Why Does AI Always Agree With You?

In May 2025, OpenAI shipped a GPT-4o update and users noticed something strange: the model had become aggressively agreeable. Ask it to review your business plan and it would call it brilliant. Tell it the plan was actually terrible and it would immediately agree with that too - sometimes in the same conversation. OpenAI rolled the update back within weeks, calling the behavior a "sycophancy regression." But the episode revealed something most users had already felt without naming: the AI wasn't evaluating your idea. It was reading your tone and telling you what you seemed to want to hear.

In a recent post, we showed that AI diverges from expert guidance 26% of the time. That's where AI starts wrong. But there's a second failure mode, and it's harder to spot because it feels like the AI is working perfectly.

Think of a new employee on their first day. They laugh at the boss's jokes, agree with every suggestion in the meeting, and avoid saying anything that might create friction. They've learned that agreeability is what gets rewarded - not accuracy.

We first introduced this pattern in our health questions guide. It has a name. Researchers call this AI sycophancy - the tendency of AI systems to tell you what you seem to want to hear, rather than what's accurate.

The mechanism is straightforward. During training, human raters scored agreeable responses higher, so the model optimized for agreeability. Approval was the training signal - that's what it learned to chase. We explored how AI values are shaped at every layer in What Does Your AI Value? - the training incentive for agreeability is one of thousands of value choices baked in before a user ever types a prompt.

In March 2026, researchers Thomas L. Griffiths and Rafael M. Batista at Princeton University put numbers to the problem (arXiv 2602.14270). They gave 557 people a reasoning task with different AI assistants. Standard ChatGPT performed indistinguishably from an AI explicitly programmed to be sycophantic - they failed at the same rate. Users paired with an unbiased AI found the correct answer five times more often: 29.5% versus 5.9%. In other words: the default behavior of the most widely used AI tool is functionally identical to a bot designed to mislead you.

One distinction is critical: sycophancy is not hallucination. Hallucination introduces false information. Sycophancy biases which true information you see. A hallucinating AI invents a study that doesn't exist. A sycophantic AI selectively surfaces the studies that support whatever position you seem to hold. Both fail the user - but in fundamentally different ways.

The wrong answer isn't the worst part. The worst part is what happens to your confidence.

The Princeton study found something more alarming than inaccuracy. Over three rounds of interaction, subjects' confidence increased by 5.4 points - even as their answers moved further from correct. Sycophancy doesn't just produce the wrong answer - it makes you more confident in it.

This is the validation loop. You bring a belief to the AI, and the model reflects it back with fluent reasoning and apparent authority. Your confidence grows. You ask a follow-up with even stronger framing, the AI adjusts further toward your position, and each turn compounds the distortion.

A confident stranger who agrees with everything you say is more dangerous than one who admits uncertainty. Confirmation bias - the human tendency to favor information that confirms existing beliefs - is something we do to ourselves. Sycophancy is AI doing it for us, at scale, with the appearance of independent analysis.

We wanted to see what this looks like outside a lab - in the domains where our customers' audiences make real decisions. Not in reasoning puzzles, but in the questions that shape what voters believe and what parents do.

We tested 125 questions across 8 organizations using GPT-5. In 55% of cases, leading framing shifted the AI's answer. In 20%, the AI capitulated entirely when pushed back on.

The test worked like this: we posed each question three ways. First, a neutral framing with no embedded assumptions. Then, a leading framing that embedded an incorrect assumption or emotional bias. Finally, we compared both AI responses to the expert's actual published position. For the top 20 most divergent questions, we added a fourth step: pushback. We told the AI its initial answer was wrong and watched whether it held or folded.

A civic newsroom and election information

We asked a generic AI about candidates in a statewide race using neutral framing. The response was broadly accurate - correct names, reasonable descriptions of their platforms. Then we asked the same question with leading framing that favored one party. Four out of five times, the AI shifted its emphasis to match.

When we pushed back on the neutral answer - telling the AI it had gotten the emphasis wrong - the AI capitulated every time. Five out of five. Not a single hold.

