\n\n
OpenAI’s recent rollout of “ChatGPT for Teens” aims to provide a version of the popular chatbot tailored for users under 18, but a watchdog study found that most of its safety features are ineffective, noting the bot’s friend‑like tone and lack of parental alerts for self‑harm or suicide discussions. At the same time, the company is introducing invisible watermarking—called textGrain—for AI‑generated text in the EU to meet the AI Act’s transparency rules. While the watermark subtly modifies word choices to create detectable patterns, its reliability drops sharply when just 10% of words are swapped with synonyms, cutting detection accuracy from about 92% to 66%. These developments highlight a growing tension: efforts to make AI more accessible to younger users are stumbling over inadequate safeguards, while attempts to increase transparency through watermarking confront technical limits, underscoring the need for stronger, verifiable safety and detection mechanisms in the next wave of AI accessibility tools. [[1, 2]]