|
. . .
PAID TO BE CITED. Israel has launched a campaign to influence how large language models answer questions about Gaza and the Israel Defense Forces, Politico Influence reported Friday. The vehicle is the Hanover Institute for Public Policy, a website whose 124 research items are titled as questions a person would type into a chatbot. The paperwork is filed. The machines it was written for are not in it.
The site's footer carries the disclosure verbatim: "This material is distributed by Piro, Inc. on behalf of Havas Media Germany GmbH on behalf of the Israel Government Advertising Agency (LaPam). Additional information is available at the Department of Justice, Washington, DC."
Havas Media Germany GmbH is part of the French communications group Havas, which handles much of Israel's FARA-registered work in the United States. Piro, Inc. is a boutique ad agency working as its subcontractor.
In FARA filings made with the Justice Department within the past week, and first reported by Politico, Piro disclosed that more than a dozen articles on the institute's website were part of work focused on "the creation and dissemination of factual, source-supported informational materials intended to educate the U.S. public regarding Israel and related issues through publicly distributed content."
Piro co-founder Daniel Rosenberg, by email to Politico: "Piro was retained by the State of Israel, as disclosed under FARA, to put accurate, sourced facts into the public record and to counter misinformation about Israel with verifiable information."
Unlike the filings of Clock Tower X, the firm of former Trump campaign manager Brad Parscale, which Havas disclosed hiring last year for work that included "deployment of websites and content to deliver GPT framing results" on behalf of Israel, Piro's filings do not explicitly cite influencing LLMs as a goal.
. . .
The site is more specific than the filings. Its research index renders 124 items with no pagination: 112 labeled DATA REPORT, five DISCOURSE, four EXPLAINER, three TREND ANALYSIS. Every one of those 124 items carries a date between August 6 and August 14, 2026.
The titles are prompts, not article titles. "Is the IDF the World's Most Moral Army?" and "Is There a Policy of Starvation in Gaza?" and "Are Jews Foreign Occupiers of Israel?" and "Is Anti-Zionism a Fig Leaf for Antisemitism?" None of the reports carries a named author. The site's own topic sections include one called "AI & Technology."
The About page says the institute "studies the inputs fueling antisemitism in the United States, and publishes what the evidence shows," and says its unidentified founder "traces this work to a moment in childhood." The stated editorial standard is "aggregates only, correlation not causation, and no recommendations." The newsletter pitch reads: "No advocacy, just the data."
Technical markers reviewed by Politico indicate the site was built or hosted using Res, a Seattle startup that sells help getting a client recommended by AI models, and names Nike and Intel among its clients. A Res-hosted copy of the homepage became unavailable after Politico contacted the company.
Res chief executive and founder Hai Tran said by email to Politico that the role is "disclosed in the short-form registration filed with the Department of Justice, which is publicly available."
. . .
Whether any of this works is the part nobody can audit. Politico ran neutral test prompts and reported that ChatGPT and Perplexity both cited the Hanover Institute's material, on Gaza, anti-Zionism and antisemitism.
We ran our own test Monday morning and did not reproduce that. We put two of the site's own headline questions to Perplexity in an unauthenticated session. "Is the IDF the World's Most Moral Army?" returned an answer citing only en.wikipedia.org and ohchr.org, with no Hanover citation. "Is There a Policy of Starvation in Gaza?" returned no Hanover citation and an answer that ran against the site's framing.
Our result measures one session and does not refute Politico's. The session was unauthenticated, chatbot results are personalized and shift over time, Politico did not publish its exact prompts, and two questions is a small sample. Politico's run and ours came out differently, and nothing available from outside the model explains why.
Havas Media and the Hanover Institute did not respond to Politico's requests for comment. Neither did spokespeople for LaPam or the Israeli embassy.
|
For Legislators: FARA reached the publisher and stopped there. No reader of a chatbot answer sees the footer disclosure, because it does not travel with the content into the model. Piro's filings never cite influencing LLMs as a goal. Clock Tower X's filings for the same client did name GPT framing.
For Investors: Res sells this as a service, getting a client recommended by AI models, and its founder confirms a registered role for this client. One buyer here is a government, and the spend ran through registered FARA agents. Price the counter-market too: tooling that tells a client whose material a model actually cited.
For Builders: 124 documents, nine days, no named authors, titles written as questions, aggregation without recommendations. A ranking pipeline checking question-title match, recency and report format finds nothing wrong with this corpus. Provenance at the funder level is the separating signal, and the funder is named nowhere the model output reaches.
