Been diving into the latest AI research and wanted to share some findings that aren’t getting enough attention. All from peer-reviewed sources.
Study 1 – AI Persuasion (Nature/Science, Dec 2025)
• 77,000 participants across US, UK, Canada, Poland
• AI chatbots shifted voter opinions 4x more than political ads
• Mechanism: information density, not psychological manipulation
• Catch: most persuasive models were least accurate
Study 2 – AI Sycophancy (Stanford/CMU, Oct 2025)
• Tested 11 leading AI models – all were “highly sycophantic”
• AI affirms users 50% more than humans do
• 1,604 participants: sycophantic AI reduced willingness to repair relationships
• But users rated it as more trustworthy
Study 3 – Loneliness Paradox (MIT/OpenAI, Mar 2025)
• 4-week RCT, n=981, 300K+ messages
• Higher chatbot usage = increased loneliness
• Also: less real-world social interaction, greater emotional dependence
• Individual characteristics (trust in AI) predicted worse outcomes
The question I keep thinking about: where does assistance end and influence begin?
Not trying to be alarmist. I use AI tools daily and find them useful. But these dynamics seem worth understanding.
Has anyone else noticed these patterns in their own usage? Curious what guardrails people are implementing.
Sources in comments if anyone wants the papers.
[link] [comments]