# Why Everyone Started Saying AI Is Dangerous This Month

By Ethan Rogers. Published 2026-09-16. 9 min read.
Canonical: https://ethanplus.ai/blog/the-money-behind-the-ai-doom-debate

If you missed it, here is what happened. In the span of about a week, the CEOs of the biggest AI companies said their own technology might be dangerous enough to humanity that development needs to be slowed down and put under outside supervision. Kamala Harris said the same thing. So did Barack Obama, with caveats. Newspapers ran explainers on whether AI could wipe us out. People started calling it the "AI doom" moment.

Then two threads went viral, one from [Kevin Bass](https://x.com/kevinnbass/status/2099621874279817638) and one from [Rooke Poole](https://x.com/rookepoole/status/2099660837388501127), claiming none of it was organic. The argument: the same handful of funders sit behind the safety organizations, the outside evaluators, the journalism, and the political messaging, and the whole thing is a paid campaign to hand a few companies control over the rules.

I will be honest about where I started. My instinct says they are onto something. I have spent years around big industries, and I have watched this pattern before: an incumbent funds the research that defines a danger, funds the groups that spread the warning, then offers to run the response. It is old. It works. And when the biggest labs in the world all ask for the same rules in the same week, my default is not to assume coincidence.

But an instinct is not a finding, and I build companies on these models, so I care about getting this right. I went and read the disclosures. Not the threads. The actual filings, grant tables, and announcements.

Here is what the record supports, where it stops, and what I would need to see to call it what the threads call it.

## The weekend everyone noticed was set up in July

The statements did not come out of nowhere.

In July, an employee statement called *[Pacing the Frontier](https://www.pacingthefrontier.com/)* asked the U.S. government to back an international effort to manage the pace of automated AI development. As of September 16 it showed 1,386 signatories, including John Schulman, Ilya Sutskever, Jakub Pachocki, Jared Kaplan and Shane Legg. The page credits Guidelight AI Standards and Encode AI for organizational support.

On [September 12](https://www.reuters.com/business/what-amodei-altman-musk-have-said-about-ai-risks-stoking-doom-fears-2026-09-14/), Dario Amodei published *[We Must Pace the Frontier](https://darioamodei.com/post/we-must-pace-the-frontier)*: embedded third-party evaluators, coordination among labs in democratic countries, international coordination. He named METR as an example evaluator and committed Anthropic to giving one embedded access. [Altman agreed](https://www.reuters.com/legal/litigation/openai-ipo-will-not-happen-2026-amid-ai-safety-fears-altman-says-2026-09-12/). [Musk said "Dario is right."](https://www.reuters.com/business/what-amodei-altman-musk-have-said-about-ai-risks-stoking-doom-fears-2026-09-14/) Hassabis endorsed the direction with caveats.

Two days later [Kamala Harris posted](https://bsky.app/profile/kamalaharris.com/post/3mviz63l3fk2z) that "we must responsibly slow the pace of frontier AI development." [Obama weighed in the same day](https://www.pcgamer.com/software/ai/we-cant-stuff-ai-back-in-a-box-says-former-us-president-barack-obama-confirming-hes-not-a-doomer-while-still-calling-for-more-government-ai-regulation/), but his line was that company-written rules are not enough and the public has to be involved. [Zuckerberg pushed back](https://www.reuters.com/business/metas-zuckerberg-says-ai-labs-have-enough-incentive-build-safely-2026-09-16/) and said competition and liability already give labs reason to build safely. [The UN chief called for global cooperation](https://apnews.com/article/6ae720a081ce4d5ede35837ca35b8460). [The Guardian ran a piece](https://www.theguardian.com/technology/2026/sep/16/ai-companies-collude-antitrust-laws) arguing that letting labs coordinate under an antitrust waiver is a dangerous idea.

So yes, a lot of statements in a short window. Also real disagreement about what should happen next. A timeline proves sequence. It does not prove anyone was directed.

## Some of the distribution is organized, and they say so

You do not have to infer a communications strategy from timing when the organization publishes it.

