why I distrust Anthropic

01

safety is a claim. power is a fact.

I distrust Anthropic because its leaders ask the public to accept a dangerous bargain: help us build increasingly powerful AI, trust us to judge the risks, and give serious weight to our proposals for governing everyone else.

The people making that request lead a company with investors, customers, competitors, and enormous spending commitments. Those interests do not disappear when the company describes its mission as serving humanity.

My case is about what Anthropic has done and what its incentives reward. Private motives cannot be read from a funding announcement. But the public does not need proof of bad intentions before demanding limits on private power.

The sharpest evidence is that, when its safety commitments became harder to sustain in a competitive race, Anthropic changed them.

02

the promise changed when competition mattered

On February 24, 2026, Anthropic rewrote its Responsible Scaling Policy. Its explanation distinguishes safeguards it can implement alone from stronger measures it wants the whole industry to adopt. It also describes its new safety-roadmap goals as nonbinding. Anthropic’s explanation.

The policy itself explains the incentive. If one developer pauses while others continue, Anthropic argues, less responsible companies could gain ground and make the world less safe. Its commitment to delay development under the specified conditions depends in part on whether it believes it has a significant lead. RSP version 3.0.

That argument may be sincere. It is also convenient for a company that needs to remain competitive: continuing to build becomes part of the safety case.

The revision did add reporting and external-review provisions. It was not an abandonment of every safeguard. But it exposed the weakness of voluntary restraint. A promise that changes when rivals advance cannot carry the same weight as a rule enforced by someone outside the race.

Competition was always going to be the test. The test was whether the commitment would still constrain the company when keeping it became expensive.

03

follow the money without inventing a conspiracy

Anthropic’s funding history shows how a safety research company became tied to some of the largest financial and technology interests in the world.

Sources: 2021 announcement, 2022 announcement, FTX investment and proposed sale, Amazon agreement, Series H announcement and participant list.

The Series H list also includes Blackstone, Brookfield, D. E. Shaw Ventures, DST Global, Fidelity, General Catalyst, Insight Partners, Jane Street, Lightspeed, MGX, and Temasek, among others. The announcement includes previously committed investments, so adding every headline round to every separate investor announcement would overcount the money.

The FTX connection deserves scrutiny. Bankman-Fried was sentenced to 25 years for fraud. But receiving an investment does not establish that Anthropic’s leaders knew about or participated in his crimes. That allegation requires evidence of its own. Justice Department sentencing announcement.

The lesson is narrower and still damaging to the moral sales pitch: money presented alongside a mission to help humanity is not evidence that the people supplying it deserve trust.

04

the investors also own the infrastructure

Amazon’s arrangement makes the conflict unusually easy to see. It invests billions in Anthropic while Anthropic commits to buying more than $100 billion of its services. Google is another investor and infrastructure partner. Anthropic announced a further expansion with Google and Broadcom in April 2026. Google and Broadcom agreement.

The FTC examined the Amazon–Anthropic and Google–Anthropic relationships alongside Microsoft–OpenAI. Its January 2025 report identified spending commitments back to cloud investors, potential barriers to changing suppliers, and access to sensitive business information. These were competition concerns, not a finding that every partnership was unlawful. FTC report summary.

My concern is dependence. If the companies financing the lab also supply the machinery and distribute its products, their interests have several routes into its decisions. A public-benefit label does not remove that pressure.

05

the leadership network needs visible boundaries

The early backers overlap with the AI-safety and effective-altruism world. Effective altruism seeks to direct resources toward doing the most good. That goal does not settle who should have authority, how conflicts should be managed, or whose judgment can be challenged.

Holden Karnofsky, a co-founder of Open Philanthropy and husband of Anthropic president Daniela Amodei, joined Anthropic in January 2025 to work on safety issues, including its scaling policy. Fortune reported that he reports to chief science officer Jared Kaplan. The family relationship is relevant to oversight. It is not evidence that he lacks expertise. Fortune reporting.

The Wall Street Journal reported in August 2026 that Dario Amodei’s wife, Cami Clark, advises him informally and helped bring in Eric Schmidt as an early investor. It also reported that her proposed investment fund, which would have formalized a financial role, did not proceed after opposition from Daniela and other co-founders. That rejection is evidence of an internal boundary and belongs in the account. WSJ investigation, archived.

The same investigation described Clark’s unsuccessful pitch to Jeffrey Epstein for an earlier business. That is not an Epstein investment in Anthropic. Nor does Claude giving a poor answer about a marriage prove that the company deliberately censored it.

Schmidt’s public-policy role raises a more direct conflict question. CNBC documented his AI investments while he chaired the National Security Commission on Artificial Intelligence. Ethics experts criticized the overlap. The report also said there was no indication he had broken ethics rules or acted unlawfully. People close to him said the investments were disclosed confidentially to the government. CNBC investigation.

