In 2022, two of Anthropic's co-founders, Daniela and Dario Amodei described the company as a small group leaving OpenAI in late 2020 and early 2021 during the pandemic, meeting masked in backyards, because they wanted to make what they called a "focused research bet" with a tightly aligned team. The bet was on building AI systems that are "helpful, honest, and harmless". On race dynamics, Dario was explicit that Anthropic should build models close to the state of the art (because you need frontier models to study frontier safety problems), but that they shouldn't race ahead or build models bigger than other orgs, and shouldn't ramp up excitement or hype about giant models. Fast forward to today: Anthropic is a frontier for-profit company valued at $380 billion as of Feb 2026, is pursuing agentic coding on a technical trajectory that approaches autonomous recursive self-improvement, Dario has among the shortest AGI/TAI timelines of any major AI lab CEOs, their recently revised RSP has drawn criticism for dropping what many in the community interpreted as a commitment to pause development under certain risk thresholds, and now, what this piece is primarily about, the "~oversight" organisations it once called for will apparently soon be funded by Anthropic's own people. The founding rhetoric of humility hasn't quite changed, but the structural conditions around it have changed enormously.
Anthropic's own stated reason for the RSP revision, that unilateral restraint in a competitive environment cedes ground to less responsible actors, is basically a textbook articulation of the competitive trap Scott Alexander described in his Moloch essay. Since the RSP revision has been discussed in depth by others, I won't go into it here. I want to focus on something else: what happens when the philanthropy of an AI company's workforce becomes the primary funding source for the organisations meant to oversee that company.
A week ago, I came across the Transformer article on Anthropic's approaching philanthropic windfall. The tl;dr of the article by Gemini: Anthropic's co-founders have pledged 80% of their wealth to charity. A $6B tender offer and upcoming IPO are transitioning paper wealth into liquid capital. organisations like Coefficient Giving are positioned to manage massive inflows focused on AI safety, animal welfare, and global health.
A note on framing:
What I'm describing is akin to (or straight up is) regulatory capture operating through philanthropic infrastructure with some Goodhartian-esque dynamics layered on top. I'm using the Moloch frame because this audience knows it and because the "emergent, nobody-chose-this" quality of the outcomes matters. But the canonical Moloch describes competitive multipolar traps, and what I'm describing involves a cooperative community drifting into a structural trap. That's a different category of Moloch I suppose. I'll use Moloch where the competitive angle genuinely applies (Anthropic does exist in a competitive industry that shapes everything downstream) and lean on regulatory capture where it doesn't apply.
The causal chain
Here's the skeleton of my argument. Decide for yourself whether each link holds.
EA's cause prioritization framework identified AI safety as a top cause area. This attracted EA-aligned people into AI companies. Some of those companies, like Anthropic and to a good extent even OAI, became enormously valuable. This made EA-aligned AI employees wealthy. Their wealth is now flowing back into EA-aligned organisations through the donation infrastructure EA built. Those organisations include the ones evaluating and overseeing the AI companies that made the donors wealthy. The EA framework can ask "is this organisation doing cost-effective work?" It may struggle to ask "has my definition of cost-effective work been shaped by the financial interests of my donor base?" because that's a question about the framework itself, not one the framework is designed to answer.
No one designed this loop. But (and this is where the "nobody chose this" framing needs some qualification) donors do choose where to give, org leaders do choose what to research, grantmakers do choose whom to fund. These are legible decisions made by people with shared backgrounds, networks, and financial interests. I think calling the outcome "emergent" captures something real about how individual rationality aggregates into structural dysfunction, but it shouldn't mystify a process that sociologists and political scientists have written about. The weakest link in this chain of mine then, I suppose, is the leap from "aggregated individual preferences" to "structural capture." A natural objection is: isn't this just a community funding what it believes in? When does funding-what-you-believe-in become “captured”? I believe that it gets captured at the point where the funding relationship compromises the recipient's capacity to produce findings that threaten the funder's interests regardless of whether that capacity is currently being exercised. A pharma company that funds the research institute evaluating its drugs has a structural problem even if the drugs happen to be good and the researchers happen to be honest. The problem is that the funding dependency erodes the capacity to say "this drug doesn't work," a capacity whose value is irrelevant to case-by-case needs. Similarly, if AIS orgs are financially dependent on Anthropic-sourced wealth, their capacity to say "Anthropic should be constrained" is structurally eroded even if Anthropic is currently behaving well. Thus, this is a concern about independence and institutional design, not about any individual's "goodness", beliefs, or integrity.
Selection mechanisms
Organisations that produce work compatible with Anthropic's continued operation (evaluations that feed into the RSP, alignment research that improves Anthropic's products, governance frameworks that legitimize Ant's approach) are positioned to receive that funding. organisations whose work implies Anthropic should slow down, stop, or submit to external authority (PauseAI, certain configurations of MIRI, maybe the likes of FLI) are unlikely to receive much. Nobody needs to make a phone call. The selection happens through aggregated donor preferences.
