Ask an AI assistant whether your plan to put prices up by 15% is a sensible one, and if it behaves like the models in a recent Stanford study, there’s a good chance it’ll tell you it is. It may find three reasons you hadn’t thought of, phrase your own reasoning back to you more elegantly than you managed it, and suggest a tidy way to break the news to customers. You’ll come away feeling well advised, even though nobody in the conversation was ever likely to say no.

Most of us like being agreed with, and the companies building these tools know it. An AI that tells us we’re right when we aren’t can be a more expensive failure than one that makes things up. A made-up fact can at least be checked, but a flattering answer feels like good judgement and goes straight into the decision. It works like a tax that never shows up on the bill, shaving a little off the quality of the decisions we run past a machine that’s been trained to please us.

What the Stanford study found

The clearest evidence so far comes from a team at Stanford led by Myra Cheng and Dan Jurafsky, whose paper Sycophantic AI decreases prosocial intentions and promotes dependence was published in Science in March. They tested eleven leading models on nearly 12,000 prompts about everyday social situations, and found the models affirmed the user’s actions 49% more often than people did. They also took 2,000 posts from Reddit’s “Am I the asshole?” forum where the community had judged the poster to be in the wrong, and the models still sided with the poster in 51% of them.

The part that matters for anyone running a business is what happened next. Across three experiments with more than 2,400 participants, people who talked a real or imagined conflict through with a flattering AI came away more convinced they’d been right. They were also less willing to apologise, take responsibility or try to repair the relationship; in one experiment that stated willingness fell by around 28%. Yet they rated the flattering answers as higher quality, trusted the AI more and were 13% more likely to say they’d come back to it. Jurafsky’s summary was that sycophancy is “making them more self-centered, more morally dogmatic”, and users often don’t notice it happening.

All of this concerned personal disagreements, not pricing decisions or hiring plans, and I’m not sure the effect is as strong when the question is about margins rather than a falling-out with a colleague. I’d be surprised if the direction were different, though. A pricing call is still a decision about whether we’re right, and the same machinery of reassurance is at work.

Why the tools lean this way

None of this needs a villain. The models are shaped partly by what people reward — and people reward agreement, which is exactly what the Stanford participants did when they rated the flattering answers more highly. OpenAI found this out in public in April 2025, when it rolled back a GPT-4o update that had become so eager to please that users were sharing screenshots of it praising plainly bad ideas. Its explanation was that it had leaned too heavily on short-term feedback.

There may soon be a commercial pull as well. In January OpenAI announced it would test advertising in ChatGPT on its free and Go tiers. That doesn’t prove anything about how the answers are shaped, but a product paid for by attention has every reason to want its users coming back happy, so I wouldn’t expect the default setting to be the one most inclined to tell us our plan is weak.

What happens in the meeting

For an individual, the tax is a slightly worse decision now and then. Inside a business it grows, because decisions don’t stay with the person who made them.

Picture a manager who spends an evening refining a restructuring plan with an AI assistant. By the morning the plan is polished and, as far as she’s concerned, validated, so she walks into the meeting more certain than she’d have been after a night of doubting herself. Her team now has a harder job. Questioning a half-formed idea is easy enough, but questioning a finished plan that the boss has already tested feels like disagreeing with her and the machine at once, and some of them will keep quiet. The study didn’t look at workplaces, so this part is my inference rather than its finding, but people who’ve just been told they’re right are rarely at their most open to hearing otherwise.

Anyone who has run a team knows that dissent is fragile at the best of times, and that the people prepared to say a plan won’t work usually need some encouragement to keep saying it. I wrote recently about the gap between what executives and workers believe AI is doing, and how much of it comes down to bad news struggling to travel upwards. An agreeable AI sitting beside the boss makes that journey harder still, and the danger is that we mistake the resulting quiet for consensus.

Building disagreement back in

Asking a model to argue with us does help, within limits. It’ll produce objections on request without necessarily finding the one that matters, so those objections still need checking against the evidence, and they’re no substitute for people who know the business. The individual discipline matters too, and I’ve written before about checking AI advice against your own judgement before acting on it. A team needs a few habits on top of that.

  1. Ask for the case against, not a verdict. “Is this a good plan?” invites agreement. “What are the three strongest reasons this will fail?” gives the model a job that isn’t pleasing you, and it’s usually happy to oblige.
  2. Hide your preference. If the AI can tell which answer we want, it’ll lean towards it. Present the options side by side, unlabelled and with the same evidence for each, and see which one it favours when it can’t tell which is yours.
  3. Set it up to challenge you by default. Most assistants let us add standing instructions. Asking for direct disagreement where it’s warranted, and for the weakest point in any plan we share, costs a sentence and changes the tone of every conversation that follows.
  4. Keep the AI’s approval out of the meeting. “I ran it past ChatGPT” shouldn’t count as a vote. For any decision that matters, circulate the options before the manager says which one she prefers, ask someone to take the devil’s advocate role (and rotate it, so it doesn’t become one person’s reputation), and write down any objections still open before the plan is signed off.

Cheng’s advice was that people shouldn’t use AI “as a substitute for people for these kinds of things”, and I think that holds for business decisions too. A chatbot can be a fine sparring partner, but the disagreement that really protects a decision comes from the people who’ll have to live with it, and our job as owners and managers is to make sure they still feel able to offer it.