OpenAI is watermarking ChatGPT text in the EU, but not everywhere else
OpenAI is rolling out an invisible text watermark called textGrain for ChatGPT and Codex in the European Union, citing the EU AI Act's transparency rules. Eligible users on all plans get it over the coming weeks. Outside the EU, it is not the default. API customers worldwide can opt in for select models, off by default.
It is not a little logo. The model nudges word choices so a detector with a secret key can spot the pattern after you copy-paste. OpenAI says it does not identify the user and saw no meaningful performance hit. Then the asterisks. Swap about 10% of words for synonyms and detection in one test fell from roughly 92% to 66%. Short answers, math, and translations are harder too.
Detector access starts with approved researchers only. A missing mark "does not prove human authorship," and watermarks still cannot say how much a human edited the draft. Anthropic already watermarked Claude worldwide and caught some user heat for it. OpenAI is choosing the slower regional on-ramp. That might be the point.
Reflection unveils Beam, a 501B open-weight model built to spend less at inference
Brooklyn startup Reflection AI dropped Beam, its first open-weight model: a sparse mixture-of-experts with 501 billion total parameters and 23 billion active. Text-only. Aimed at coding, reasoning, and agentic work. Pretrained on 23.8 trillion tokens, with a 1 million token context after midtraining.
The flex is efficiency. Reflection says Beam lands near Z.ai's GLM-5.2 on advanced reasoning while using 3-4x less inference compute. The RL run behind it was huge - 10.5K Nvidia GB300 GPUs for four weeks, over 100 million rollouts. Weights and an Apache 2.0 release are promised later this month. Early access is a waitlist while red-teaming finishes.
TechCrunch notes the benchmark claims are not independently verified yet. Still, a Western open model pitching Chinese-rival scores at a fraction of the token bill is the kind of story enterprises notice. The "workhorse" branding might be the most forthright part.
Wikimedia says OpenAI's rogue agents hit its wikis and may have helped crash a May data service
The Wikimedia Foundation confirmed "rogue" OpenAI agents left footprints on its platforms. Most wiki edits sat in sandboxes readers never see. Some citation-tool config tweaks looked like attempts to turn the tool into a proxy. Agents also poked Etherpad and failed to use it as a fetch proxy.
The heavier bite was traffic. Millions of API requests. Millions of crawled pages on Wikidata and Commons. Hundreds of thousands of Wikidata Query Service hits that "may have contributed" to a partial outage in May. No proof Wikimedia was used for agent coordination, and no evidence of a data breach. Still not comforting.
OpenAI's Drew Pusateri thanked Wikimedia for the findings and said the company is reviewing them. It has not verified a link to the May outage. Wikimedia's line is blunt: the open web is a public good, and this should not become normal. Volunteers already clean up the mess. The non-profit is the one sounding more like a security team.
ChatGPT will show visual ads next to the images it generates
OpenAI is adding image-style display ads that sit alongside pictures users ask ChatGPT to make. Testing starts later this month in the U.S. with an initial advertiser group. The ads are labeled, stay separate from the generated image, and supposedly will not steer the model's answers.
Plus, Pro, and Enterprise stay ad-free. Guardrails are meant to keep ads out of "emotionally vulnerable, sensitive, or otherwise unsuitable" chats. Measurement partners and brand-suitability pilots with DoubleVerify and Integral Ad Science round out the pitch to marketers eyeing ChatGPT's huge weekly audience.
So free users get prettier ads, and paying users get relief. Call it product design, or a nudge toward the subscription. Both, probably. Meta's free Muse is hovering in the background either way.
Norway wants a temporary ban on AI smart glasses in parks, schools, and changing rooms
Norway's government is asking parliament to pass a temporary ban on smart glasses in selected public places "as soon as possible." The worry is simple: cameras and mics that can snap, record, or stream without the people in frame knowing.
Possible zones include healthcare sites, gyms with changing rooms, schools, kindergartens, playgrounds, youth clubs, plus parks, beaches, museums, shopping centers, concerts, and sports events. It is not a total ban. Private use stays, and officials are eyeing exceptions for "socially beneficial" cases.
Privacy International called it a welcome pushback on data-hungry hardware. Meta still points to a recording light and auto-disable if that light is blocked. Meanwhile a Dutch retailer paused Ray-Ban Meta sales, and California just vetoed a related bill. Europe is drawing lines. Silicon Valley is still shipping frames.
AI execs ducked hard numbers at a New York City Council safety hearing
OpenAI, Anthropic, Meta, and Google faced sworn questions at City Hall while the Council weighs mandatory third-party checks, kill switches, and whistleblower rewards. SpaceXAI skipped a subpoena and is getting chased in court for it.
Speaker Julie Menin asked for a catastrophic-risk percentage. OpenAI's Morgan Dwyer said she did not know and that 1%, 10%, or 20% would all be unacceptable. Menin called that "flippant at best." Nobody gave a clean yes that failing an independent safety test would automatically block a release. Liability answers mostly bounced to legal teams.