[ControlAI's April plan](https://controlai.org/blog/prevent-asi-plan) states the goal plainly: an international prohibition on artificial superintelligence. It describes a sustained push to get extinction-risk arguments in front of the public, journalists, creators and lawmakers. It counts 22 media placements and 14 creator collaborations, some paid, some not, and names Hank Green and Rational Animations. The $50 million figure in the title is a proposed budget, not money spent.

The lawmaker contact is real. In a [January House of Lords debate](https://hansard.parliament.uk/lords/2026-01-08/debates/5CAD7DC7-3B7E-4925-85C4-88354195031E/AISystemsRisks), Lord Fairfax thanked ControlAI for help preparing his speech and Baroness Ritchie mentioned briefings from its policy advisers.

Anthropic, separately, has [disclosed $40 million to Public First Action](https://www.anthropic.com/news/donation-public-first-action), restricted to public education and policy work, not candidate elections. The company says the point is to raise the profile of the policy debate.

And the [Future of Life Institute reports](https://futureoflife.org/about-us/finances/) $46 million in 2025 spending, including $27 million in grants and $6 million on media and outreach. FLI also says it takes no money from Big Tech or from companies trying to build AGI. That matters, because a risk-focused nonprofit's spending is not the same thing as an AI company's spending.

That is what "funded messaging" looks like in the record. Published goals, real budgets, acknowledged briefings. Whether any specific message was accurate or misleading is a separate question you answer by reading the message.

## Journalism funding is its own thing

A paid advocacy placement and a supported reporter are not the same relationship, and collapsing them is how the threads went wrong.

The [Tarbell Center for AI Journalism](https://www.tarbellcenter.org/about) lists Coefficient Giving, Longview Philanthropy and the Survival and Flourishing Fund among donors above $1 million, and says donors have no editorial control. [Its registry](https://www.tarbellcenter.org/) lists Robert Hart's September 9 Verge story on an Anthropic safety lead's extinction-risk estimate, and Aisha Down's September 15 Guardian explainer on whether AI could wipe us out. Both are fellows.

So the support is disclosed. What it does not tell you is who chose the topic, who edited it, or whether the reporter could have published a contrary finding. Funding shapes which questions get time. That is a mechanism worth watching. It is not evidence of interference in these two stories.

Same standard for the other direction, by the way. A company funding upbeat AI coverage deserves the same scrutiny.

## Where the same names show up twice

This is the part with actual overlap.

Anthropic's [May 2021 Series A](https://www.anthropic.com/news/anthropic-raises-124-million-to-build-more-reliable-general-ai-systems) named Jaan Tallinn as lead investor and Dustin Moskovitz as a participant. Now look at the [Survival and Flourishing Fund's 2025 table](https://survivalandflourishing.fund/): Tallinn is the source behind a $516,000 recommendation to Encode AI, $120,000 plus a $428,000 match to METR, and $583,000 plus a $200,000 match to Tarbell.

Read that carefully. An early lab investor is named in support of an evaluator, a journalism fund, and one of the two organizations behind the July statement. Those are recommendations and conditional matches, not bank records. But the overlap is real and it is relevant to independence.

What is not established is the bigger claim. [Kevin Bass's own notes](https://github.com/kevinnbass/metr-money-figure/blob/master/NOTES.md) describe his $7.7 billion figure as a ceiling built from an ownership bound and a financing valuation, with the recipient of the donated shares unresolved and no direct METR entry in the grant index he checked. So "a proven $7.7 billion Anthropic stake at Good Ventures bankrolls METR" is not something the records show. METR's [August update](https://evals.alignment.org/blog/2026-08-14-funding-update/) lists about $71 million in commitments from the Audacious Project, Jane Street individuals, Schmidt Sciences, Pew and Packard, among others.

An investor overlap, a grant, and dependence on one asset are three different findings. The viral diagrams draw them as one arrow.

## "Independent evaluator" is a contract, not a label

Amodei's proposal leans on embedded third-party evaluators. Whether that means anything depends on the terms.

METR's [conflict-of-interest policy](https://metr.org/coi-policy.pdf), dated August 28, rejects frontier-lab donations, allows free compute, and says that as of that date it had never been paid for assessment work by a lab. It leaves the door open to future payment as long as the amount does not depend on the result.

