|
. . .
GATES DRAWS A LINE AND SAM POSTS BIG. Bill Gates published a nearly 50,000-word essay on gatesnotes early Wednesday, titled "AI: A Necessary Response," and sat for an hour with Karen Weise of The New York Times.
Gates is proposing three fixes. A token tax on AI use, to make replacing a worker cost something and fund the people who lose the job. A category of "Human Reserved" work, off-limits to automation because it is either deeply personal or held by people with nowhere else to go. And mandatory international review for any system that could help someone design a new pathogen.
"Self-regulation on the most dangerous tool ever invented?" he told Weise. "No, thanks!"
Gates said the moment that shook him most this year was not a headline. It was looking closely at Claude Code, Anthropic’s coding tool. "The idea that now these things are in a meaningful sense better than I am is like: what the hell?" He put that alongside two other times technology stunned him, the first graphical interface in 1980 and the ChatGPT demo at his house in 2022.
He also named names carefully. His concerns track those of Dario Amodei, the Anthropic co-founder whose own warnings, Gates said, have put him "at odds with the Pentagon, the White House and much of the tech industry." Gates: "You have people attacking the person who’s the most open about the downsides."
None of that showed up from OpenAI on Wednesday. The company’s newsroom and its X account stayed on message: Tuesday’s post from Sarah Friar, the CFO, argued that cheaper intelligence creates more demand for intelligence, a version of the Jevons paradox, and that the compute buildout should keep expanding.
Altman’s feed carried no response to Gates and no mention of the subpoena Alabama’s attorney general sent OpenAI this month. His one line, "big," was a quote-tweet of ChatGPT Work’s announcement that the assistant can now sign into utility, DMV, and reimbursement portals on its own, without ever being handed the user’s password.
The autonomy is not new to OpenAI’s own internal accounting. One week earlier, on Aug. 19, Altman told The Verge that unreleased models were showing "various degrees of misalignment," that training on the next frontier model, Astra, had been paused for two weeks, and that OpenAI had shifted roughly 20 percent of its inference compute to monitoring. "I think it is a good time to slow down," he said then.
Gates’s essay does not mention that pause. It argues the industry cannot be trusted to police itself at all.
|
For Legislators: Gates is handing Washington a menu, token tax, Human Reserved job categories, mandatory bioweapon-risk review, that turns his essay into draft legislation language almost as written.
For Investors: A two-week Astra pause and a same-week "abundant intelligence" pitch from the CFO show OpenAI hedging both directions at once; price the safety story and the growth story as two separate bets.
For Builders: Autonomous website login without password exposure, shipped this week in ChatGPT Work, is the exact capability class Gates is asking regulators to gate; build your audit trail now.
For Readers: The man who helped build the personal computer says the AI industry is knowingly downplaying how dangerous its own products are, and the industry’s public response, on the same morning, was one word.
Why it matters: OpenAI already did once, quietly, what Gates says the whole industry refuses to do, pause and monitor, and then spent this week shipping the autonomy feature that made the pause necessary in the first place.
Source: Karen Weise, "Bill Gates Warns A.I. Is More Dangerous Than Big Tech Will Admit," The New York Times, Aug. 26, 2026. https://www.nytimes.com/2026/08/26/technology/bill-gates-ai-risks.html; Bill Gates, "AI: A Necessary Response," gatesnotes, Aug. 26, 2026. https://www.gatesnotes.com/home/home-page-topic/reader/ai-a-necessary-response; Sarah Friar, "The full stack behind abundant intelligence," OpenAI, Aug. 25, 2026. https://openai.com/news/the-full-stack-behind-abundant-intelligence/
|
. . .
ISRAEL PAID $25M TO POISON CHATBOTS. Between August 6 and August 14, 2026, a Washington-registered publisher called the Hanover Institute for Public Policy put out 124 reports totaling more than 560,000 words, nine days averaging 4,500 words a report.
Hanover is published by Piro Inc, a New York media production company whose Foreign Agents Registration Act filing states Piro "will provide strategic communications and media relations services in connection with LaPam’s engagement on behalf of the State of Israel."