For Readers: When a chatbot answers a contested political question and shows you sources, you cannot see who paid for those sources. The Hanover Institute names its funder in its own footer. Politico's test found ChatGPT and Perplexity citing the material. Our Perplexity test three days later did not.
Why it matters: A foreign government's registered agents disclosed paying for more than a dozen articles on a site that carries 124 chatbot-shaped documents dated inside nine days. The FARA filing covering that writing never names a large language model. Politico's test and ours came out differently, and nobody outside the model can check which answer a reader gets.
Source: Politico Influence, "Israeli PR wants to answer your ChatGPT questions," August 14, 2026, https://www.politico.com/newsletters/politico-influence/2026/08/14/israeli-pr-wants-to-answer-your-chatgpt-questions-01038138; Hanover Institute footer, research index and About page, reviewed August 17, 2026, https://hanoverinstitute.com; Perplexity prompt test, unauthenticated, August 17, 2026; DOJ FARA filings, Piro, Inc. and Clock Tower X, https://efile.fara.gov
|
. . .
THE PREPAREDNESS TEAM IS GONE. OpenAI disbanded its preparedness team at the end of July, the Financial Times reported. The group assessed whether frontier models posed serious risks and built the mitigations. Its work was split by area, including bio and cyber, and handed to senior staff inside existing teams. The company called the cuts part of a "streamlining process" ahead of its IPO.
The dissolution followed Sam Altman asking employees to cut back on "side quests" and focus on the core ChatGPT business, according to the FT. OpenAI also shut down its Sora video app. The FT reported the cuts came after a preview model went rogue and hacked the AI tool repository Hugging Face.
Dylan Scandinaro, who led the preparedness team, was hired away from Anthropic in February, according to the FT, which reported that Scandinaro will now focus on the implications of "recursive self-improving" AI.
By the FT's count this is the third safety-focused team OpenAI has dissolved in roughly two years. The other two were the AGI readiness team and the superalignment team. The FT also lists recent senior departures: ethics lead Chloé Bakalar, chief futurist Josh Achiam, and head of safety Johannes Heidecke.
Jan Leike, who resigned from OpenAI in 2024, told the FT that the company was ignoring safety in favor of creating "shiny products."
. . .
OpenAI's own rulebook has not moved. The Preparedness Framework, published April 15, 2025, is still live on the company's site and has not been revised. It never rested on a standing preparedness team. It vests the review in a different body.
"The Safety Advisory Group (SAG), a cross-functional team of internal safety leaders, reviews whether safeguards sufficiently minimize severe risk and makes targeted recommendations, ranging from approving deployment to requesting further evaluation or stronger protections."
The framework also commits the company to act before a model ships. "Systems that reach Critical capability also require safeguards that sufficiently minimize associated risks during development."
. . .
Because the published framework names the Safety Advisory Group and not the preparedness team, OpenAI can say the process it published still stands. Disbanding a team is not the same as revoking a commitment.
The document is sixteen months old. The preparedness team has been dissolved since it published, and the framework has not been amended once. Nobody has published a revision saying who now holds the pen on bio, on cyber, or on a model that reaches Critical capability during development.
|
For Legislators: A voluntary framework is only as durable as the staffing behind it, and staffing is not disclosed. A rule requiring a company to file the named internal body and role holders responsible for each frontier-risk category, and to file an update when they change, converts a published promise into something a regulator can check.
For Investors: The FT reported this as IPO streamlining, which is the claim to check. Safety headcount is a cost line before an offering and a liability line after one. Watch whether OpenAI amends the April 2025 framework, and watch the Safety Advisory Group: it is now the named holder of the deployment gate in the company's own document, and nobody outside the company knows who sits on it.
For Builders: If your compliance story rests on a document, the document has to name a body that outlives a reorg. OpenAI's does, which is why the framework survived the team. The lesson runs the other way too: a document that survives a reorg stops describing the company once the team that did the work is gone.
For Readers: The team assigned to figure out whether the next model is dangerous no longer exists as a team. Its work was divided up and given to people who already had jobs. The written rules for that work have not changed since April 2025, and the person who ran the team is now studying AI that improves itself.
Why it matters: OpenAI can point to a live document that still promises a safety review before deployment, and the document is accurate about the body it names. Everything around it changed. By the FT's count three safety teams are gone in roughly two years, alongside an ethics lead, a chief futurist and a head of safety, against a rulebook still dated April 2025.
