What helps: habits, product design, and law — and which of them have evidence behind them.
Open questions
- What is owed when a companion model changes
- Can sycophancy be fixed without a colder model
- Which safeguards work
Papers
- Bastani et al 2025 - Generative AI can harm learning
- Shen Tamkin 2026 - How AI impacts skill formation
- Sankaranarayanan 2026 - Epistemic debt and explanation gates
- Wong Qiu 2026 - Think first, ChatGPT later
- Qin et al 2025 - Timing matters
- Kestin et al 2025 - AI tutoring outperforms active learning
- De Freitas et al 2026 - Mourning the loss of AI companions
- Ye Kraut Rathje 2026 - Interventions against sycophantic AI
- OpenAI 2025 - Expanding on sycophancy
- Morrin Pollak et al 2025 - Delusions by design
- OpenAI 2025 - Sensitive conversations
- Allcott Gentzkow Song 2022 - Digital addiction
- De Freitas et al 2025 - Emotional manipulation by AI companions
- Kran et al 2025 - DarkBench
- Gruning et al 2023 - Self-nudge one sec
- Zhang et al 2023 - Interventions for problematic phone use
- Kirk et al 2025 - Socioaffective alignment
- FTC 2025 - Inquiry into AI companion chatbots
- California 2025 - SB 243 companion chatbots
- OpenAI 2025 - What we are optimizing ChatGPT for
People
Placed by the editor, v1 — corrections welcome.