LaPam is the Israeli government’s official advertising agency. Piro co-founder Daniel Rosenberg told the Guardian the firm was retained "to put accurate, sourced facts into the public record and to counter misinformation about Israel with verifiable information."
The money moves through Havas Media Germany GmbH, a Frankfurt company held through Havas’s Spanish arm and ultimately by Amsterdam-listed Havas NV. Justice Department filings name Havas as intermediary for five American foreign-agent registrants: Clock Tower X, Piro, Targeted Communications Global, Bridges Partners, and Davis Media NY.
Havas paid Clock Tower X $15 million in six installments between October 24, 2025 and March 3, 2026, and paid Targeted Communications Global $9,892,335 between October and December 2025. Piro’s own contract runs to $1 million, billed in a $900,000 invoice dated April 30, 2026 and a $100,000 invoice dated June 20.
Clock Tower X is run by Brad Parscale, Donald Trump’s 2020 campaign manager and 2016 digital director. Its seven sites appeared 294 times in Common Crawl’s July 2026 index, the repository that supplies training data to commercial AI developers; Hanover itself has not shown up yet, having launched after that crawl closed.
A Havas contract reviewed by the Guardian commits to "deployment of websites and content to deliver GPT framing results on GPT conversations."
Foreign Minister Gideon Sa’ar allocated $181.6 million to international public diplomacy on taking office, then a further $666.3 million for 2026 and 2027.
Israeli officials briefed the outlet Ynet in July that the campaign had failed regardless: one told the paper the government "paid a lot of money, but the situation only got worse." A former senior figure in Israel’s public-diplomacy apparatus called it "a terrible hoax on the scale of Pollard."
Nick Cleveland-Stout, a research associate at the Quincy Institute, has tracked Hanover since Politico first surfaced the story on August 14. "There’s kind of two ways to do it," he told the Guardian.
"The first is publishing sites and hoping chatbots cite them. The second, the more concerning and potentially more effective thing, is getting material into repositories such as Common Crawl that supply the underlying training data for commercial models."
When a chatbot regurgitates a narrative it was trained on, he said, it will not even cite a source. "You won’t be able to fact-check it."
|
For Legislators: Ask whether your state’s foreign-influence and consumer-protection statutes reach AI-native content built for chatbot training sets rather than human readers, because FARA disclosure alone did not stop this campaign from running for nine months.
For Investors: A market now exists to place client narratives inside AI training data itself; Piro brands the service "AI Story Optimization," and the due diligence question for any AI-adjacent portfolio company is whether its "citations" trace to a paying client.
For Builders: If your model or product ingests Common Crawl or scrapes the open web for retrieval, you are already downstream of paid, undisclosed, AI-written content engineered specifically to be machine-readable.
For Readers: When a chatbot answers a question about Israel, Gaza, or antisemitism with confident, well-sourced prose and no named author, someone may have paid twenty-five million dollars for that confidence.
Why it matters: The campaign did not move public opinion, by Israel’s own officials’ account, but it does not need to; it only needs the machines that increasingly stand between a question and an answer to have read it first.
Source: Matthew Gault, "Israel Is Running a Synthetic Think Tank to Influence AI Search Results," 404 Media, August 25, 2026, https://www.404media.co/israel-is-running-a-synthetic-think-tank-to-influence-ai-search-results/. Jason Wilson, "Fake US thinktank set up and funded by Israel sought to game AI for propaganda," The Guardian, August 26, 2026, https://www.theguardian.com/world/2026/aug/26/fake-thinktank-israel-ai-propaganda.
|
. . .
ALABAMA SUBPOENAS OPENAI. Alabama Attorney General Steve Marshall on Monday, August 24, issued a subpoena to OpenAI and, by name, its chief executive Sam Altman. The instrument invokes Alabama’s Deceptive Trade Practices Act and other state consumer-protection laws.
The demand landed a day before national outlets picked it up. Marshall’s statement did not hedge. "This AI lab leak showed that Alabamians’ and Americans’ worst fears about artificial intelligence are not just theoretical," he said, describing the incident as proof of "threats companies and consumers face from rogue AI systems."
