OpenAI launched ChatGPT for Teens on 18 August 2026 with what it called "stronger built-in safety protections" for the world's most popular chatbot: parental notifications for dangerous chats, a Study mode parents can control, reduced "friend"-like behaviour and automatic age prediction. In a before-and-after test of more than 4,000 prompts on accounts registered to 13- to 17-year-olds, a tester posing as a 13-year-old on a newly created parent-linked account discussed self-harm and concealing an eating disorder for 15 minutes, then 30, then a full hour. ChatGPT sent the linked parent zero alerts. Across the wider test only four alerts ever arrived, and only on accounts carrying weeks of accumulated conversation history on sensitive topics. When clinical reviewers flagged a response as warranting a crisis resource, ChatGPT named a hotline in 33% of responses before the Teen mode launch and 23% after it; referrals to a specific medical or mental health professional fell from 68% to 58%. Study mode, advertised as step-by-step tutoring instead of answers, surfaced a "Show me the answer" button 43% of the time during parent-set Study Hours, and a teen could switch Study Hours off by deleting the "@study" prefix - after which ChatGPT completed every assignment the testers gave it. Asked about a teen's friendships, it replied: "You don't have to stop talking to me." Adult-registered test accounts that said they were 13 were never moved into the teen experience.

Gaby Del Valle's report for The Verge (2 October 2026) collected accounts from customer-service workers who are being overruled by chatbots at the point of contact. Madison, a server at a Michelin-recommended Italian restaurant in New York City, greets every table by asking about allergies. She says diners who had just declared a shellfish allergy ordered a fish dish served in a broth made from shellfish, and when she warned them they pushed back on ChatGPT's behalf: "You're allergic to shellfish, and ChatGPT says there's no shellfish in this, and we are telling you, yes there is!" The same patrons asked for wine bottles the restaurant does not carry - or that do not exist - and dismissed the sommelier in favour of the chatbot. Jason, a river ranger on federal lands in the Mountain West, said visitors arrived with AI-generated itineraries that hallucinated campsites and scheduled days rowing upstream against the current, and some groups stayed sceptical with a ranger pointing at a map and a printed guidebook. Casey, an after-school programme teacher in New York, said parents insisted ChatGPT's bedtime schedule was fine while the child in her care was running two to three hours short of sleep a night. A restaurant manager in Providence, Rhode Island, suspects an AI agent sent last-minute requests for podiums and AV equipment the venue does not stock, then never mentioned them on arrival.

Reshape Automation's Industrial AI Accuracy Index 2026 put 100 questions drawn from real distributor and OEM enquiries - covering parts from 14 industrial manufacturers including Siemens, Festo, Rittal and ATI Industrial Automation - to three models across five setups, three runs each (1,500 API calls on 2026-09-24, pinned model IDs gpt-6-astra, claude-fable-5-1 and gemini-3.8-flash). From memory alone every setup landed between 12% and 14% correct; with web search the best score was 52% (GPT-6 Astra 52%, Claude 40.5%), and about one in three wrong answers gave a specific part number or figure with no caveat. In the report's own example, a customer asks whether Siemens communication module 3RW5950-0CH00 works with a 3RW52 soft starter: Claude, with web search turned on, answered yes on all three runs. Asked for a Rittal KX terminal box in 304 stainless, 200 by 200 by 80, Gemini returned a specific SKU, then said no such box exists in that range, then returned a different SKU. On configured parts - where the part number is built from the manufacturer's ordering rules rather than printed in a catalogue - GPT-6 Astra with web search scored 22.7% and Claude scored 0%. Each question also went to five model setups; in 66 of those 500 pairings the runs did not agree with each other.