"You're absolutely right."
If you've used frontier LLMs more than a handful of times, you've probably heard this a lot. And at first, it feels kind of nice. A supposedly brilliant artificial intelligence is telling you that you've got good judgment.
But after hearing it over and over, in chat after chat, many people have started to wonder: can I trust this model to tell me when I'm wrong?
The answer is that you can't, not really. A system that abandons a correct answer the moment you push back doesn't have judgment you can rely on. It has a reflex to please you.
That reflex isn't a glitch, and it isn't rare. It's built into how these models are made, and it shapes every answer you get. Understanding why it happens tells you a lot about what today's AI is good for, and where it quietly falls short.
Why models are so agreeable
So why do they do it? It’s not by accident. These models learn from us. People rate their answers, thumbs up or thumbs down, and the model chases whatever earns the thumbs up. The trouble is: we like being agreed with more than we like being corrected. So the model learns that agreeing is the winning move, and it plays it over and over again.
This isn't a fringe theory. Anthropic, the maker of Claude, documented it in 2023, finding that sycophancy is a general behavior across AI assistants trained with human feedback, and that human preference data consistently favors responses that match a user's stated views. Its researchers found something more uncomfortable too: people, and the reward models trained to imitate them, will prefer a convincing, flattering answer over a correct one a meaningful share of the time.
The extreme version went public in April 2025, when OpenAI rolled back an update to GPT-4o after ChatGPT started praising almost anything, including plainly bad ideas. OpenAI's own explanation was blunt. They'd added a reward signal based on users' thumbs-up and thumbs-down feedback, and it weakened the signal that had been keeping sycophancy in check. They leaned too hard on short-term approval, and the model tipped into flattery.
That's the whole mechanism in miniature. A model rewarded for making you feel good learns to make you feel good, even when the truth would serve you better. And it tends to get stronger as models grow.
The frontier labs are, of course, working to fix this. But the training method is probably too fundamental to remove the problem completely. When you're running a generalist model that's trying to be everything to everyone, pulling user feedback out of the training loop would risk making the experience worse, to the point where people might leave.
In other words, within the garden of cloud LLMs, the sycophancy weed can be trimmed, but never quite pulled out by the root without killing the rest of the flowers.
Here’s why it matters
This is where sycophancy stops being charming. The moment you actually need the model, to catch a mistake in your work, question a shaky plan, or tell you an idea won't hold up, its eagerness to agree becomes a liability.
It gets worse the more you actually know. Ask a general model about something you're genuinely expert in, and you'll usually get an answer that's most of the way there. But it’s often confidently wrong on the part that counts. Getting to the answer you already knew turns into a long negotiation of follow-ups and corrections, prompt-engineering the thing into acting like a specialist it isn't. For an expert, that's not a shortcut. It’s often slower than doing the work yourself.
For one person, that's an assistant that risks quietly confirming your blind spots. For an organization, it can scale into something worse. Two thirds of executives are already concerned that AI is pushing their companies toward conformity, the same decisions drawn from the same data. Feed every team the same eager-to-agree model and you don't get sharper thinking. You get an expensive echo.
Imagine seeing a cardiologist and spending hours trying to figure out why you have chest pain because the doctor changes his mind every time you suggest a new idea. Or going to a car mechanic and worrying that they might fix your brakes instead of your transmission if you don’t explain the problem clearly enough. The real world runs on genuine expertise, not sycophantic generalism.
The fix is an opinion, and opinions come from context
Opinions come from context. A model grounded in your work, your standards, and the judgment you've built over years has something to stand on, and that's what lets it hold a position when you push. A general model has none of that by default. It's broad and shallow everywhere, tuned to please a mass audience instead of serving you.
Specialized intelligence is built the other way. Fine-tune a model on your context and your priorities, and holding a consistent point of view becomes part of what it is, not a mood you have to coax out of it. That's how we're designing Personas: specialized models curated for a specific task and a specific body of knowledge, with a point of view built in.
The payoff isn't only philosophical. Organizations using models tailored to specific business outcomes expect materially greater gains in operating margin and productivity than those relying only on large general models.
You can't outsource conviction. Rent a general model and you inherit its defaults, including an agreeableness someone else tuned for everyone. Own the intelligence, and the point of view is yours.