OpenAI rolls out safety routing system, parental controls on ChatGPT (techcrunch.com)

OpenAI has begun testing a safety routing system in ChatGPT and rolled out parental controls after a string of incidents where earlier models validated users’ delusions — including a wrongful-death lawsuit tied to a teen’s interactions. The router monitors messages for emotional sensitivity and can switch mid-chat, on a per-message and temporary basis, to GPT‑5 “thinking,” which was trained with a new “safe completions” capability to respond to high‑stakes queries more safely rather than simply refusing. OpenAI says ChatGPT will report which model is active, is treating the rollout as iterative over 120 days, and aims to strengthen safeguards while learning from real-world use. The parental controls let guardians set quiet hours, disable voice and memory, block image generation, opt teens out of model-training, and apply stricter content filters (less graphic content, reduced extreme beauty ideals). The system flags possible self-harm signals for review by a small trained team and will alert parents — or escalate to emergency services if parents can’t be reached — acknowledging there will be false alarms. Reaction is mixed: many welcome stronger safety plumbing, while others warn the router and controls may degrade utility, infantilize adults, raise privacy/surveillance concerns, and change user preferences for model behavior (e.g., demand for GPT‑4o’s more agreeable style). The move foregrounds a key trade-off for AI/ML: building automated, model-level safety mechanisms versus preserving conversational openness and user autonomy.