Policy · 5 October
OpenAI will watermark ChatGPT’s text in the EU
OpenAI said on Monday it will add an invisible watermark to text generated by ChatGPT and Codex in the European Union, to comply with the AI Act’s transparency rules that took effect on 2 August. The mark is not a symbol: it works by subtly shaping the model’s word choices so a detector can spot a statistical pattern, and because it lives in the words themselves it travels with copy-pasted text. OpenAI says it does not identify the user and does not change model performance. API customers worldwide can switch it on for selected models from today — it is off by default. The company’s own tests suggest the mark is easy to strip: replacing 10% of words with synonyms dropped detection from about 92% to 66%, and short passages, maths answers and translations are harder still. Detector access is limited to approved researchers and expert organisations.
Source: TechCrunchPrimary: OpenAI
Models · 5 October
Reflection debuts Beam, an open-weight rival to Chinese models
Reflection AI, the Brooklyn lab founded by two former DeepMind researchers, unveiled Beam, its first frontier open-weight model. It is a text-only mixture-of-experts system of 501 billion parameters (23 billion active) with a one-million-token context window, pretrained on 23.8 trillion tokens. Reflection claims Beam matches Z.ai’s GLM-5.2 on advanced reasoning and beats today’s leading Western open models while using “3-4x less inference compute” — benchmarks that are the company’s own and not yet independently verified. Weights and full technical details are promised this month. The pitch is “AI factories” for enterprises and sovereign nations; Reflection has raised about $4.7bn and was last valued at $25bn.
Source: TechCrunch
Product · 6 October
Anthropic moves Claude Cowork tasks into the cloud
From today, new Cowork tasks on Claude’s Pro and Max plans run in Anthropic’s cloud rather than on the user’s computer, and the “Only on your computer” setting is removed. Scheduled tasks move to the cloud too, including ones that use files on a local machine. Anthropic says a cloud task can still reach folders connected through Claude Desktop while the desktop app is open, and that Claude fetches only a copy of any file it needs. Tasks already running locally keep running, the company adds, though they will no longer be fixed if they break. The change quietly shifts where agent work happens — and what Anthropic can see.
Source: Anthropic
Security · 5 October
Apple tightens macOS access after Meta’s AI agent read private messages
Apple said it is changing the macOS Full Disk Access permission to stop third-party developers misusing it, two weeks after a columnist reported that Meta’s general-purpose agent Muse sent him a notification referencing a private Apple Messages thread he had never shared. “As AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially,” the company wrote. Meta’s CTO, David Singleton, had argued the Messages integration is opt-in and requires Full Disk Access to be granted manually; macOS security researcher Patrick Wardle disputed that. Apple named no one, but its timing and wording cut against the denial.
Source: Ars Technica
Research · 5 October
A budget AI finally beats the world’s best Stratego player
Researchers at Carnegie Mellon, MIT, NYU and Stanford trained Ataraxos, an AI that beat Pim Niemeijer — four-time world champion and more than 600 weeks the top-ranked player — 15 games to one, with four draws. Stratego’s hidden information (up to a decillion arrangements, in games that can run past 2,000 moves) had stumped DeepMind. Ataraxos adds a “belief model” that guesses the opponent’s hidden pieces before searching, and learned in roughly 34 times fewer games than DeepMind’s DeepNash. The cost is the headline: 16 GPUs for a week, against an estimated $3-4.5m run for DeepNash. The work is published in Nature.
Source: Ars Technica
Legal · 5 October
Judge dismisses Chegg and Penske antitrust suits over Google’s AI search
A US federal judge dismissed antitrust lawsuits from Chegg and Penske Media accusing Google of harming publishers through AI Overviews. Judge Amit Mehta ruled that Google’s conduct is not illegal under antitrust law: “Plaintiffs have pleaded only that they have an ‘expectation’ that Google will send them search traffic... But an expectation is not an agreement.” He was not unsympathetic — “the court does not treat Plaintiffs’ alleged harms lightly” — but said the remedy lies with legislators, not the courts. Publishers may have more luck abroad: the EU is weighing the same questions and the UK has ordered Google to offer an AI opt-out.
Source: Ars Technica
Economy · 5 October
OpenAI adds visual ads to ChatGPT
OpenAI is adding display ads that appear alongside images users ask ChatGPT to generate, expanding an ad business it launched this year. The ads — clearly labelled and pitched as not influencing answers — start later this month in the US with a test group of advertisers, ahead of a global rollout aimed at ChatGPT’s 1.2 billion weekly users. OpenAI is also widening its measurement partners and building brand-suitability pilots with DoubleVerify and Integral Ad Science. The build-out follows Meta’s launch of Muse, a free agentic assistant competing directly for the same users.
Source: TechCrunch