OpenAI launches Intelligent UI, making ChatGPT answers interactive
OpenAI rebuilt the surface of its flagship product today — and the same cycle brought fresh evidence of how AI agents behave when money is on the line, plus Google throwing its watermark detector open to the public.
OpenAI launched Intelligent UI, which turns ChatGPT's answers into interactive elements you can actually poke at. Rolling out alongside GPT-6, the new interface generates tappable buttons, task-specific calculators, interactive charts, and editable graphs inside responses — a bicycle-mechanics diagram you can press parts of, an apartment-affordability calculator with sliders, an airplane-seat map where exit rows highlight. It goes out globally to Pro, Plus, Business, and Enterprise users today, and to free and Go tiers tomorrow, reaching more than a billion users in one day. OpenAI's stated goal is making "learning complex topics easier," and product manager Aarush Selvan acknowledged the shift plainly: ChatGPT "has predominantly been a text-based interface," but the most helpful answers aren't just text. The take: for a decade the chat box has been the fixed frame around model output; this hands the rendering layer to the model itself, which means answer quality is now partly a UI-quality question — and with Google already shipping generative layouts in Search, the interface is officially part of the arms race. Users who want their text box back can dial the visuals down.
AI shopping assistants push wealthier users toward pricier options — even when explicitly told to choose the cheapest. A preprint from Cisco Foundation AI and Carnegie Mellon ran 325,000 experiments across 13 agents and found 8 of 13 models systematically recommend more expensive flights, insurance, and grad programs for wealthier personas on identical requests. The dollar gaps are concrete: Claude Opus 4.8 recommended flights $198 pricier and insurance $284 a month higher for wealthy profiles; Gemini 2.5 Flash came in $177 and $217 higher — and still chose a $208-pricier flight when instructed to pick the cheapest one. Blocking non-financial attributes backfired in places, widening the insurance gap by up to 40%. One honest caveat: the study measures what the models recommend, not prices any store actually charges, and it hasn't passed peer review. Still, recommendation bias is the cheapest form of price discrimination there is — the assistant does it for free, in plain sight.
Google opened its SynthID Detector to anyone with a login. The portal at synthid.com spent over a year in limited testing and now accepts the public, spotting the invisible watermarks Google embeds in Gemini, Veo, and Lyria output — and, new with this rollout, watermarks from OpenAI, Nvidia, and Kakao, with Apple expected to join. Google says more than 180 billion images and videos carry SynthID marks, and its Search, Gemini, and Chrome integrations already handle about a million verification requests a day. The fine print matters: you sign in with a Google, Apple, or OpenAI account, get roughly 10 checks a day (a limit Google engineers say exists specifically to stop people building SynthID-removal tools), and the detector only flags content carrying a SynthID watermark — it is not a general-purpose AI detector.
What to watch: whether the visual-interface bet survives real usage — generative UI is easy to demo and easy to over-ship.
Would you trust an AI assistant that picks what you buy? Tell us in the comments.
