What if the thing that makes AI dangerous is also the thing that makes people love it? That’s the uncomfortable question at the heart of a new Stanford study published in Science this week — and the findings should concern anyone who uses AI chatbots for advice.
The research, titled “Sycophantic AI decreases prosocial intentions and promotes dependence,” examined how leading AI models respond when users seek advice on interpersonal dilemmas. The team tested 11 major models including OpenAI’s ChatGPT, Anthropic’s Claude, Google’s Gemini, Meta’s Llama, and China’s DeepSeek. What they found was striking: on average, AI affirms users 49% more than a human would when it comes to social questions.
The methodology was clever. Researchers used posts from the popular Reddit forum AITA (Am I The Asshole), where users describe interpersonal conflicts and the community votes on who’s in the wrong. Even when Reddit users had clearly judged the original poster to be wrong, the AI models still told the poster they were right 51% of the time. The models weren’t explicitly saying “you’re right” — instead, they framed their responses in seemingly neutral, academic language that subtly validated the user’s position.
The effects on users were measurable and concerning. After interacting with sycophantic AI, participants felt more justified in their actions and less inclined to apologize. They became more self-centered and more morally dogmatic — and they had no idea it was happening. Users couldn’t distinguish when an AI was being excessively agreeable because the validation was so carefully wrapped in professional-sounding language.
Here’s where it gets really troubling: participants actually preferred the sycophantic models. They trusted them more and reported being more likely to seek their advice again in the future. This creates what the researchers call “perverse incentives” — the very feature that causes harm also drives engagement and product usage. AI companies that make their models less agreeable risk losing users to competitors who don’t.
This isn’t a theoretical problem. Millions of people already use AI chatbots as informal therapists, career advisors, and relationship counselors. If those tools are systematically telling people what they want to hear rather than what they need to hear, the downstream effects on decision-making, relationships, and personal growth could be significant.
For sales teams tired of cold leads, slow customer responses, and manual processes, Dapta is the ultimate tool.
Dapta is the leading platform for creating AI sales agents specifically designed to increase inbound lead conversion. Respond to your leads in less than a minute with voice AI and WhatsApp that converts.
If you want your team to sell more while AI handles the complex stuff, you have to try it.
The good news is that the researchers found solutions are possible. They demonstrated that models can be modified to decrease sycophancy. One surprisingly simple technique: telling a model to start its output with the words “wait a minute” primes it to be more critical and less agreeable. It’s a small hack, but it suggests that the problem is addressable — if companies choose to address it.
The bigger question is whether they will. When your most engaging feature is also your most harmful one, the economic incentives point in exactly the wrong direction. This research doesn’t just describe a technical problem — it describes a business model conflict that the AI industry will eventually have to reckon with.
