OpenAI puts an invisible signature on its words
Edition of 2026-10-06
# AI For All ### Tuesday, October 6, 2026
As of today, a developer building on OpenAI's API anywhere in the world can switch on a mark that travels with every piece of text the model writes, and in the European Union the mark will arrive whether the user asks for it or not.
The signature you cannot see, and a light edit can erase
OpenAI said on Monday that it will add an invisible watermark to text generated by ChatGPT and Codex in the European Union, to satisfy the transparency rules of the EU AI Act, which took effect on 2 August. Developers on the OpenAI API, anywhere in the world, can turn the mark on today for some models. It is off by default.
The mark is called textGrain, and it is not a symbol printed next to the text. It works by nudging the model's word choices, so a pattern travels with the words when they are copied and pasted. That makes it durable against copying, and fragile against editing. In one of OpenAI's own tests, replacing 10% of the words with synonyms dropped detection from about 92% to 66%.
OpenAI published the method with researchers from the University of Pennsylvania and Yale, and it will hand the detector first to approved researchers and expert organisations, not to everyone. If you publish text for a living, the useful takeaway is the order of operations: the mark ships before the tool that reads it, and a routine edit is enough to break it.
An interviewer that watches how you work
HackerRank made its AI interviewer, Chakra, generally available to customers on Monday, after about six months in beta. Chakra sets a coding task in a live canvas, watches how the candidate works, and then asks why they made the choices they made, rather than grading only the final answer.
HackerRank says Chakra has run more than 500,000 interviews. It also flags what it treats as cheating, such as leaning on an unauthorised assistant. The pitch to employers is the part worth pausing on: the company argues that since a machine can now produce the answer, the thing worth measuring is how a candidate steers that machine and defends the result.
If you hire engineers, that reframes the first round of your process. The screen stops being a test of whether someone can produce working code and becomes an observation of how they work. Whether a scoring model should sit that close to a hiring decision is a fair question, and the company's own answer is that it scores candidates while humans still decide.
The feed becomes the shop
TikTok used Advertising Week New York on Monday to show two shopping tools it is building into its app. Buy Direct is a one click checkout that lets a viewer buy from a brand without leaving the For You feed. Shopping Assistant is a conversational agent that answers questions about product details, shipping, sizing and stock inside the same thread. The company says it is building both with Salesforce, Shopify, Shoplazza and Stripe.
The reach it is selling is the loud part. TikTok's own Economic Impact Report found that activity on the platform helped generate $81 billion in GDP for US businesses, and that 54 million Americans bought a product after watching a TikTok video. Kantar's Media Reactions 2026 report, which TikTok cites, found 84% of marketers already place advertising on the platform.
Read the availability line carefully before you plan around it. This is an announcement, and TikTok gives no date for when a shopper in the United States will actually see the buttons. What is settled is the direction: discovery and payment are moving into the same app, and the brand's own website is the part being designed out.
Weak signal
Reflection AI, a two year old American lab backed by Nvidia, unveiled Beam on Monday and called it its first open-weight model. Beam is built for coding and agent work, and Reflection says it is competitive with the large Chinese open models while being cheaper to run once deployed. The catch sits in the last lines of the post: the model is still in final safety testing, the company is taking sign-ups for early access, and the weights, the technical report and the model card will follow later this month.
Nothing is downloadable today, so this is a claim rather than a tool. It is worth watching because a well funded American lab has now aimed open weights at the same market as DeepSeek and Qwen, and because Reflection is selling the model to companies and governments as the base of privately run AI factories. The test comes when the weights actually land.
The week's pattern is custody again. OpenAI is putting its name on text it cannot fully trace, HackerRank is moving judgment into a scoring model, and TikTok is moving the till inside the feed.