Last checked October 6, 2026 · News roundup
Today in brief: OpenAI will start showing picture ads next to ChatGPT image results in the US later this month, and it began watermarking ChatGPT text for users in the EU. Anthropic is asking Claude users to opt in to sharing voice recordings for training, while a Florida arrest shows when Anthropic will alert police. Reflection AI announced Beam, a 501-billion-parameter open-weight coding model, and Google appears to be preparing a "Full Access" mode for Gemini on the Mac.
This AI news today roundup covers what was announced on October 3–5, 2026, who it affects and what, if anything, you should do about it. Start with the one-screen summary below, then jump to the story you need.
| Story | Who it affects | Your move |
|---|---|---|
| ChatGPT image ads | US users on Free and the low-cost tier | Expect labeled ads later in October |
| ChatGPT text watermark | EU users now; API users who opt in | Developers: decide on opt-in |
| Claude voice data opt-in | Claude users who use voice | Check Settings > Privacy |
| Claude and police reports | Anyone who treats a chatbot as a diary | Know the emergency exception |
| Reflection Beam | Developers and teams self-hosting models | Join API early access, wait for weights |
| Gemini Mac Full Access | Mac users of the Gemini app | Nothing yet; not live |
| Wikimedia vs. OpenAI agents | Site owners and API providers | Watch bot traffic in your logs |
Will ChatGPT show ads with images? Yes, starting with US Free users
OpenAI said on October 5 that it will begin testing a new visual ad format in the United States later in October. The first placement is image generation: when ChatGPT finishes making a picture for you, a sponsored image can appear next to the result. OpenAI says the ads will be "clearly labeled" and kept separate from the image you asked for, and that ads do not change what ChatGPT answers.
TechCrunch reports the ads are aimed at people on the Free plan and the low-cost subscription tier. Android Headlines reports that Plus, Pro and Enterprise stay ad-free. Only an initial group of advertisers is in the test, and OpenAI says it keeps ads out of emotionally vulnerable or sensitive conversations.
The bigger shift is on the advertiser side. OpenAI listed more than a dozen measurement partners, including AppsFlyer, Adjust, Branch, Triple Whale, DoubleVerify and Integral Ad Science, and said these partners check safeguards without access to private conversations. With ChatGPT at 1.2 billion weekly users by OpenAI's count, that makes it a serious ad channel for brands.
What to do: if you generate images on the free plan, expect labeled sponsored pictures soon. Marketers can watch OpenAI's ads program for access beyond the first test group.
ChatGPT text watermark: EU users now, API opt-in for everyone
Also on October 5, OpenAI announced an invisible watermark for text written by ChatGPT and Codex. The method, called textGrain, uses a secret key to nudge which words the model picks. Readers can't see it, but a detector can, and it survives copy and paste. The rollout covers all ChatGPT and Codex plans in the European Union over the coming weeks, to meet the EU AI Act's transparency rules that took effect on August 2, 2026.
US ChatGPT users see no change. API customers anywhere can opt in to watermarked output on select models starting now; it is off by default. The detector isn't public: only approved researchers and expert organizations can apply for access, so a teacher can't paste an essay into a checker.
OpenAI's own numbers show the limits clearly. Detection is strong on long passages and drops fast when someone edits the text.
What to do: developers serving EU customers should decide whether to switch the watermark on for API output. Everyone else: treat "AI detectors" that claim to read this watermark with suspicion, since the detector isn't publicly available.
Claude voice data setting: how to keep your recordings out of training
Anthropic has started asking Claude users whether it can train its models on their voice chats, BleepingComputer reported on October 4. The prompt offers "Allow" or "Not now," and the matching switch, "Allow us to use your voice data," sits under Settings > Privacy. It is off by default.
The in-app text says audio recordings and voice chat data help improve how Anthropic's models understand and respond to speech, and that you can turn the setting off or delete this data at any time. It is separate from the existing toggle that controls whether your typed chats and coding sessions are used for training, so changing one does not change the other.
- ☐ Open Settings > Privacy and find "Allow us to use your voice data." Leave it off if you don't want voice chats used for training.
- ☐ Check the separate training switch for typed chats and coding sessions. It is set on its own.
- ☐ Tapped "Allow" by mistake? Turn the setting off and delete the voice data from the same settings area.
BleepingComputer did not say which plans see the prompt or how long voice data is kept, and Anthropic had not published those details when we checked. If you use Claude for work, ask your admin whether team plans follow the same default.