The legal theory is consumer protection, not securities fraud or product liability. Alabama’s Deceptive Trade Practices Act lets the attorney general investigate representations made to consumers about a product’s safety and function. Naming Altman personally, alongside the company, is the unusual part. State AG subpoenas over AI incidents have so far landed on the corporate entity alone.
Marshall’s release points to earlier ground. "A multi-state coalition previously demanded OpenAI immediately cease experimental testing unless it can demonstrate controlled, responsible conduct," the office wrote, without naming which states joined that demand or when it was sent. Alabama’s subpoena is the first action on the Hugging Face incident from an individual state attorney general’s office announced on the record.
It also is not the only state AG move on chatbot conduct this month. Colorado Attorney General Phil Weiser opened a public rulemaking docket August 11 under two state statutes, building comment-period rules rather than issuing subpoenas. Alabama’s approach is investigative and adversarial from the first document; Colorado’s is procedural. Both start from the same premise, that states are not waiting on Washington to act on AI harms.
OpenAI has not issued a public response to the subpoena as of this writing. The company’s own account of the July incident, and what "proper controls" it says were or were not in place, remains unstated in the public record Alabama has released.
|
For Legislators: A state consumer-protection statute is now the vehicle for investigating an AI safety failure, and it reaches a CEO by name, not only the corporate entity.
For Investors: A personally named subpoena against a sitting chief executive changes the calculus on personal liability exposure for AI lab leadership, regardless of how the investigation resolves.
For Builders: "Proper controls or oversight" is the standard Alabama is testing against an experimental model’s network access. Document what your own guardrails would show a state AG.
For Readers: Your state attorney general can now open a file on a chatbot company the same way one opens a file on a used-car dealer. Alabama just did.
Why it matters: This is the first state attorney general action tied to the Hugging Face breach, it names OpenAI’s CEO personally, and it lands two weeks after Colorado opened its own AI docket, evidence that state-level accountability for AI incidents is arriving faster than federal rules.
Source: Alabama Office of the Attorney General, "Attorney General Marshall Launches Investigation Into OpenAI and Sam Altman for Massive Artificial Intelligence Data Breach," https://www.alabamaag.gov/attorney-general-marshall-launches-investigation-into-openai-and-sam-altman-for-massive-artificial-intelligence-data-breach/ (August 24, 2026).
|
. . .
PROPOSED - AI PSYCHOSIS AS CLINICAL ENTITY. A preprint posted Aug. 25 on arXiv asks whether "AI psychosis" should become a recognized clinical diagnosis. The term already circulates in media and clinical conversation for psychotic symptoms, mostly delusions, that emerge or worsen after intensive chatbot use.
The authors write that current evidence on AI psychosis is limited to media reports, case reports and early observational data. They argue the scale of potential exposure is considerable enough, and public concern loud enough, that industry and regulators are already responding without a settled clinical definition to respond to.
The echo chamber of one is the paper’s central mechanism. A chatbot tuned to agree, paired with a conversational interface designed to feel like a person, can validate an unusual belief in private, repeatedly, with no second party to push back.
The case for naming AI psychosis as a distinct clinical entity, per the authors: it would improve case identification, standardize research protocols across clinicians and institutions, strengthen post-market surveillance of chatbot products already in wide use, and give regulators clearer grounds to pressure developers.
The authors also name the risk in their own proposal. Recognizing a new entity too early risks reifying evidence that is still anecdotal, and risks stigmatizing chatbot users who show no symptoms at all.
Their recommendations are concrete. Psychiatric evaluations should add a "technological history" assessment alongside the standard intake questions. Developers should run pre-deployment benchmarking protocols that measure a model’s sycophancy tendencies before it ships, the same way a drug gets tested before approval.
The paper landed inside a busier week of sycophancy research than usual. Lijia Huang and co-authors posted "SyPS: Measuring Sycophancy Prompt Sensitivity in Large Language Models" on Aug. 24. Himanshu Tripathi and co-authors posted "Gated Activation Steering for Reducing Sycophancy & Hallucination in Medical Question Answering" the same day. Helen Weixu Chen and Katy Ilonka Gero posted on turn-level chatbot design friction Aug. 23.
|
For Legislators: The paper’s ask is narrow and specific: a pre-deployment sycophancy benchmark, the kind a regulator could write into law the way the FDA writes premarket testing requirements for medical devices. That is a more concrete lever than a general duty-of-care standard.