The public deserves clear disclosure of financial interests, adviser roles, and recusals. Access through family, philanthropy, or government service should not be treated as a substitute for independent oversight.

06

political influence works across party lines

By July 21, 2026, Anthropic had announced $40 million for Public First Action. It says the money is restricted to public education and policy work and cannot be used to influence candidate elections. That restriction matters. It would be inaccurate to describe the full amount as campaign contributions. Anthropic’s disclosure.

Public First Action is led by former lawmakers Brad Carson and Chris Stewart. CNBC reported that the group launched advertising supporting Republican senators Marsha Blackburn and Pete Ricketts. Separately, The Hill reported a $1 million contribution from Dario Amodei to the associated super PAC. Corporate policy spending, personal campaign spending, and the activities of related groups should be tracked separately. CNBC, The Hill.

In California, Anthropic endorsed Scott Wiener’s SB 53. Wiener described the company as constructive in the legislative process. That establishes cooperation on policy, not that he is controlled by Anthropic or endorses everything it does. Interview with Wiener.

Some of these policies may be good. Independent testing, whistleblower protection, and public disclosure can protect people. The conflict arises because Anthropic is both a company subject to the rules and a well-funded participant in shaping them.

Regulation can also favor large incumbents if compliance is costly or access depends on infrastructure they already control. That is a risk to examine in the actual rules, not proof that every safety proposal is a secret attempt to eliminate competitors.

07

look at conduct when rights cost money

In the June 2025 Bartz ruling, Judge William Alsup found that Anthropic had downloaded millions of pirated books for a permanent library. He rejected fair use as a defense for that acquisition and retention. The same ruling found the training use at issue transformative and fair, as well as certain scanning of purchased books. Those are distinct findings. Court order.

Anthropic subsequently agreed to a $1.5 billion settlement, which received preliminary approval in September 2025. That is an agreement to resolve claims, not a criminal conviction. AP reporting.

This matters because it is conduct, not branding. Other people’s rights stood between the company and something it wanted. Its response gives us a concrete reason to question whether its declared concern for humanity reliably protects individual humans.

08

humanity includes people outside America

In November 2024, Anthropic, Palantir, and AWS announced access to Claude for U.S. intelligence and defense agencies, including classified environments. Palantir is a partner in this account, not an investor established by that announcement. Partnership announcement.

There is significant evidence in Anthropic’s favor here. In February 2026, it publicly resisted demands to remove restrictions on mass domestic surveillance of Americans and fully autonomous weapons. That stance cannot fairly be erased from a critical account. Amodei’s statement.

But the statement also supports lawful foreign intelligence work. Its weapons objection concerns present reliability and missing safeguards. It is not a permanent rejection of autonomous weapons. Protecting American national interests and protecting every person affected by American power are different obligations.

For people elsewhere, the question is practical: who can challenge a harmful use of these systems, inspect what happened, and obtain a remedy?

09

a public-benefit company is not public control

Anthropic’s Long-Term Benefit Trust gives financially disinterested trustees rights to appoint and remove directors, with power designed to increase over time. This is a real governance mechanism. The company also describes provisions allowing sufficiently large stockholder supermajorities to change the trust without trustee consent. Anthropic’s governance explanation.

The same explanation says the board selected the original trustees and future trustees are selected by trustees. It identifies later appointments including national-security policy leader Richard Fontaine, Carnegie president Tino Cuéllar, and former Federal Reserve chair Ben Bernanke.

Expertise may improve judgment. It does not give the affected public a vote. The trust represents an attempt to act for humanity. Humanity did not choose it.

10

why this can become a worldwide risk

The danger is the combination: a company building powerful systems, dependence on enormous commercial partners, influence over the rules, and a leadership team able to revise its own safety commitments.

If such systems become central to work, information, and state decisions, mistakes and self-serving choices can affect people who never agreed to use them. The company's own catastrophic-risk claims make independent constraints more urgent, not less.

I cannot establish that Anthropic’s leaders secretly want to harm the world. I can identify a structure in which they can sincerely believe that helping humanity requires more resources and more discretion for their own company.

That is enough reason to distrust their assurances as a substitute for oversight. It is not a reason to give another AI company a free pass.

What would change my view? Independent access to meaningful safety evidence. Enforceable conditions for stopping dangerous deployments. Clear conflict disclosures. Protection for people who report failures. And ways for affected people to challenge decisions and obtain remedies.

Until then, I will judge Anthropic by the limits it accepts when those limits cost it money, speed, or power.

11

trails from here

• Tribe roles: how groups allocate authority and decide who can speak for them.
• Consequential tests: why a commitment matters when it can change the next decision.