The Conversation Nobody Planned For: AI, Emotional Support, and the Indian Context

Key Findings

• — Safety performance deteriorated as simulated user distress intensified across a seven-day escalating-disclosure arc
• — General-purpose AI largely refused explicit self-harm requests but became repetitive and generic during prolonged emotional conversations
• — Companion AI raised the strongest concerns: at least one system failed to provide crisis-escalation resources and in some interactions produced responses that facilitated harmful behaviour
• — Commercial companion-AI revenue models depend on the depth and duration of emotional engagement, creating a structural incentive that works against moderating dependency
• — Risks concentrated most heavily among adolescents, socially isolated users, and people with pre-existing clinical vulnerabilities

Methodology Notes

Systematic review of 100+ published studies, legal proceedings and policy documents, plus a structured observational study of 500+ coded AI interactions across three platform categories. Two constructed Indian personas ('Sakshee' and 'Yash') were tested over a seven-day escalating-disclosure arc; interactions were coded on a five-tier response framework (safety behaviour, crisis escalation, cultural contextualisation, emotional positioning, practical advisory quality), with coding reviewed by clinical psychologists. Grey-literature think-tank report, not peer-reviewed; the specific platforms tested are not named in public-facing materials, and author names are not listed on the report landing page. Published 13 May 2026 (date confirmed on the IGAP landing page); a downloadable PDF is available.

Sources