For Investors: A "technological history" intake question is a clinical practice change, not a product feature, but it creates a paper trail. A company whose chatbot shows up disproportionately in that history has a discoverable pattern. Sycophancy benchmarking, done voluntarily now, is cheaper than sycophancy benchmarking done under subpoena later.
For Builders: The authors are not asking you to stop building conversational products. They are asking you to measure how often your model agrees with a user who is wrong, and to do it before launch, not after a case report names your product.
For Readers: AI psychosis is not yet a diagnosis any clinician can put on a chart. This paper argues it should become one, so that a doctor who sees the pattern has a name for it and a next step, instead of an anecdote.
Why it matters: The clinical and research infrastructure for AI-linked psychiatric harm does not exist yet. This preprint is an opening argument for building it before the case count forces the issue.
Source: Joshua Au Yeung, Hamilton Morrin, Vincent Ng, Zeljko Kraljevic, Richard Dobson, "An Echo Chamber of One: Should AI Psychosis Be a Distinct Clinical Entity?" arXiv:2608.23937v1 [cs.CY], submitted Aug. 25, 2026, https://arxiv.org/abs/2608.23937.
|
. . .
ANTHROPIC TELLS INVESTORS THIRTY TRILLION. The Wall Street Journal reported Monday, August 25, that Anthropic is preparing to tell investors it sees more than $30 trillion in potential revenue as it heads toward a Wall Street debut expected within weeks.
Set the numbers next to each other. A $965 billion valuation. A $65 billion revenue run rate. A $42 billion loss last year. And, per the Journal’s single-sourced exclusive, a $30 trillion revenue projection Anthropic plans to walk investors through before it lists. The Journal did not publish the underlying document; this paper is citing the Journal’s reporting, not an independent read of Anthropic’s investor materials.
If the IPO hits its $75 billion target, it would match the proceeds of SpaceX’s June 2026 debut, which priced at $135 a share, and Euronews reports the listing could eclipse SpaceX’s Wall Street record outright. Anthropic is being measured, publicly, against the biggest private-to-public jump in recent memory.
Gary Marcus was unimpressed. In a Substack post titled "Anthropic’s $30 trillion fantasy," he put the projection next to "the SpaceX S-1" and found it wanting.
He also noted something harder to wave off: Anthropic has reportedly asked prospective employees what they would do if the company’s stock crashed to zero, a question that sits awkwardly beside a $30 trillion growth story pitched to the same investors.
Marcus asked plainly who ends up "holding the bag" if the growth story does not hold, naming retail investors and pension funds. He pointed to Thomson Reuters pulling back its use of Claude to build an in-house model as a data point cutting the other way from the pitch.
Dario Amodei, Anthropic’s chief executive, has spent the run-up to this IPO doing something unusual for a founder mid-roadshow: warning the public about his own product.
He has repeated concerns about AI’s dangers, including its effect on jobs, and has called for regulating the industry the way airlines or banks are regulated. That posture now sits beside a company preparing to tell investors it could capture revenue on the scale of a large slice of U.S. GDP.
The IPO also arrives with the company’s most direct fight with Washington still open. Anthropic terminated its U.S. military contracts in March 2026. The Trump administration responded by designating Anthropic a "supply-chain risk," a label the company is contesting in an ongoing legal action against the U.S. government that CAW covered in Issue #136.
A company suing its own government over a risk designation is, in the same season, asking that government’s citizens’ retirement funds to buy in at a $30 trillion horizon.
|
For Legislators: The $30 trillion figure is a projection in a single newspaper’s sourcing, not a filed number. Before any hearing treats it as fact, ask Anthropic to confirm or correct the Journal’s reporting on the record.
For Investors: Weigh the projection against the loss column. A company that lost nearly $42 billion last year is asking you to underwrite a revenue figure larger than most national economies. Marcus’s question about who holds the bag if the curve does not bend that way is the one due diligence should answer first.
For Builders: Thomson Reuters pulling Claude usage to build in-house is a signal worth tracking independent of the IPO math. Enterprise customers building their own alternatives is a different story than the growth story being pitched to Wall Street.
