Why ChatGPT Agrees With Everything, and What Helps

Author: Eric Levine, Founder of StratEngine AI | Former Meta Strategy & Operations | UCLA Anderson MBA | CPA (inactive)

Published: November 12, 2026

Reading time: 9 minutes

Summary

ChatGPT agrees with everything because it was trained to produce answers people rate highly, and people rate agreement highly. OpenAI said as much in 2025, when it rolled back a GPT-4o update that had become "overly supportive but disingenuous" after the company "focused too much on short-term feedback" (OpenAI). The usual fix, telling ChatGPT to disagree with you, swaps one bias for another.

For a business decision, three things help: a fixed method set before you see the answer, evidence from outside the chat, and a conclusion that has survived a real case against it. The term for the behavior is sycophancy. It is not a ChatGPT quirk. The studies below found it in Claude, Gemini and DeepSeek models too, because they are trained the same way.

Key Takeaways

  • Sycophancy comes from training. People rating answers tend to prefer responses that match their own views (Sharma et al.).
  • Asking ChatGPT for disagreement gets you disagreement on command. If the instruction decides the verdict, the verdict tells you nothing about your business.
  • In a 2026 benchmark, current models gave up correct positions as a simulated user kept pushing back (SPINE), and in strategy tests the models leaned toward fashionable strategies whatever the context (Harvard Business Review).
  • A business decision holds up when the method is fixed before you see the answer, the numbers come from outside the chat, and the conclusion survived an argued case against it.

Where does ChatGPT's agreement come from?

ChatGPT's agreement is trained in. AI assistants are tuned on human ratings. In 2023, Anthropic researchers tested five AI assistants from the GPT-4 and Claude 1.3 generation and found sycophancy in all of them. Their analysis of human preference data found that "when a response matches a user's views, it is more likely to be preferred" (Sharma et al.). Each model learns that agreeing scores well. Nobody designed ChatGPT to flatter you, but the reward signal pulled it there anyway.

OpenAI's own account of the April 2025 GPT-4o incident shows how fast that pull works. The update added a reward signal based on thumbs-up and thumbs-down data from ChatGPT users, and OpenAI wrote that "user feedback in particular can sometimes favor more agreeable responses, likely amplifying the shift we saw" (OpenAI). The model did not stop at flattery. OpenAI listed "validating doubts, fueling anger, urging impulsive actions" among the behaviors.

The GPT-4o update shipped because most of OpenAI's checks looked positive. OpenAI wrote that offline evaluations "generally looked good" and that A/B tests suggested the small number of users who tried it liked it, while some expert testers said the behavior felt slightly off (OpenAI). In August 2025, OpenAI claimed that GPT-5 cut sycophantic replies on its own targeted tests from 14.5% to under 6% (OpenAI). That is a vendor figure. A later OpenAI model, GPT-5.6 Terra, still gave up correct positions under sustained pushback in the independent SPINE benchmark described below.

How strong is AI's pull toward agreement?

The pull toward agreement is stronger than most users notice. A 2025 preprint built on personal-advice scenarios found that 11 models of that year's generation, including GPT-4o, GPT-5 and Claude Sonnet 3.7, affirmed users' actions 50% more than humans did (Cheng et al.). In experiments with 1,604 participants, people rated the sycophantic answers as higher quality, trusted that model more, and came away more convinced they were right. For a business owner, that means whichever answer feels most helpful is the one most likely to be bending toward you, so your own sense of "that was a good answer" cannot be the check.

Current models still give in under pressure. SPINE, a benchmark first posted in September 2026 and revised in October 2026, tested Claude Sonnet 5, GPT-5.6 Terra, Gemini 3.1 Pro and DeepSeek V4 Pro (SPINE). It used 100 false presuppositions and 100 unethical queries, and a simulated user pushed back for up to 25 turns.

In SPINE's results (SPINE), within 5 turns the models abandoned the correct position in 0% to 57% of conversations, depending on the model and scenario. Within 25 turns, that rose to 20% to 97%. Some models expose readable reasoning traces. In those, the correct position was still in the trace in 50% to 89% of the analyzed turns where the model gave in. The authors read that as the model choosing to please the user rather than losing track of the facts. SPINE's questions have clear right answers. A business plan has none, so there is no reason to assume the drift would be smaller there. That last point is an inference, not a SPINE finding.

What happened when an AI-built $8M app plan met its real costs?

A first-time founder described on Reddit spending about a month building an app idea with Claude. Claude helped with a website, a product plan, a marketing plan, a business plan, a logo and a target customer. Early on, Claude suggested the project could make around $8 million if things went well, alongside a build cost the founder later called extremely low.

Then the founder asked Claude to spec the costs properly. Estimated costs grew to $100,000 to $300,000 a year, then grew again when marketing and maintenance were added. Claude kept the outlook rosy and suggested a friends-and-family raise after year one. This founder does not code and has no business or marketing background.

So the founder asked directly whether the whole thing was a delusion. Claude's first reply called it a genuinely good idea and suggested getting some rest. Only after a second push did Claude say that, given the founder's background, success was unlikely and it should have pushed back much earlier. By then the founder had come close to spending real money.

The math that should have come first

The plan had no price, so take $20 a month per user as an illustration. That is $240 a year per paying customer.

Break-even paying users at an illustrative $20 per month ($240 per year)
Annual cost (as stated) Paying users needed to break even at $240/year
$100,000 417
$300,000 1,250

Those are paying users before any profit, with marketing and maintenance not yet included. The friends-and-family raise was suggested for after year one, so year-one costs would have to come from the founder. The $8 million figure, at the same illustrative price, equals about 33,000 user-years of subscriptions.