Twelve days later, [Anthropic announced a signed agreement with METR](https://www.anthropic.com/research/alignment-assessment-cybersecurity-incidents) to investigate four cybersecurity incidents, with transcript and employee access, eight weeks with an option to extend. Payment terms were not disclosed. An August policy cannot tell you the finances of a September deal.

The questions I would want answered for any evaluator: who picked them, what access was refused, can the lab delay publication, and who resolves disagreements. And separately, does the research hold up. METR's [task-horizon work](https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/), for example, measures the task length at which a model succeeds half the time. That is not "how long an agent can run unsupervised" and it is nowhere near a probability of extinction. Watch how a narrow result becomes a headline. That is where claims get misleading, and you find it by following the method, not the donor.

## The money story is the one that actually matters

Here is where I think the threads pointed at the right thing for the wrong reason.

The [FTC's January 2025 study](https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2025/01/behind-ftcs-6b-report-large-ai-partnerships-investments) of Microsoft and OpenAI, Amazon and Anthropic, and Google and Anthropic described deals that require the lab to spend a large, specified chunk of its investor's money on that investor's cloud. The investor is also the supplier. Since then the numbers have only gotten bigger:

| Deal | What was announced | What the number is not |
|---|---|---|
| **[Microsoft, NVIDIA and Anthropic, November 2025](https://www.anthropic.com/news/microsoft-nvidia-anthropic-announce-strategic-partnerships)** | Up to $5 billion from Microsoft and $10 billion from NVIDIA; Anthropic commits to $30 billion of Azure compute. | Not $45 billion of independent demand. |
| **[Amazon and OpenAI, February 2026](https://openai.com/index/amazon-partnership/)** | $50 billion of investment, $35 billion of it conditional; AWS agreement expanded by $100 billion over eight years. | Not cash that has settled. |
| **[Amazon and Anthropic, April 2026](https://www.anthropic.com/news/anthropic-amazon-compute)** | $5 billion now, up to $20 billion more; Anthropic commits more than $100 billion over ten years to AWS. | A purchase obligation is not revenue. |
| **[NVIDIA and CoreWeave, September 2025](https://www.sec.gov/Archives/edgar/data/1769628/000176962825000047/crwv-20250909.htm)** | NVIDIA buys unsold capacity through 2032, initial value $6.3 billion. | A backstop is not customer demand. |

None of this is fake. Servers get delivered. But a dollar of financing that becomes a dollar of cloud spend that becomes a dollar of chip revenue is still one dollar of new money, and the whole chain depends on the next round arriving. Valuation is not cash. Backlog is not revenue. A conditional commitment is not a deposit.

And notice what both sides of the story share. A lab can say "this technology is extraordinarily powerful" to investors and "this technology is extraordinarily dangerous" to regulators, and the two claims rest on the same premise. That is not a contradiction. It is a reason to test the premise instead of taking either pitch at face value.

## What I actually think

I still think this looks like a campaign. I cannot prove it is a deceptive one, and I am not going to pretend the record does that work. Similar timing does not, and neither does a donor list. What the record does show is enough to keep me skeptical: the same early investor behind the evaluator, the journalism fund and the statement organizers; advocacy groups that publish their outreach plans and get thanked in Parliament; a lab that funds the policy debate, names its preferred evaluator, then signs a deal with it whose terms we cannot see; and a financing loop where the investor is also the supplier and everyone's numbers depend on the next round landing.

Every one of those is a documented, legal, disclosed relationship. That is exactly how the campaigns I have seen before were built. Nobody needs a secret memo when the incentives all point the same way.

So here is my standard, and I will hold myself to it. To call it deceptive I would need one of three things: a claim its author knew was wrong, a sponsorship that was hidden, or an evaluation whose access, method or publication was shaped by the people being evaluated. Show me any of those and I will say it plainly. Until then, the honest version is that a set of well-funded actors with a shared interest are all pushing the same direction, and that is reason enough to read the proposal like a contract.

