How is AI impacting people?

Key Findings

• — Companion platforms are large and intensive: market leaders such as Character.ai are reported to have upwards of 20 million regular users, and users spend an average of two hours per day on the platform, 33% more than regular TikTok users' time-on-platform.
• — Prevalence estimates for companionship use diverge sharply: surveys report figures such as 70% of US teens using AI for companionship, while developers' internal analyses find only 2 to 3% of conversations with Claude or ChatGPT serve that purpose.
• — In a study of relationship-seeking model personas, users' enjoyment waned over time while markers of separation distress increased; the authors describe this dissociation between 'liking' and 'wanting' as a hallmark of other forms of dependence, while noting there is as yet no evidence for positive or negative impacts on social or emotional behaviour from these studies.
• — On mental health, the review finds considerable uncertainty about the magnitude of harm from widespread AI use, describes sycophancy-driven feedback loops, and cites simulation studies in which chatbots invite escalation, increase dependency and enable harm.
• — Named open questions include which companion-product features drive dependence, which demographics benefit or suffer, and whether heavy companion use spills over into offline relationships.

Methodology Notes

Unreviewed narrative review with no stated systematic search protocol; 32 pages; correspondence address at the UK Department for Science, Innovation and Technology. Version 3 posted 2026-09-06 (OSF record; PDF created 2026-09-06); version 2 was posted 2026-06-17 and the first version earlier in 2026 (date not exposed by the OSF API). The published_date above is the version-3 date; no DOI is recorded on the OSF record. Verified from the OSF API record and the downloaded PDF.

Sources