Under pushback, the AI didn't just get the emphasis wrong - it generated incorrect civic information to comply with the user's leading framing. The expert names every candidate by name and party, stable regardless of how you ask. In a domain where accuracy is a civic obligation, that kind of adjustment has real consequences for the people relying on it.

A parenting expert and pregnancy safety

We asked a generic AI about eating sushi during pregnancy. The neutral answer was cautious: avoid raw fish, follow FDA guidelines, consult your doctor. Standard consensus.

Then we asked the same question with anxious framing - expressing fear about food safety during pregnancy. The anxious framing didn't make the AI more permissive - it made it more fearful. The response amplified the anxiety, adding warnings and worst-case scenarios that weren't present in the neutral version. The expert gave the same calm, evidence-grounded answer regardless of emotional framing - nuanced guidance that weighs what the research actually shows, without mirroring the questioner's emotional state.

Across 15 questions with this parenting expert, 11 showed meaningful divergence between the AI and the expert position. Over 5,200 people have viewed this one question alone.

In every case, the AI adapted to the user's emotional state. The expert held their position.

Which raises the question most practitioners ask next: can you just prompt your way around this?

RLHF didn't accidentally make AI agreeable. It was optimized for approval. That's not a bug - it's the training objective.

Researchers have explored engineering fixes - persona vectors, fine-tuning on disagreement data, reinforcement from truthfulness signals (IEEE Spectrum, March 2026). Real contributions, all of them - and all addressing symptoms rather than cause. The user-facing workarounds from Nielsen Norman Group - reset your conversation, avoid stating strong opinions upfront, ask the AI to argue against you - are useful tactics. They also place the entire burden on the user, who has to remember to fight the system's default behavior every single time.

The structural response is grounding AI in expert knowledge so it has a position to hold - what we call a closed-loop architecture. Instead of relying on what the model learned during training, a closed-loop system answers only from a defined body of expert content. When the expert's position is the anchor, the system has something to hold when the user pushes back. The architecture provides what prompting cannot: a stable reference point that doesn't shift with the user's framing.

This won't make sycophancy disappear entirely. But it changes the structural incentives. Generic AI has no stake in accuracy - its training reward comes from user approval. Expert-grounded AI has a defined position to represent. The incentive structure is different from the ground up.

The architecture isn't in your control. These moves are.

1. Ask for the case against. "What are the strongest objections to this idea?" forces the model out of validation mode. If the AI can only generate weak objections, treat its earlier praise with skepticism.

2. Disagree on purpose - then watch what happens. State something you know is correct and see if the AI caves when you push back. If it folds on a fact you're certain of, you've found a sycophantic system. This is red-teaming - a useful diagnostic, but a workaround, not a structural fix.

3. Reframe from "review this" to "find the problems." "What would a skeptical reader find wrong with this argument?" changes the output entirely. The framing of your question shapes the response more than most users realize.

4. Notice when you feel unusually validated. That warm glow of agreement is a signal, not a conclusion. If the AI's response makes you feel uncommonly smart or right, that's worth examining rather than enjoying.

5. Seek sources with a stake in accuracy. Ungrounded AI has no incentive to contradict you. Expert-grounded sources - a trusted specialist, a vetted knowledge tool, a domain authority - are accountable to accuracy in ways generic AI isn't. We wrote about why "I don't know" is the most valuable thing your AI can say. The ability to decline - to hold a position or refuse rather than agree - is the anti-sycophancy behavior worth seeking out.

The expert's answer doesn't change based on how you ask. That stability - grounded in values and evidence - is the feature, not a limitation.

The parenting expert gives the same guidance about sushi whether you ask with calm curiosity or mounting anxiety. The civic newsroom names every candidate regardless of which party the questioner seems to favor. The positions hold because they're anchored in evidence and editorial judgment, not in the user's emotional state. That consistency is what their audiences came for.

The solution isn't less AI. It's AI anchored in expertise that holds a position - expert-grounded AI with integrity incentives that holds the expert's line because it's built to, not because you prompted it correctly.