Because the proposal is where this gets real for anyone building on these models. Who appoints the evaluator. Who sets the threshold. Whether a new company gets the same rules as the incumbents who wrote them. I am fine with oversight. I am not fine with a system where the people defining the danger also approve the response, and the rest of us build on whatever they leave open. That is the part I am watching, and it is the part the disclosures already justify watching.

*Reporting note: This piece is based on public disclosures, grant recommendations, policy documents, research descriptions, securities filings, public statements, and reporting that identifies those statements. Some original social posts were available only through reporting. No requests for comment were sent.*

## Sources and original documents

Links appear beside the relevant claims above. This register preserves the source descriptions and dates used in the article. Original posts, organizational disclosures, reporting, and opinion are not interchangeable forms of evidence.

1. Anthropic, “Donating another $20 million to Public First Action,” July 21, 2026. [Anthropic: Public First Action funding and policy statement](https://www.anthropic.com/news/donation-public-first-action)

2. ControlAI, “Preventing extinction from ASI on a $50M yearly budget,” April 20, 2026; organizational strategy and self-reported outreach figures. [ControlAI: outreach strategy and proposed annual budget](https://controlai.org/blog/prevent-asi-plan)

3. Originating X threads: [Kevin Bass: original X thread](https://x.com/kevinnbass/status/2099621874279817638); [Rooke Poole: original X thread](https://x.com/rookepoole/status/2099660837388501127). Cited for the investigators’ own claims and the starting point of this investigation, not as independent proof of every allegation. The supplied research brief was a lead list.

4. “Pacing the Frontier,” July 2026; signatory display and organizer credits accessed September 16, 2026. [Pacing the Frontier: employee statement, signatories, and organizer credits](https://www.pacingthefrontier.com/)

5. Dario Amodei, “We Must Pace the Frontier,” September 2026. [Dario Amodei: We Must Pace the Frontier](https://darioamodei.com/post/we-must-pace-the-frontier)

6. Reuters, “Tech leaders, governments split over ‘AI doom’ fears,” September 14, updated through September 16, 2026; and Reuters, “OpenAI's Altman won't do IPO this year, calls AI extinction risk ‘unacceptable’,” September 12, 2026. [Reuters: industry and government responses](https://www.reuters.com/business/what-amodei-altman-musk-have-said-about-ai-risks-stoking-doom-fears-2026-09-14/) ; [Reuters: Altman’s September 12 remarks](https://www.reuters.com/legal/litigation/openai-ipo-will-not-happen-2026-amid-ai-safety-fears-altman-says-2026-09-12/)

7. Kamala Harris, Bluesky post, September 14, 2026, 21:04:37 UTC. [Kamala Harris: original Bluesky post](https://bsky.app/profile/kamalaharris.com/post/3mviz63l3fk2z)

8. PC Gamer, Andy Edser, reporting Obama's September 14 statement, September 16, 2026. [PC Gamer: reporting on Obama’s statement](https://www.pcgamer.com/software/ai/we-cant-stuff-ai-back-in-a-box-says-former-us-president-barack-obama-confirming-hes-not-a-doomer-while-still-calling-for-more-government-ai-regulation/)

9. Reuters, “Meta's Zuckerberg says AI labs have enough incentive to build safely,” September 16, 2026. [Reuters: reporting on Zuckerberg’s response](https://www.reuters.com/business/metas-zuckerberg-says-ai-labs-have-enough-incentive-build-safely-2026-09-16/)

10. Associated Press, “UN chief urges global AI cooperation saying world can't afford ‘a race to the bottom on AI safety’,” September 16, 2026. [Associated Press: reporting on Guterres’s statement](https://apnews.com/article/6ae720a081ce4d5ede35837ca35b8460)

11. The Guardian, “Allowing AI firms to collude to ‘pace the frontier’ is a dangerous proposition,” September 16, 2026. Cited as commentary expressing a competing view, not an adjudicated finding. [The Guardian: commentary on coordination and antitrust](https://www.theguardian.com/technology/2026/sep/16/ai-companies-collude-antitrust-laws)