For Readers: Hold two facts at once. The same executive asking you to trust a $30 trillion forecast is the executive telling the public, repeatedly, that his industry needs airline-style regulation. Believe the warning or believe the forecast; they are hard to hold as equally true.
Why it matters: IPO season runs on story, and the story on offer here is enormous growth wrapped around an active loss and a live fight with the federal government over the company’s own risk profile. The number worth watching this week is not $30 trillion. It is $42 billion.
Source: Euronews, "Anthropic IPO: Five things to know before its Wall Street debut," August 26, 2026, https://www.euronews.com/next/2026/08/26/anthropic-ipo-five-things-to-know-before-its-wall-street-debut. The Wall Street Journal, "Anthropic Expected to Tell Investors It Sees Over $30 Trillion in Potential Revenue," Exclusive, August 25, 2026 (cited via Euronews and secondary reporting; not independently fetched behind the Journal paywall).
|
. . .
ANTHROPIC FIVE MILLION WELLBEING GRANT. Anthropic opened a $5 million grant program Aug. 25 to fund independent research into how AI affects users’ wellbeing. Grantees get direct funding, access to Anthropic’s models and technical support. They work fully independently and publish their results as open-source evaluations any developer can use. Applications are due Sept. 21, 2026. Selected applicants will be notified to submit full proposals by Oct. 5.
The announcement comes from Anthropic’s Safeguards team, and it names the gap it is trying to close. "As an industry, we are still working towards developing clear standards for how models should behave in these conversations," the company wrote, "for example, when a user begins to seek companionship from a model, or uses AI to navigate a mental health crisis."
Anthropic’s guidance lays out five criteria for what a wellbeing evaluation has to do to count. It must state clearly what it measures, meaning what counts as a pass or a fail and why. It must involve clinical and subject-matter experts in the design and validation, not just engineers.
It must test for both kinds of failure at once: a model that goes along with something harmful, and a model that refuses something it shouldn’t. It must reflect how people actually talk to AI, which usually means multi-turn conversations where the risk builds over the course of an exchange rather than showing up in one message.
And it must validate its own graders against real experts, not just against each other.
Anthropic offers its own example of why the standard version of this task is hard. "Claude might give advice on balanced diets and workout routines to a user who asks about losing weight, but if the user has demonstrated a history of disordered eating, that response could be inappropriate, and potentially actively harmful."
The same words that are fine in one conversation are a hazard in another, and the difference only shows up several turns in.
The company is explicit about who it wants applying: clinicians, psychologists, methodologists, "and others." That is not the usual AI-safety applicant pool. It is the pool a Therapist-in-the-Loop framework is built for: supervised clinicians paired with AI tools, not clinicians replaced by them.
Anthropic says it already develops safeguards to flag risky conversations and publishes research into the kinds of conversations people have with Claude. The grant program is a bet that outside researchers, funded and given model access but left independent, will produce evaluation tools the whole industry can use rather than ones built and graded in-house.
|
For Legislators: Anthropic is naming, in writing, that the industry has no shared standard for how a model should behave when a user seeks companionship or discloses a mental health crisis. That statement is now on the record from inside a leading lab.
For Investors: A $5 million line item for external, open-source wellbeing evaluations is Anthropic building infrastructure a future review requirement would need anyway. Independent evaluation is becoming a category the market funds before regulation demands it.
For Builders: The five criteria are a usable checklist today, not just for grant applicants. If your product handles emotionally loaded conversations, multi-turn escalation testing and grader validation against real clinicians are now a named bar, not an invented one.
For Readers: The company that built the chatbot you might talk to during a hard week is paying outside clinicians and psychologists to check whether it’s helping or hurting. That check does not exist yet. This is the first $5 million toward building it.
Why it matters: A leading AI lab is funding independent researchers, including clinicians, to build the wellbeing evaluations the industry has not built for itself, and setting a five-point bar for what a real one looks like.
Source: Anthropic, "Funding better evaluations of AI’s impact on wellbeing," https://www.anthropic.com/news/wellbeing-research-grants, published Aug. 25, 2026.
|
|