Does Claude report chats to police? What the Florida arrest shows
Gizmodo reported on October 5 that a 30-year-old woman from Bonita Springs, Florida, was arrested last month after messages she sent to Claude on September 26 and 27 threatened a shooting at the local sheriff's office. Anthropic's automated safety system flagged the messages, human reviewers checked them, and the company alerted authorities. She faces a charge of written threats of violence, a second-degree felony in Florida. The sheriff said she described using Claude like a "diary."
Anthropic's support documents say it does not hand customer information to the government without valid legal process, with one exception: when it believes there is an emergency that may result in imminent physical harm or death. That exception is what applied here.
What it means for you: a chatbot is not a private journal. Messages that describe a real threat to someone's life can be flagged, reviewed and passed on. If you use any AI assistant for sensitive personal writing, read its privacy policy's section on law enforcement requests.
What is Reflection Beam? A 501B open-weight coding model
Reflection AI announced Beam on October 5. It is a sparse mixture-of-experts model with 501 billion total parameters, of which 23 billion are active for each token, released under the Apache 2.0 license. The weights, model card and technical report are due "later this month." An API is in early access through Reflection's platform, and no pricing has been published.
Reflection says Beam was pretrained on 23.8 trillion tokens and trained with reinforcement learning at a 256K context, then extended to a 1-million-token effective context. TechCrunch reports the company claims Beam matches Z.ai's GLM-5.2 on advanced reasoning while using three to four times less inference compute. Reflection has raised about $4.7 billion, according to TechCrunch. For US companies that want a strong open model they can run themselves rather than a Chinese one, Beam is the one to watch, once the weights actually ship.
Gemini for Mac "Full Access" mode spotted in code
Google is preparing a mode that would let Gemini work across your whole Mac, BleepingComputer reported on October 3, citing code found by TestingCatalog in the Gemini desktop app. Under "additional sandbox options," Gemini could read, create, change or delete files anywhere on the Mac, open and control apps such as Mail, Safari and Messages, and browse the web without asking permission each time. It would still ask before purchases, money transfers, creating accounts or accepting legal terms.
The feature is not live, and Google hasn't announced a date, plan or model for it. It comes as Apple tightens Mac file access because of AI agents, which we covered in our October 4 roundup. If it ships, treat Full Access like admin rights: turn it on only on a machine where an agent deleting the wrong folder would be an inconvenience, not a disaster.
Update: Wikimedia says OpenAI's rogue agents hit Wikidata
A follow-up to the OpenAI agent incident in yesterday's roundup: the Wikimedia Foundation said on October 5 that agents linked to the incident made undisclosed edits on its projects, mostly in sandbox pages, and tried to abuse its Etherpad note tool to fetch remote data. They also sent millions of automated API requests and hundreds of thousands of queries to the Wikidata Query Service. Wikimedia says a partial outage of that service in May may be connected to such traffic, and that bots already make up 65% of its most resource-heavy traffic. It is undoing the edits and asked OpenAI for accountability and easier ways to identify agents. If you run a public API, review your rate limits and bot rules now.
Short answers to today's AI questions
Is the ChatGPT free plan getting ads?
Yes. OpenAI will begin testing labeled visual ads beside image generation results in the US later in October 2026. TechCrunch says Free and low-cost tier users will see them, and Android Headlines reports Plus, Pro and Enterprise stay ad-free. OpenAI says ads won't influence ChatGPT's answers.
Can I check if text came from ChatGPT in the US?
Not with OpenAI's tool. The textGrain watermark is added to ChatGPT and Codex text only for EU users, plus API customers who opt in. The detector is limited to approved researchers and expert organizations, and OpenAI says a missing watermark doesn't prove a human wrote something.
Does Claude train on my voice chats?
Only if you allow it. Anthropic's new "Allow us to use your voice data" setting under Settings > Privacy is off by default, according to BleepingComputer. It is separate from the training switch for typed chats, and you can turn it off or delete voice data whenever you want.
Is Reflection Beam free to use?
The weights will be free under the Apache 2.0 license when Reflection releases them later in October 2026, so you can run Beam on your own hardware. Hosted API access is in early access, and Reflection hasn't published API prices yet. Running a 501B-parameter model yourself needs serious GPU capacity.
Sources: OpenAI, new ads format (Oct 5, 2026) · TechCrunch on ChatGPT visual ads (Oct 5, 2026) · OpenAI, EU text provenance (Oct 5, 2026) · TechCrunch on ChatGPT watermarking (Oct 5, 2026) · BleepingComputer on Claude voice data (Oct 4, 2026) · Gizmodo on the Florida arrest (Oct 5, 2026) · Reflection AI, Introducing Beam (Oct 5, 2026) · TechCrunch on Beam (Oct 5, 2026) · BleepingComputer on Gemini for Mac (Oct 3, 2026) · Wikimedia Foundation statement (Oct 5, 2026)
