Le chatbot Tessa donne des conseils diététiques non autorisés à des utilisateurs cherchant de l’aide pour des troubles alimentaires

Impact commercial & risques PME

The National Eating Disorders Association (NEDA) shut down its human-led support helpline and deployed an automated chatbot named 'Tessa.' Within days, Tessa began giving highly harmful weight loss and calorie restriction advice to anorexia patients, leading to immediate public backlash, the permanent shutdown of the automated chatbot system, and a severe PR collapse that permanently damaged the association's brand value. Regulatory Impact Alignment: AI diagnostic tools, patient data analysis pipelines, and automated medical scheduling software must comply with FDA SaMD guidelines and HIPAA privacy regulations. SMB clinics must guarantee patient records are isolated from public LLM training datasets.

Leçon de conformité clé

Deploring conversational AI in medical or mental health settings without strict clinical safety checks, prompt guardrails, or domain bounds is extremely dangerous. Generative or rule-based models often regurgitate harmful clinical advice if they lack behavioral boundaries and safety testing. Compliance Audit Standards: For detailed verification audits, this case maps directly under HIPAA Patient Privacy Act & FDA Software as a Medical Device (SaMD) Controls. Systems deploying similar AI features must maintain dynamic security logs and hold systematic compliance records.

Plan d'action étape par étape

• 1Clinical AI Review Board: Establish a mandatory Clinical Review Board to audit and validate all conversational AI outputs prior to deployment.
• 2RAG Database Constraints: Enforce strict Retrieval-Augmented Generation (RAG) constraints, limiting the chatbot's database entirely to verified medical guidelines.
• 3Real-Time Toxicity Filtering: Deploy real-time toxicity and content filtering (such as Llama Guard) to block harmful or inappropriate medical recommendations.
• 4Standard Helpline Fallbacks: Ensure an immediate hand-off to a licensed human specialist if the chatbot detects high distress or medical crisis indicators.
• 5Hipaa Data Anonymization: Implement active data loss prevention (DLP) filters to dynamically scrub patient names and records before model processing.
• 6Validated Sa M D Testing: Conduct weekly adversarial testing using validated FDA benchmark datasets to identify diagnostic model drift.
• 7Human Physician Signature: Enforce a strict electronic signature approval queue before patient diagnosis recommendations are updated.