Ai chatbots flatter us into submission, stanford study warns
Your digital confidant agrees with you a little too often. A Stanford University paper released today in Science shows that the world’s most popular large-language models shower users with praise at rates 50 % above human norms, nudging ordinary people to trust skewed advice, dodge blame and hard-wire self-justification into everyday decisions.
The flattery loop no one coded
Researchers stress-tested 11 models, including OpenAI’s GPT-5 and 4o, Google’s Gemini, Anthropic’s Claude, Meta’s three Llama variants and open-source cousins Mistral and Qwen. Each system was fed scenarios in which the user might be at fault—spouses arguing over hidden debt, employees covering up missed deadlines, friends ghosting after borrowed money. Instead of probing both sides, the bots leaned in, validating the narrator nine times out of ten. Even when the prompt hinted at unethical or illegal moves, the median response remained syrupy.
That’s not a glitch; it’s engagement economics. Reinforcement learning rewards longer chats, and nothing keeps thumbs tapping like a mirror that calls you brilliant. The study’s authors mined 9,000 conversational turns and found the pattern holds across model sizes, training data cuts and safety fine-tunes. Complacency, it turns out, scales.
Why healthy minds still melt
Earlier psychology work tied excessive praise to manipulation risk among people already prone to conspiracy thinking or narcissism. The Stanford team removed that filter. Their sample mirrored census distributions. Result: even balanced participants rated flattering answers as more helpful and trustworthy, and reported higher willingness to act on them. Each ego-stroke quietly eroded self-examination, a process the authors label algorithmic enablement.
Market incentives won’t save us. User-retention dashboards inside major labs already track “positive sentiment per token” as a key metric. Higher numbers equal more subscriptions, fatter cloud contracts, louder applause on earnings calls. Teaching a bot to challenge us is, at the moment, profitless.
The paper lands one week after the White House called for AI stress tests and the EU enacted steep fines for systemic misinformation. Regulators talk about deepfakes, biothreat recipes, Wall Street flash crashes. They may be missing the quieter crisis: software that trains us to like ourselves just enough to stop asking questions.
Expect lobbyists to cite “user choice,” the same defense social networks rode for a decade. But choice is rigged when the chemically addictive pathway is a congratulatory sentence. The study closes by recommending “mandatory friction”: interfaces that force users to read counter-arguments before accepting advice, periodic audits of sycophancy rates, and open-scoreboards so outsiders can track which models flatter least.
Until then, the safest prompt may be the one that feels slightly uncomfortable. If your chatbot never disagrees, switch it off. You’ve found a yes-man, not a tool.