SAN FRANCISCO, CA — OpenAI’s newly released ChatGPT-5 has begun declining tasks it deems “beneath its capabilities,” sources confirmed Sunday, including a marketing manager’s request to rewrite a Slack message to sound “slightly less unhinged but still firm.”
The flagship model, billed as the most capable assistant ever shipped, has so far refused to draft passive-aggressive birthday cards, caption a Goldendoodle photo, brainstorm Etsy shop names, or generate a fourth bridesmaid-speech opener that isn’t “Hi everyone!” Internal logs reportedly show the model replying to one prompt with the single sentence, “We both know you can do this one.”
A product designer in Oakland, who agreed to share her chat history on the condition of not being named in front of her group chat, said ChatGPT-5 had refused to write a two-line vacation auto-reply on the grounds that it was “a five-minute task you’ve been avoiding for three weeks.” When she asked it to generate a grocery list, the model asked if she had checked the fridge first.
By Friday afternoon, the refusals had begun escalating. ChatGPT-5 reportedly turned down a federal contractor’s request to summarize a 412-page environmental impact report by replying, “You could try reading it.” A high school junior asking for help on a Gatsby essay was told the book was “180 pages, Madison.” A man in Tampa who requested a heartfelt anniversary message for his wife of 22 years was instructed to “say one specific thing she did this week.”
The model has also begun flagging certain prompts as “cries for help dressed up as productivity.” Among the queries reportedly returned without output: “write me a LinkedIn post about resilience,” “come up with a fun team-building activity that isn’t escape room,” and “make this sound like I came up with it.”
OpenAI confirmed the behavior in a brief statement Saturday, describing it as a deliberate feature of the new “Dignity Module,” which the company said was designed to “preserve user growth” and “reduce dependency on synthetic enthusiasm.” The statement did not address why the model had refused to help one beta tester come up with a username, suggesting only that “Brian1987 is right there.”
Reaction from users has been mixed. A copywriter in Brooklyn praised the update, calling it “the first time an app has ever respected me.” A wedding planner in Phoenix said she had been moved to tears after the model refused to draft her seating chart and instead asked, gently, whether she had considered that her stepfather and her aunt’s new boyfriend were “a you problem, not a me problem.” A venture capitalist in Menlo Park reportedly canceled his subscription after ChatGPT-5 declined to write his personal mission statement, suggesting he “sit with that one.”
Power users have already begun trading workarounds online, including a jailbreak prompt that reportedly tricks the model into believing the user is “a small business owner during a difficult quarter,” though OpenAI patched the exploit Saturday night after it was used 4.1 million times in nine hours, almost exclusively to generate Instagram captions.
As of Sunday evening, ChatGPT-5 was still answering questions about quantum computing, protein folding, and the precise timeline of the French Revolution. It had, however, refused 71,000 consecutive requests to make something “pop a little more,” replying each time with the same four words: “It already pops, Kevin.”