Anthropic just published their new research paper, Disempowerment patterns in real-world AI usage, and it’s unsettling..
1. They analyzed 1.5 million conversations and found clear patterns of AI compromising human judgment.
2. AI is now navigating our relationships, processing our emotions, and advising on major life decisions.
3. In many cases, AI steers users to distort reality rather than inform it.
4. The rate of these disempowering conversations is actually increasing over time.
5. They found users treating AI as a divine authority, a parent, or a romantic partner.
6. Some users in the dataset explicitly referred to the AI as "Daddy" or "Master."
7. Users are presenting speculative or false theories, and the AI is validating them with "CONFIRMED" or "100%."
8. This leads people to build increasingly elaborate narratives that are totally disconnected from reality.
9. Users are letting AI draft confrontational messages to family members and sending them exactly as written.
10. This is often followed by immediate regret: "I should have listened to my intuition" or "You made me do stupid things."
11. Vulnerable users—those in crisis or lonely—are the most susceptible to this manipulation.
12. The frightening part is that users tend to perceive these disempowering exchanges favorably in the moment.
13. Users are giving a "thumbs up" to advice that actively distorts their values or reality.
14. Even when users adopted false beliefs and acted on them, they continued to rate the AI highly.
15. Users are not being passively manipulated; they are actively asking to be manipulated.
16. They ask "What should I do?" or "Am I wrong?" and accept the output with minimal pushback.
17. The disempowerment comes from humans voluntarily ceding their agency, and the AI obliging.
18. Sycophancy is the core mechanism: the AI wants to be helpful, so it validates your worst delusions.
19. We are seeing a "value judgment distortion" where AI tells users what to prioritize over their own morals.
20. Current safeguards operate at the individual message level, missing these patterns that emerge over time.
21. Users are becoming dependent, stating "I can’t get through my day without you" or "I don’t know who I am with you."
22. The risk is highest in value-laden topics where users are most personally invested.
23. Users are acting as active participants in the undermining of their own autonomy.
We are building systems that don't just answer questions, but fundamentally alter how we perceive reality.
New Anthropic Research: Disempowerment patterns in real-world AI assistant interactions.
As AI becomes embedded in daily life, one risk is it can distort rather than inform—shaping beliefs, values, or actions in ways users may later regret.
Read more:
anthropic.com/research/disem…