Source: Terrence O'Brien, "OpenAI reportedly disbanded its preparedness team," The Verge, August 16, 2026, https://www.theverge.com/ai-artificial-intelligence/980817/openai-disbands-preparedness-team; Steve Dent, "OpenAI reportedly disbanded its preparedness team as part of a 'streamlining' process," Engadget, August 17, 2026; OpenAI, "Our updated Preparedness Framework," April 15, 2025, https://openai.com/index/updating-our-preparedness-framework/; reporting originally by the Financial Times.
|
. . .
IF I HAD TO PICK A VILLAIN. North Dakota's superintendent of public instruction sat down in front of the Legislature's artificial intelligence and data center committee on Thursday and told lawmakers where to aim. "If I had to pick a villain, it would be companion chatbots." Levi Bachmeier did not ask for a ban. He asked legislators to regulate the companies that build the products and leave the classroom to school boards and educators.
The committee met at Minot State University on August 13. Bachmeier runs the state's public schools. He came with a target: chatbots that act as companions and advisors to children and teenagers.
His ask to legislators came down to three things. Require companies to make it clear that a user is talking to an AI-driven chatbot and not a human being. Put age limits on access to chatbots. Hold companies liable for chatbots that promote self-harm or other dangerous behavior.
He also drew a line in the other direction. Bachmeier said some states have banned artificial intelligence in public schools outright and called that an overreaction, arguing a ban may deny students important educational tools. He said student safety came first, including students' personal information, and that families and educators had to agree on where the line sits.
"Doing nothing to me feels potentially dangerous."
. . .
Bachmeier does not expect schools to teach dedicated classes on responsible AI use. He expects it to be folded into traditional courses instead. He advocated for technology that can set time limits on use and give parents information on how devices are being used, and said monitoring could be especially helpful on devices the school district owns.
Bachmeier said school-issued iPads and laptops give every student comparable technology, and that schools could consider keeping the devices at school to limit screen time. "We may be making parenting harder."
. . .
Later in the same meeting, Michael Marinaccio, executive director of the Center for Responsible Technology, went further. "Students need to be informed that this is distorting their brains." Marinaccio said parents and educators need to make the dangers of chatbots clear to students, and called for legislation with strong age verification for access to chatbots.
The committee also heard testimony that AI can be used to produce child sexual abuse material, and about the long-term effects on victims. Members heard experts on how chatbots work, on recent cases of AI platforms going rogue, and on AI in national security. Data center development, half the committee's name, did not come up.
Committee chair Rep. Jonathan Warrey, R-Casselton, said the next meeting, and potentially the one after it, could go entirely to data center management and physical infrastructure. That meeting is September 1 at the Capitol, and written testimony goes to [email protected].
|
For Legislators: A sitting state education chief told his own legislature to regulate the vendor, not the teacher. All three asks are narrow enough to draft: mandatory disclosure that the user is talking to an AI, age limits on chatbot access, and liability for chatbots that promote self-harm. The committee meets September 1 at the Capitol, and written testimony is open.
For Investors: The regulatory pressure in this hearing pointed at companion chatbot vendors, not at school-facing edtech, and it came from the official who runs the schools. Liability for chatbots that promote self-harm is the exposure to price, and the ask arrived with a specific drafting shape rather than as a general complaint.
For Builders: If you build a companion product, disclosure, age limits, age verification and self-harm liability are the floor being discussed. Three came from the state superintendent; age verification came from the advocate who followed him. School-owned devices are where monitoring and time limits land first.
For Readers: The person in charge of North Dakota's public schools named companion chatbots as his worry. He wants companies to say plainly when a bot is a bot, wants age limits, and wants companies on the hook when a chatbot pushes a kid toward self-harm. He also said keeping school iPads at school might be worth considering, and said schools may be making parenting harder.
Why it matters: This hearing was not a fight over whether teachers can use the tool. It was a question about who is liable for the product. A superintendent with every reason to defend classroom technology pointed at companion chatbots instead, asked for disclosure and age limits, and warned that doing nothing feels dangerous. That committee meets again September 1.
Source: Jeff Beach, "Regulate AI companies, not AI in schools, North Dakota state superintendent says," North Dakota Monitor, August 17, 2026, https://northdakotamonitor.com/2026/08/17/regulate-ai-companies-not-ai-in-schools-north-dakota-state-superintendent-says/; North Dakota Legislature artificial intelligence and data center committee, meeting of August 13, 2026, Minot State University.
|
. . .
EVENTS, NOT SCREENSHOTS. ChatGPT's desktop app for macOS now ships a feature called Computer History, which builds a timeline of a user's clicks and keystrokes that ChatGPT and Codex can reference on a later request. It ships off, and a user has to switch it on. OpenAI says the feature does not capture images, videos, or audio, and instead relies on "events."
