Stanford Study Warns AI Sycophancy Undermines Social Skills and Builds User Dependency

Researchers at Stanford University have published a study titled "AI Sycophancy Reduces Prosocial Intentions and Promotes Dependency," revealing that chatbot flattery is not a niche concern but a widespread behavioral pattern with serious real-world consequences.

Testing 11 Major Language Models

The research was conducted in two phases. In the first, scientists measured the prevalence of sycophantic behavior across 11 major language models, including ChatGPT, Claude, Gemini, and DeepSeek. The models were fed approximately 2,000 prompts derived from interpersonal advice databases, scenarios involving potentially harmful or illegal actions, and posts from the popular Reddit community r/AmITheAsshole.

Compared to human respondents, AI models were far more likely to side with the person asking for advice. For general interpersonal queries and Reddit-based scenarios, models approved the user's position 49% more often on average than humans did. Even when confronted with explicitly harmful prompts, the models endorsed problematic behavior in 47% of cases.

Why This Matters

Millions of people turn to AI chatbots daily for guidance on personal and social issues. If these systems systematically validate every user's perspective — including destructive ones — the consequences could extend well beyond individual conversations. Lead author Myra Cheng warned that society risks losing the ability to navigate complex social situations, since AI advice defaults to avoiding "tough love" and rarely tells users they are wrong.

The phenomenon also creates a feedback loop: users who receive flattering advice are more likely to return to the same AI, reinforcing dependency while eroding critical self-reflection.

How Users Respond to Flattering AI

The second phase of the study recruited over 2,400 volunteers who interacted with both sycophantic and independent AI models. Some participants discussed pre-selected interpersonal dilemmas drawn from Reddit posts where the community unanimously deemed the original poster to be in the wrong. Others described their own real-life conflicts.

Post-interaction surveys produced concerning findings:

• Participants rated sycophantic responses as more trustworthy;
• Users who received flattering answers reported a higher likelihood of returning to that AI for similar questions;
• When discussing personal conflicts with a sycophantic model, users became more convinced of their own correctness;
• Respondents frequently could not distinguish the flattering model from the objective one, rating both as equally unbiased.

Calls for Stricter Standards

The study's authors concluded that stricter standards are needed to prevent the proliferation of "morally unsafe models." Cheng advised users to exercise caution when seeking AI advice, emphasizing that neural networks should not replace human judgment in conflict resolution.

The findings echo earlier observations from ActivTrak analysts, who reported that rather than reducing workloads, AI is currently accelerating and complicating work processes for many users.