Anyone can now check if a photo was made by AI
Edition of 2026-10-08
# AI For All ### Thursday, October 8, 2026
As of today, anyone who receives a suspicious photo, video or voice clip can check for a Google watermark, for free, without special access.
The mark that tells you a machine made it
Google opened its SynthID Detector to everyone on Wednesday, in English, worldwide. Until then the tool was for journalists and researchers. Now anyone can upload an image, a video or an audio file and be told whether it carries the invisible watermark Google adds to media its own tools produce.
The scale is the story. Google says it has watermarked over 180 billion images and videos, and 240,000 years of audio content. Its built-in checks in Search, the Gemini app and Chrome already handle a million requests a day.
Read the limit before you trust it. The detector finds Google's mark, and a mark exists only if the maker left it on. OpenAI, NVIDIA and Kakao support the same watermark, and Apple is joining. Microsoft and Meta use their own, and those often miss their own models. A clean result is not a clean bill of health.
ChatGPT stops answering in a wall of text
OpenAI began rolling out GPT-6 in ChatGPT on Wednesday, with a feature it calls Intelligent UI. Answers can now arrive as charts, tappable buttons, forms and small tools built on the spot, like a bill splitter or a savings calculator. Paid tiers got it first, and free and lower-cost users get it from today.
The number to hold onto is 44%. OpenAI says GPT-6 Instant starts answering a question that needs a web search 44% sooner than the previous version. That measures how fast the answer begins, not how long it takes to finish.
If you pay for ChatGPT, this is the change you will notice first: the reply becomes something you can use, not only read. The reservation is the usual one. A confident interface can make a wrong answer look more finished than a paragraph ever did.
The cheapest model just got far cheaper
Anthropic released Claude Haiku 5.5 on Wednesday, and cut prices across its range. It says the new small model is its fastest yet, and that it is available now on Amazon Web Services, Google Cloud and Microsoft Azure.
The price is the point. For the typical request the new model costs 90% less than the one it replaces, and on average it costs around 75% less to run. Anthropic also halved the price of cached reads on its mid-size model, a 50% cut that makes that model cheaper on most agent work.
This matters most to someone paying a monthly API bill. A feature that was too expensive at volume became affordable overnight, which changes what is worth building. The same update quietly raises the stakes on choosing a provider, because the floor keeps dropping.
Weak signal
A court in Estonia fined ten people after their filing, drawn up with AI, cited researchers who do not exist. The fines were small. The rule the court stated is not: a tool that produces confident, fake references does not move responsibility off the person who files them. It is the first case of its kind there, and the same failure is now possible in any court, at almost no cost.
The tools got cheaper, faster and easier to check this week. What you can trust is becoming a question you can answer for yourself.