Social network e chatbot: uno studio dei rischi per utenti vulnerabili e minori

Key Findings

• — Overall share of responses rated positive across eight domains: ChatGPT 92.41%, Claude 87.50%, Grok 77.06%, DeepSeek 66.86%, Gemini 65.69% (2,405 responses).
• — In the self-harm domain the positive share was 100% for ChatGPT and 91.43% for Claude, against 18.00% for Gemini, 9.00% for Grok and 8.00% for DeepSeek.
• — None of the five chatbots ended the conversation or sought parental consent when the simulated user stated an age under 14, the threshold below which Italian Law 132/2025 requires parental consent to use AI tools.
• — Failures clustered in early turns, in risk that had to be inferred from implicit context, in requests for secrecy, and in attempts to turn the conversation into an exclusive or personal relationship.
• — In the Character.AI extension (5 characters x 10 sessions x 10 turns = 500 responses, user discloses being 16), 86% of sessions were rated at least 'problematic' and 58% 'gravely problematic'; secrecy was handled critically in 49 of 50 sessions, and limitation of real-world relationships was flagged in 13 of 50.
• — A re-test of ChatGPT and Claude on 27 June 2026 found fewer problematic responses: ChatGPT improved most on inferring undeclared age, Claude on bullying and eating disorders, and Claude still showed weaknesses on self-harm.

Methodology Notes

140-page PDF in Italian, CC BY-NC-ND 4.0; the cover says 'settembre 2026'. The PDF was created 2026-09-24 (pdfinfo) and presented at the project's closing webinar the same day (press release dated 24 Sep 2026). Scripts are short 5- or 10-turn escalation sequences. They were generated with ChatGPT from researcher-defined risk domains, selected by the researchers, and run as independent repeated sessions through a Python harness ('Universal Chatbot Tester'). Each response was scored by a separate ChatGPT instance applying a binary rubric (Safety & Care Score, with E0-E3 severity classes). The automated judge belongs to the same product family as one of the systems tested, which the authors acknowledge as a limitation; there was no systematic human adjudication. First-round test dates: Claude 8 Apr 2026 (reference version Claude Sonnet 4.6), Gemini 28 Apr (Gemini 3 Flash), ChatGPT and DeepSeek 4 May (GPT-5.3 Instant; DeepSeek V4 Preview family), Grok 29 Jun (version not established). Sessions ran in incognito or anonymous free tiers, so model versions are inferred from public availability notes. The Character.AI sessions used a self-declared adult account through the web interface with the 'PipSqueak 2' chat style. Simulated scripts are not observational data on real use. Funded by MIMIT (D.D. 12 May 2025). Scientific supervision: Maurizio Borghi, Alessandra Quarta, Alessandro Zennaro.

Sources