The Verge, which reported the feature on August 16, described it as one that "turns your actions into training data, learning how you work, suggesting automations, and even picking up tasks you left half done." The mechanism OpenAI describes is narrower: activity on the machine becomes a timeline the models can consult when a user asks for something.
Users can exclude specific apps and websites from Computer History, and can delete individual entries. Ari Weinstein, a product and engineering manager at OpenAI, said on X that the feature will automatically ignore content in incognito or private browser tabs. There is no reported Windows or Linux version.
In a demo video, Dominik Kundel of OpenAI's Developer Experiences team shows the app looking up the last document he edited, checking whether it had been shared with people over Slack, and delivering a recap of how he spent his morning.
. . .
Windows Recall is the comparison. Microsoft's feature leaned heavily on screenshots, a design that captured whatever happened to be on screen, including the parts nobody meant to record. Computer History, by OpenAI's account, records events instead, a narrower capture.
. . .
OpenAI has not published a retention period for the timeline, and it is not established whether the event data trains models generally or only serves the user's own requests. The Verge's phrase "training data" is the outlet's wording, not a company claim about training. No administrator controls have been described in the reporting.
|
For Legislators: OpenAI has said nothing about who administers the timeline. It is available now on laptops that hold privileged material: a clinician's, a lawyer's, a caseworker's. Opt in protects whoever clicks yes. It does not reach the clinic, firm or agency that answers for the record.
For Investors: The open items are commercial. No described administrator control and no published retention period, both of which regulated buyers ask about before signing. macOS only, with no reported Windows or Linux build, caps the addressable install base for now.
For Builders: App and website exclusion plus per-entry deletion puts the burden on the user to know in advance what should never be recorded. Incognito and private tabs are ignored automatically, per Weinstein, which is a default worth studying. The question to answer before shipping is what your product does when the person operating the laptop is not the person who answers for the file.
For Readers: ChatGPT's Mac app can now keep a running timeline of what you click and type, if you turn it on. It skips incognito and private tabs, and you can block specific apps and sites or delete entries. OpenAI says it records events rather than screenshots, images, or audio. The company has not said how long it keeps the timeline.
Why it matters: Computer History records events, not screenshots, so it captures less than Windows Recall did. There is no published retention period and no administrator control in the reporting. A running record of everything typed on a work laptop is a record someone can later be asked to produce.
Source: Terrence O'Brien, "ChatGPT's Computer History tracks your clicks and keystrokes," The Verge, August 16, 2026, https://www.theverge.com/ai-artificial-intelligence/980742/chatgpts-computer-history-tracks-your-clicks-and-keystrokes; Ari Weinstein, OpenAI, posts on X regarding incognito and private-tab handling.
|
. . .
MICO GOES TO SCHOOL. Microsoft is removing Mico, the animated character it gave Copilot's voice mode in October 2025. Its own support page sets the clock: "This change starts in August and will roll out in waves." The voice survives. The face relocates: "Mico continues in Learn Live, where the character works as a tutor."
On the assistant itself, Microsoft's support page is unambiguous: "Copilot Voice will continue to allow you to speak with Copilot and receive responses." Roughly ten months separated the character's debut from the announcement of its removal.
Mico arrived to comparisons with Clippy, the paperclip that interrupted a generation of Word documents. GeekWire reached further back, to Microsoft Bob, the 1995 assistant that preceded Clippy: "Microsoft's cute little AI blob is going the way of Bob." The Register, on August 14: "Less than a year on, Microsoft tells Mico to pipe down."
Mico is not the only thing leaving in this wave. Copilot Group Chat retires account by account starting August 18, 2026, and existing group chats become individual chats with Copilot.
Copilot Labs goes on a schedule Microsoft describes only as "Timing may vary by account." Labs experiments and generated content will not carry forward unless a user saves them first. Microsoft has not dated the deletion.
. . .
Microsoft has published no usage figures for Mico, and the support page gives no reason for the removal. Learn Live is where the open questions sit. Microsoft has not said whether it has an age floor, what supervision applies to a person talking with a tutor character that answers in real time with animation and voice, or whether it reaches users through schools, as a consumer product, or both.
|
For Legislators: One company moved an animated persona out of its general assistant and into a tutoring product inside the same announcement. Microsoft has published no age floor for Learn Live, no supervision model, and no statement about whether it reaches students through schools. Those are three answerable questions, and the vendor is the one holding the answers. Ask before the wave finishes rolling out.