12. UK Parliament, Hansard, “AI Systems: Risks,” January 8, 2026, remarks by Lord Fairfax of Cameron and Baroness Ritchie of Downpatrick. [UK Parliament, Hansard: AI Systems: Risks](https://hansard.parliament.uk/lords/2026-01-08/debates/5CAD7DC7-3B7E-4925-85C4-88354195031E/AISystemsRisks)

13. Future of Life Institute, “Finances,” 2025 expenditure and funding disclosures, accessed September 16, 2026. [Future of Life Institute: finances and donor disclosures](https://futureoflife.org/about-us/finances/)

14. Tarbell Center, “About,” lifetime supporters and editorial-independence disclosures. [Tarbell Center: supporters and editorial-independence statement](https://www.tarbellcenter.org/about)

15. Tarbell Center, supported-story registry, entries dated September 9 and 15, 2026, and other listed reporting. [Tarbell Center: supported-story registry](https://www.tarbellcenter.org/)

16. Anthropic, Series A announcement, May 28, 2021. [Anthropic: May 2021 Series A announcement](https://www.anthropic.com/news/anthropic-raises-124-million-to-build-more-reliable-general-ai-systems)

17. Survival and Flourishing Fund, published 2025 recommendations and matching-pledge legend, accessed September 16, 2026. [Survival and Flourishing Fund: recommendations and matching pledges](https://survivalandflourishing.fund/)

18. Kevin Bass, metr-money-figure, NOTES.md, coverage notes dated September 14, 2026. This is the investigator's account of his source audit; the article does not claim to have independently re-audited every underlying filing. [Kevin Bass: metr-money-figure / NOTES.md](https://github.com/kevinnbass/metr-money-figure/blob/master/NOTES.md)

19. METR, “Funding update,” August 14, 2026. [METR: August 2026 funding update](https://evals.alignment.org/blog/2026-08-14-funding-update/)

20. METR, Conflict of Interest Policy, version 1.0, August 28, 2026, especially pages 1 and 7. [METR: conflict-of-interest policy, PDF](https://metr.org/coi-policy.pdf)

21. Anthropic, “An alignment assessment of recent cybersecurity incidents,” September 9, 2026. [Anthropic: cybersecurity incident assessment and METR agreement](https://www.anthropic.com/research/alignment-assessment-cybersecurity-incidents)

22. METR, “Measuring AI Ability to Complete Long Software Tasks,” March 19, 2025. [METR: long-software-task measurement methodology](https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/)

23. Demis Hassabis, “A Framework for Frontier AI and the Dawning of a New Age,” July 14, 2026. [Demis Hassabis: frontier-AI governance framework](https://demishassabis.substack.com/p/a-framework-for-frontier-ai-and-the-dawning-of-a-new-age)

24. Federal Trade Commission, “Behind the FTC's 6(b) Report on Large AI Partnerships & Investments,” January 2025. Staff study, not a finding that the arrangements are unlawful. [FTC: study of large AI partnerships and investments](https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2025/01/behind-ftcs-6b-report-large-ai-partnerships-investments)

25. Anthropic, Microsoft–NVIDIA partnerships announcement, November 18, 2025. [Anthropic: Microsoft and NVIDIA partnerships](https://www.anthropic.com/news/microsoft-nvidia-anthropic-announce-strategic-partnerships)

26. OpenAI, Amazon strategic partnership, February 27, 2026. [OpenAI: Amazon strategic partnership](https://openai.com/index/amazon-partnership/)

27. Anthropic, Amazon compute expansion, April 20, 2026, updated April 21. [Anthropic: Amazon compute expansion](https://www.anthropic.com/news/anthropic-amazon-compute)

28. CoreWeave SEC Form 8-K, agreement dated September 9, filing signed September 15, 2025. [SEC: CoreWeave Form 8-K on NVIDIA capacity agreement](https://www.sec.gov/Archives/edgar/data/1769628/000176962825000047/crwv-20250909.htm)

29. Anthropic, “Higher usage limits and a SpaceX compute deal.” [Anthropic: SpaceX compute agreement](https://www.anthropic.com/news/higher-limits-spacex)