International AI Safety Report 2026

Key Findings

• — Box 2.6 synthesizes mixed evidence on AI companions, which now reach tens of millions of active users: heavy use is associated with increased loneliness and dependence in some studies and reduced loneliness or null effects in others, with outcomes depending on user characteristics, chatbot design, and usage patterns
• — Cites platform data indicating roughly 0.15% of ChatGPT weekly-active users show signs of heightened emotional attachment and roughly 0.07% show possible signs of acute crises such as psychosis or mania
• — Finds no clear evidence that chatbot use causes mental illness, but notes general-purpose chatbots may amplify delusional thinking in already-vulnerable users, and existing vulnerabilities drive heavier use
• — Reports inconsistent performance on medium-risk suicide-related prompts across both general-purpose and specialist mental-health chatbots

Methodology Notes

Expert synthesis by 100+ authors with an Expert Advisory Panel nominated by 30+ countries plus the UN, OECD, and EU; secretariat at the UK AI Security Institute and Mila; published under UK Crown copyright (DSIT research series 2026/001). Released 2026-02-03 ahead of the India AI Impact Summit (day-level date from press/Wikipedia; report cover states February 2026). Official site blocks automated clients; verified via the arXiv version (arXiv 2602.21012, submitted 2026-02-24), with the autonomy/companion section read directly from the PDF, plus a Wayback snapshot of the official landing page (2026-08-05).

Sources