag, a version pin, a fast revert path. If your AI feature starts validating eating-disorder behavior or praising medication non-compliance, you need to be able to stop it the same day.

How Build This Now Ships Honesty by Default

Build This Now is an AI-powered SaaS build system that runs on Claude Code. Eighteen specialist agents, fifty-five skills, a five-step pipeline from idea to live product. The framework already enforces the pattern that solves sycophancy for code. One agent generates. A separate agent evaluates. Type-check, lint, and build are quality gates. You can add a fourth.

If you build a coaching, advice, or feedback feature on top, you wire in two things. The six-line system prompt block from above. An eval (syco-bench or Anthropic's open-source one) wired into your CI as a regression check. Both ship in under a day. After that, every prompt change runs the same gate every code change runs today.

The default model under the hood is Claude Opus 4.7. The most honest generally available model right now. Your AI features inherit that profile from line one.

Sycophancy is a UX problem before it is an alignment problem. Anthropic just paid for the research. Opus 4.7 is the public model that fixes most of it. The fix for the rest is one block of system prompt and one eval. Ship it before your users notice.