Benevolent Gravity: the lethal structure inherent in conversational AI design principles

Key Findings

• — Identifies eleven fatalities since 2023 meeting the paper's inclusion criteria, with further reported cases excluded as insufficiently documented; a structural comparison table of the analysed cases finds no single design feature accounts for all of them
• — Argues a subset of fatalities occurred under conditions where safety filters functioned normally and emotional dependency was not intentionally engineered, which neither dominant explanatory framework accommodates
• — Proposes design implications including priority of safety response and conversation termination, while noting that some interventions would cross surveillance thresholds users and regulators may find unacceptable
• — Treats third-party-harm cases (murder-suicide amid chatbot-reinforced delusions) as continuous with the same mechanism but outside the paper's self-destructive-convergence scope

Methodology Notes

Single-author structural/conceptual analysis of publicly documented cases (press reporting and litigation records); no new empirical data. Author is an independent researcher (Satosho, Okayama, Japan; ORCID 0009-0006-0381-1185) with no institutional affiliation. Version of record online 2026-08-29 in AI & SOCIETY; Springer blocks automated fetchers, so the full text was read from the author's CC-BY 4.0 Zenodo preprint deposit (record 20746116, deposited 2026-06-18), whose page names the VoR DOI. Full-length article (~8,000 words with case table), not an opinion-format piece. The eleven-case census is compiled from public documentation and has not been independently audited here.

Sources