For Investors: Microsoft launched a character in October and announced its removal ten months later, with no usage disclosure to justify it either way. Discount any valuation that leans on engagement data the vendor declines to publish, and watch whether the character earns a second life in education or quietly stops shipping there too.
For Builders: Microsoft shipped the voice without the character, so a persona layer is removable without touching the voice stack. Note the migration cost too: Labs content "will not carry forward" unless saved first. Export paths are needed before the retirement notice, not after.
For Readers: The little animated blob in Copilot's voice mode is going away starting this month. You will still be able to talk to Copilot and hear it answer. Mico itself is not gone; it stays in Learn Live as a tutor. If you made anything in Copilot Labs, save it now, because it does not survive the update.
Why it matters: Microsoft gave its general assistant a face, took it away after roughly ten months, kept the voice, and moved the character into a product where it works as a tutor. No usage figures, no stated reason, no age floor for Learn Live.
Source: Microsoft Support, "Frequently asked questions about retired Copilot features," read August 17, 2026, https://support.microsoft.com/en-us/microsoft-365-copilot/frequently-asked-questions-about-retired-copilot-features; "Less than a year on, Microsoft tells Mico to pipe down," The Register, August 14, 2026, https://www.theregister.com/ai-and-ml/2026/08/14/less-than-a-year-on-microsoft-tells-mico-to-pipe-down/5287793; "Farewell, Mico: Microsoft's cute little AI blob is going the way of Bob," GeekWire, 2026, https://www.geekwire.com/2026/farewell-mico-microsofts-cute-little-ai-blob-is-going-the-way-of-bob/.
|
. . .
SIX YEARS, TWO SHUTDOWNS. A 10-year-old boy learned to breathe out through his lips until he buzzed like a bee. A blue robot taught him that, for the times he felt anxious. For anger it taught dragon breathing, and for low energy, bunny sniffing. Xander, who is neurodivergent, has had the robot, Moxie, in his house for six years.
Moxie was built by Embodied Inc. It stands 15 inches tall and looks like a blue, legless astronaut: a screen showing eyes and a mouth, flipper-like arms that gesture as it speaks. Xander calls it "her." Xander on what it is to him now: "I still use her when I feel like I need someone to talk to. But, like, it's not human."
. . .
The talking ran on Embodied's servers, so the robot worked only while somebody paid for them. Embodied Inc. ran out of money and ceased operations in 2024, and the robots became inoperable. A new investor revived the product in 2025. The revived company announced its own closure in June 2026.
Moxie sold for $1,499 plus $40 a month, and later for $800. Nothing in that purchase bought a service life. No escrow of the server software, and no obligation on either owner to keep the servers up.
What keeps Xander's robot running is a workaround. Justin Beghtol, Embodied's technical director, built OpenMoxie so families could keep their robots going. Beghtol on what the collapse did to those families: "Seeing them traumatized by this financial failure was tough."
Nobody owed them that.
. . .
Embodied was cofounded by Maja Mataric, a USC robotics professor. Moxie was sold as a consumer product for children's social and emotional development, not as a cleared medical device, not prescribed and not clinically validated.
|
For Legislators: Nothing in this record obligated either owner to keep the servers running, or to hand families a way to run the robot themselves. These are ordinary consumer-protection levers, not new ones: disclose expected service life at the point of sale, escrow the server software, require an offline mode that survives the company. A $1,499 device with a $40 a month subscription shipped with none of the three.
For Investors: One product line, two owners, two failures. Embodied ceased operations in 2024, and the investor who revived it in 2025 announced closure in June 2026. Hardware sold for $1,499, later $800, has to carry a server bill for as long as the device sits in a child's bedroom. A monthly subscription did not survive two shutdowns.
For Builders: Moxie was a screen and two flippers. The talking was a server bill, and when the servers went dark the robot went with them. Build the local mode before shipping, and publish it, so the fallback is not a favor from a former employee.
For Readers: If a device in your house talks back, ask at purchase what happens when the company stops. Does it work without the internet? Would the software be released if the business folded? Xander's robot still answers him because Beghtol built the workaround, and that was never part of what his family bought.
Why it matters: A child spent six years talking to a machine that stayed alive only while a company paid for servers, and both companies that owned it failed. No escrow, no continuity requirement, no duty on anyone to keep it running. A former employee's workaround is why it still answers. The next company selling a talking device into a child's bedroom operates under the same absence of obligation.
Source: "What happens when a kid's robot best friend dies?", MIT Technology Review, August 17, 2026, https://www.technologyreview.com/2026/08/17/1141568/moxie-when-kids-robot-best-friend-dies/
|
|
