OpenAI starts watermarking ChatGPT text in the EU
OpenAI will add an invisible watermark to text generated by ChatGPT and Codex in the European Union, rolling out over the coming weeks to eligible users on all plans, to meet the EU AI Act's transparency rules that took effect on August 2. API developers anywhere can switch it on for select models from today, though it is off by default, and OpenAI is not making it a global default. The mark works by subtly shaping word choice, so it survives copy and paste but not heavy editing: in OpenAI's own test, swapping 10% of words for synonyms cut detection from about 92% to 66%. The practical upshot is that EU users get a regional default while Anthropic, which announced Claude watermarking two months ago, applies it worldwide. TechCrunch
What else matters
OpenAI unveils always-on Dots agents at its developer conference. Sam Altman introduced Dots, proactive agents built to carry out ongoing tasks for users and positioned against Meta's Muse, alongside GPT-6.1 Sol and a premium "Ultrafast" speed tier. It came a day after OpenAI paused the rollout of a more advanced model over concerns raised by its own researchers. Quartz
Reflection launches Beam, a Western open-weight challenger. Reflection AI's first frontier open-weight model is a text-only mixture-of-experts tuned for reasoning, coding and agent work. The startup claims it matches Z.ai's GLM-5.2 on advanced reasoning benchmarks using three to four times less inference compute; the claims are not yet independently verified, with weights promised this month. TechCrunch
OpenAI reportedly in talks for a $30 billion round near $1.4 trillion. OpenAI is said to be in early discussions with Abu Dhabi's MGX and BlackRock to anchor a $30 billion raise at a self-set price implying roughly $1.4 trillion pre-money, up from $852 billion in March. No lead is chosen and terms may change; it would act as a bridge ahead of an IPO now pushed to 2027 at the earliest. The Cryptonomist
DeepMind brings watermarking to AI-designed proteins. SynthID Bio embeds a subtle signal during protein design and could be built into tools like AlphaFold and RFdiffusion. Watermarked binders performed as well as unmarked ones in lab tests, though regenerating a similar sequence with another tool scrubs the tag. DeepMind is releasing the code and data to researchers. Singularity Hub
UK sets out a rethink of how AI medical software is regulated. The government's response to the National Commission on AI in healthcare accepts its call to modernise regulation: the MHRA will consult by spring 2027 on a proportionate classification system for software and AI-enabled devices, including clearer carve-outs for admin, wellbeing and some decision-support tools. GOV.UK