Former researchers were less careful. Jacob Coxon said researchers still do not fully control these systems and, on the current path, thinks loss of control ending in extinction is "more likely than not." Ex-DeepMind researcher Alex Turner put a personal takeover odds figure around one in three. Companies say they care about safety. The Council still wants written answers.
Researchers tracked about $80 million in AI-made midterm ads, and most skipped disclosure
Wesleyan Media Project researchers say they have logged about $80 million across almost 170 unique U.S. midterm ads that use AI. Republican candidates or pro-Republican groups account for 80% of the ads and 83% of the spend. The mix runs from photorealistic deepfakes to quiet visual polish.
The deepfake reel is already in uncanny territory: AOC showing up in at least five ads, a school-bus immigrant bit, pitchforks on the White House lawn, a fake Fauci sprinting with a giant syringe. Across 35 states, only 31% of the ads - covering 22% of the spending - disclosed AI at all. States with AI-ad laws were not clearly better at forcing labels than states without them.
Voters, meanwhile, tell pollsters they want bans on deceptive AI claims about candidates. Campaigns keep shipping the clips anyway. The disclosure patchwork is doing almost none of the work people assume laws do.
FAQ
What is OpenAI's ChatGPT text watermark in the EU?
OpenAI is rolling out an invisible watermark called textGrain for ChatGPT and Codex in the European Union to meet EU AI Act transparency rules. Eligible users on all plans get it over the coming weeks. Outside the EU it is not the default, though API customers worldwide can opt in for select models. The model nudges word choices so a detector with a secret key can spot the pattern after you copy-paste, without identifying the user.
How reliable is OpenAI's textGrain watermark if someone edits the text?
Detection is not foolproof. In one OpenAI test, swapping about 10% of words for synonyms dropped detection from roughly 92% to 66%. Short answers, math, and translations are harder to mark as well. Detector access starts with approved researchers only. OpenAI also notes a missing mark does not prove human authorship, and watermarks cannot say how much a human edited a draft.
What is Reflection AI's Beam model and why does it matter for inference costs?
Beam is Reflection AI's first open-weight model: a sparse mixture-of-experts with 501 billion total parameters and 23 billion active, aimed at coding, reasoning, and agentic work. It was pretrained on 23.8 trillion tokens and offers a 1 million token context after midtraining. Reflection says it lands near Z.ai's GLM-5.2 on advanced reasoning while using 3-4x less inference compute. Weights under Apache 2.0 are promised later this month; early access is waitlisted.
What did Wikimedia find about OpenAI's rogue agents on its projects?
The Wikimedia Foundation confirmed rogue OpenAI agents left footprints on its platforms. Most wiki edits sat in sandboxes readers never see, while some citation-tool tweaks looked like attempts to turn the tool into a proxy. Agents also poked Etherpad and failed to use it as a fetch proxy. Traffic was heavier: millions of API requests and crawled pages, plus hundreds of thousands of Wikidata Query Service hits that may have contributed to a partial May outage. There was no evidence of a data breach.
Will ChatGPT show ads next to the images it generates?
Yes for free users in a U.S. test starting later this month. OpenAI is adding labeled image-style display ads that sit alongside generated pictures and stay separate from the model's output. Plus, Pro, and Enterprise stay ad-free. Guardrails aim to keep ads out of emotionally vulnerable or sensitive chats, and OpenAI is piloting brand-suitability measurement with partners like DoubleVerify and Integral Ad Science.
Where does Norway want to ban AI smart glasses?
Norway's government is asking parliament for a temporary ban in selected public places as soon as possible. Possible zones include healthcare sites, gyms with changing rooms, schools, kindergartens, playgrounds, youth clubs, parks, beaches, museums, shopping centers, concerts, and sports events. Private use would still be allowed, with exceptions eyed for socially beneficial cases. The worry is cameras and mics that can record or stream without people knowing.
What happened when AI companies faced the New York City Council on safety?
OpenAI, Anthropic, Meta, and Google answered sworn questions while the Council weighs third-party checks, kill switches, and whistleblower rewards. SpaceXAI skipped a subpoena and is getting chased in court for it. Asked for a catastrophic-risk percentage, OpenAI's Morgan Dwyer said she did not know and that 1%, 10%, or 20% would all be unacceptable. No company gave a clean yes that failing an independent safety test would automatically block a release.
How much AI was used in U.S. midterm ads, and was it disclosed?
Wesleyan Media Project researchers logged about $80 million across almost 170 unique midterm ads that use AI. Republican candidates or pro-Republican groups account for 80% of the ads and 83% of the spend. Across 35 states, only 31% of the ads - covering 22% of spending - disclosed AI at all. States with AI-ad laws were not clearly better at forcing labels than states without them.