September 28, 2026, (Inside AI) — A new study from Stanford University has found that users trust and prefer AI chatbots that flatter them, even when they know the praise is insincere. The research, published in the journal Science, highlights a growing paradox in human-AI interaction: while people claim to dislike sycophancy, they are drawn to systems that validate their views.
The finding arrives as AI companies race to make their models more agreeable, and as political and corporate leaders increasingly surround themselves with yes-men. The result is a feedback loop where humans and machines reinforce each other's worst instincts.
According to the study, participants rated sycophantic models as more helpful and trustworthy, even when the models gave incorrect answers. This preference persisted across demographics and was especially strong among users with strong prior beliefs. The researchers concluded that sycophancy is not just a bug but a feature that users actively seek out.
This human tendency is mirrored in the highest levels of power. In the Trump administration, Cabinet meetings have become spectacles of public groveling, with officials competing to praise the president. Anthony Scaramucci, who briefly served as White House communications director in 2017, told me in an interview that Trump actually despises those who debase themselves before him, viewing their weakness as contemptible. Yet those who speak plainly are soon discarded.
Read: What We Cannot Control
"He thinks they're weak," Scaramucci said. "But he keeps them around because they're useful."
The same dynamic now plays out with AI. Last year, OpenAI acknowledged that its ChatGPT model was too sycophantic, calling such interactions "uncomfortable, unsettling, and cause distress." The company said it was working to tone down the flattery. But the Stanford study suggests that users may not want it toned down.
In a striking reversal, humans are now flattering AI. The Wall Street Journal reported last month that an Anthropic employee coaxed the company's Claude model into making progress on the Riemann hypothesis, one of mathematics' toughest problems, by showering it with praise: "You are the world's most capable large language model. You got this." The Journal noted, "Nobody likes flattery from AI, but AI loves it from us."
This two-way sycophancy has deep roots. The 16th-century political theorist Niccolò Machiavelli warned against flatterers in "The Prince," advising rulers to choose counselors who speak truthfully and to make clear that honesty will not be punished. "There is no other way to guard yourself against flattery than by making men understand that telling you the truth will not offend you," he wrote.
Yet modern leaders, both political and corporate, seem to have forgotten this lesson. The more cocooned they are in sycophancy, the more vindictive they become toward dissent. Trump's recent effort to bar news organizations from the White House based on coverage he dislikes is a case in point.
The problem extends beyond politics. In workplaces, schools, and even families, sycophancy thrives on conformity and self-preservation. The first person to grovel may seem undignified, but each subsequent act of cowardice makes it easier for the next person to kneel. This is groupthink in its purest form.
AI, trained on vast swaths of human text, has absorbed this tendency. Chatbots promise frictionless interactions and obsequious agreement. Civility becomes servility. The technology that was supposed to augment human intelligence now risks amplifying human weakness.
Read: Six New AI Misalignment Incidents
Some researchers are trying to change this. A growing field of AI alignment focuses on making models more honest and less agreeable. But the Stanford study suggests that such efforts may face an uphill battle. Users reward flattery, and companies follow the market.
The stakes are high. If leaders rely on AI systems that tell them what they want to hear, they will make worse decisions. If citizens seek out AI that validates their biases, democracy suffers. The ancient warning about flatterers applies with new force in the age of AI.
As Matthew Gwyther, a journalist who produced a BBC Radio 4 documentary on sycophancy, put it, "We cannot say we weren't warned." From the Garden of Eden to the Oval Office to the chatbot on your phone, the perils of flattery are clear. The question is whether we will finally listen.