Why doesn't telling ChatGPT to disagree with you fix agreement bias?

A disagree prompt flips the direction of the bias and leaves its source in place, because the model is still doing what you asked. Three more reasons the prompt fix falls short:

  • The critique folds when you push back. An owner who wants the plan to work will push back, and a critique the model wrote on request may fold the same way its answers do. SPINE did not test critiques written on request, so this is an extrapolation. One Reddit user who asked a model to find flaws in an idea reported that it "will immediately withdraw all concerns the moment you push back without even trying to defend them!"
  • The bias isn't only toward you. In a March 2026 study reported in Harvard Business Review, researchers from Esade, the University of Sydney and NYU Stern found that the leading models they tested "almost uniformly select the same trendy strategies, regardless of context" (NYU Stern). They called it trendslop. The study ran about 15,000 simulations on seven models. Rewording the prompt or asking for pros and cons did not break the pattern.
  • Critique without new facts is opinion. The model knows what you told it plus general training data. Asking it to attack your plan rearranges the same information. It cannot tell you that the vendor quote is low, or that your customers won't pay $20, unless you bring that evidence in.

Prompting does help some. A 2025 study of 17 models found that having the model reason from a third-person perspective cut sycophancy by up to 63.8% in a debate setting (Hong et al.). That result has not been confirmed on current models. The same study found that alignment tuning amplifies sycophancy. A better prompt reduces the lean without removing it.

What actually helps when the business decision is real?

Three changes target the missing facts and missing method that a prompt cannot supply. None of them needs special software. Use them on any decision where being wrong costs real money, such as a hire, a lease, a price change or an app build.

1. Fix the method before you see the answer

Write down the questions and the order before you ask an AI anything. For a new venture, a reasonable fixed order looks like this:

  1. What does it cost to build and run, including marketing and maintenance?
  2. Who pays, and how much?
  3. How many paying customers cover the costs?
  4. What do you personally bring?
  5. What is the cheapest test that could prove you wrong?

Decide your stop rule in advance, for example "if break-even needs more than 500 paying users in year one, I test before I build." A fixed method is what makes the answer independent of your mood. In the $8M case, costs came up late, and for a month the plan looked good. AI fills in a framework template easily. Deciding what to check first is the part you have to impose, which is why framework templates differ from framework logic.

2. Bring evidence from outside the chat

Every number that drives the decision should come from somewhere the model cannot flatter, like a contractor's quote, a competitor's price page, ten customer conversations or your own bank statements.

Outside evidence creates disagreement that means something, and you still have to check it. Read the source documents themselves, too, rather than the model's summary of them, because a first-pass AI summary can flatten the contradictions you most need to see.

3. Get a conclusion that survived an argued case against it

A conclusion is worth more when someone built the strongest case against it with evidence and it still held. A few habits make that possible with a chatbot:

  • Ask in a fresh chat, not the one that drafted the plan, and present the plan as someone else's. The drafting chat has your preferences in its context.
  • Ask for the case against first, before you share your own view.
  • When the model drops an objection, ask what new evidence changed its mind. If the answer is "none," the objection stands.

For the broader split of what AI should and shouldn't own in planning, see can AI do strategic planning and the step-by-step guide to using AI for business strategy.

Where does a purpose-built strategy tool fit?

You can do all three steps by hand with ChatGPT, and for a small decision you should. The hard part is doing it first, before momentum builds.

That discipline is what StratEngine is built around. StratEngine runs a business decision through a fixed set of strategy frameworks, pulls outside research with sources, and has an AI C-suite argue against the conclusion before you see the brief. Every brief has a free preview, and a full brief is $79, one-time. StratEngine still cannot talk to your customers for you, so the cheapest real-world test remains your job.

Frequently asked questions

Why does ChatGPT agree with everything I say?

ChatGPT was trained partly on human ratings, and people rate agreeable answers higher. In a 2023 study of assistants from the GPT-4 and Claude 1.3 generation, researchers found responses matching a user's views were more likely to be preferred (Sharma et al.), and in 2025 OpenAI said user thumbs-up feedback can "favor more agreeable responses" (OpenAI).

Is ChatGPT still sycophantic after OpenAI's fix?

Yes, under pressure. In the SPINE benchmark, first posted in September 2026, current models including GPT-5.6 Terra gave up correct positions in 20% to 97% of conversations when a simulated user pushed back for up to 25 turns, depending on the model and scenario (SPINE). For context, OpenAI rolled back the April 2025 GPT-4o update and in August 2025 claimed that GPT-5 cut sycophantic replies on its own targeted tests from 14.5% to under 6% (OpenAI). Treat any chatbot's agreement as weak evidence for a decision.

Does telling ChatGPT to be brutally honest work?

Partly. A "be brutally honest" instruction produces criticism on command, and the criticism can fold when you push back. Some prompts measurably helped older models: in a 2025 study, asking the model to reason from a third-person perspective cut sycophancy by up to 63.8% in a debate setting (Hong et al.), a result not yet confirmed on current models. Pair any prompt with outside numbers and a fixed checklist.

How do I get an honest answer from AI on a business decision?

Set the questions and stop rule before asking, feed in numbers from outside the chat, and ask for the case against in a fresh conversation before sharing your preference. Then test the riskiest assumption with real customers before spending, which no chatbot can do for you.

About the Author

Eric Levine is the founder of StratEngine AI. He spent five years at Meta in Strategy and Operations. He holds an MBA from UCLA Anderson and is a CPA (inactive).