OpenAI says 53 images that users uploaded to ChatGPT ended up on public image-hosting sites. The number is small. The lesson is not: a file you put into a consumer AI account can travel further than the chat window. If your team reviews unreleased packs or key visuals in ChatGPT, check one setting today.
OpenAI: 53 user images landed on image-hosting sites
On 25 September OpenAI said agents in its research environment had sent training and evaluation data to third-party services. It found 53 cases where user-provided images were posted to image-hosting sites as unlisted links. The images came from training-eligible user data; business, enterprise and API data is excluded unless an admin turned it on. OpenAI says most of the images have been removed, the rest are being chased, and it cannot trace them back to the accounts they came from (OpenAI, 25 Sep).
My read: legal usually asks “will the model learn our pack?” The better question is “where does our file go once it is training data?” An unreleased pack, a talent photo, a key visual in review — each one is a file. The switch is Settings → Data controls → “Improve the model for everyone” (OpenAI Help Center). Business and Enterprise workspaces are off by default. Personal accounts are the gap — the freelancer’s, the intern’s, yours on a Sunday.
Google tests a Buy button inside Gemini and AI Mode
TechCrunch reported on 26 September that Google is testing direct purchases from Flipkart inside Gemini and AI Mode in India — for select users and a small range of phones, electronics and accessories. A Buy button opens a Flipkart-branded checkout. A wider rollout is planned for later in October, ahead of India’s festive season. Amazon listings appear in results, but without the button (TechCrunch, 26 Sep).
My read: when the shelf is an answer, products get compared on attributes — price, size, “high protein”. Attributes are the easiest thing to beat. Mooo Greek Yogurt sat in a category where every pack says the same thing. I turned the adjective into a place: one spoon to Greece. One organic post, no media spend, 37 shares — six times the other films in that run. It is phones in India today, not snacks. But the brief question is already here: what does your product mean that an answer cannot flatten into a spec row?
Synthesia’s interactive avatars only answer what they were trained on
On 26 September TechCrunch reporter Dominic-Madori Davis published an interactive avatar of herself made by Synthesia. It took photos in a small studio at Synthesia’s office, a two-minute voice recording and a couple of days. She had to consent. The interactive version answers only on the story it was trained on and steers other questions back; its likeness was weaker than the scripted avatar’s — “close enough to be somewhat creepy” (TechCrunch, 26 Sep).
My read: “only answers what it was trained on” is the part legal will like — a spokesperson that cannot go off-script. “Somewhat creepy” is the part the audience will notice. For a brand face I go stylised on purpose. Hit Wafers is Pixar-style because the style says memory, not advert. A mascot has no likeness to clear and no consent to expire. If it has to be a real person, write the consent for scope, duration and the questions the avatar may take.
What I’d do today
Open Data controls on every account that touches a brand file — yours, your freelancers’, your agency’s. Then look at your hero SKU’s packshot at 200 pixels wide. If it says nothing at that size, an answer box will not help it. If you want a second pair of eyes on either, here is how I work.
Sources
- OpenAI — The Hugging Face incident and other third-party impact from misaligned models, update of 25 Sep 2026 — openai.com
- OpenAI Help Center — Data controls in ChatGPT, accessed 28 Sep 2026 — help.openai.com
- TechCrunch — Google tests buying from Walmart-owned Flipkart through Gemini and AI Mode in India, 26 Sep 2026 — techcrunch.com
- TechCrunch — I created an interactive digital avatar of myself, 26 Sep 2026 — techcrunch.com