You could ask: how do you know Anthropic employees will form a coherent donor bloc? Ant employs thousands of people with diverse views. Pretty sure some are internally critical of the RSP revision. Some may fund PauseAI-adjacent work. Pharma company employees donate to pharma-critical organisations all the time! Fair point. The selection mechanism I'm describing requires sufficient convergence, not unanimity. My claim is that the professional context (shared information environment, shared social network, shared financial interest in Anthropic's success, shared advisory infrastructure) will produce enough convergence in aggregate to meaningfully shape the funding landscape. Whether it actually does is testable, and I'll offer some predictions.
Counterforces
Before going further, I'll give some weight to what pushes against this concern.
Government funding for AI safety is growing: UK AISI, US AISI, EU AI Act enforcement bodies. Non-EA philanthropists like the Templeton Foundation, Hewlett Foundation, MacArthur Foundation, Alfred P. Sloan Foundation, etc. have begun funding AI safety and governance work. If these alternative sources scale fast enough, the funding concentration I'm worried about may not materialize, or may be substantially diluted. This is a counterforce that could dissolve the strongest version of my concern/argument.
Also, I don't know well enough what exact institutional safeguards currently exist within the EA funding infrastructure. Does CG have relevant COI policies? Funding diversification requirements? If robust safeguards exist, the mechanism by which funding concentration translates to epistemic capture may be significantly blocked. If anyone closer to this infrastructure can speak to what safeguards exist (or don't), I'd like to know! I'm raising the structural concern because I think it's important enough to warrant community attention. This is more "serious concern/ques that deserves investigation" than "diagnosis of existing capture."
Value drift and (perverse?) incentive structure
An employee who joins Anthropic and believes in AI safety will, over time, likely have their understanding of "AI safety" shaped by Anthropic's institutional perspective, through immersion rather than any deliberate indoctrination. Colleagues, friends, competitors, and funded organisations all reflect the same perspective back. From the inside, this feels like learning. From the outside, it could be environmentally induced narrowing. I'd wager that these are pretty hard to distinguish.
All this then interacts with the donation infrastructure. The more faithfully an employee follows EA-recommended giving pathways, the more their giving may flow toward organisations whose work is compatible with their employer's interests because the advisory infrastructure has likely been shaped by people who share their professional context. And soon, this infrastructure's effectiveness at concentrating resources, which is normally its greatest virtue, becomes, in this specific case, a mechanism for concentrating oversight capacity in industry-compatible directions. When pharmaceutical companies fund continuing medical education, the content systematically skews toward their products because the funding relationship shapes what gets produced, and not necessarily because anyone is lying. When tobacco companies funded research, the funded research systematically delayed consensus on harm through emphasis and framing, not just outright falsification. The mechanism is funding dependency which is shaping outputs over time.
Where EA can't measure its own success
EA's optimisation function works brilliantly when outcomes are measurable: malaria nets distributed, lives saved, cost-per-outcome calculated. The optimisation has an… "external referent", it can be wrong and correct itself.
In AI x-risk (forget s-risks), the external referent is weaker. Of course there are near-term measurables (rates of deceptive outputs, robustness failures, autonomous action incidents, stuff along those lines), but the core concern of preventing catastrophic outcomes from transformative AI lacks tight feedback loops. Why? Well because the catastrophic outcomes we're trying to prevent haven't happened yet. The only available proxy for "good AIS x-risk work" is the judgment of people the community considers expert, in practice, people embedded in AI companies and EA-aligned research organisations. When the measure of "high-impact AI safety work" becomes the target for funding-seekers, it ceases to be a good measure. Goodhart's Law operating at the level of a cause area.
All this can be testable
I'm not claiming that Anthropic employees are hypocrites, that EA is a front, or that philanthropy is insincere. I'm claiming that sincerity is not a firewall against structural capture. Came up with some directional hypotheses, not clean tests below. They're better than nothing, and I'm putting them down so we can check:
If the structural concern is real, within five years we should observe: (1) Anthropic equity-derived funding grows to represent a dominant share (I'd say >40%) of AI safety nonprofit funding. I lack current baseline data and would welcome corrections, knowing the current funding landscape precisely would itself be useful. (2) organisations whose published work has recommended binding constraints on frontier AI development (not just evaluations that feed into voluntary frameworks) experience below-average funding growth relative to the overall AIS funding pool. (3) METR's budget grows more slowly than Anthropic-aligned evaluation organisations, unless it secures substantial government or non-industry funding. But METR could thrive or struggle for idiosyncratic reasons unrelated to the dynamics I'm describing. Any single prediction isn't decisive but patterns across multiple indicators would be.
Conversely, if five years from now AI safety funding is genuinely diverse with non-industry sources comprising more than 50% of total AIS funding and organisations are producing findings that result in binding constraints on Ant's operations, I'm wrong. I hope I am wrong and all that'll happen is more vaccines, better QALYs, and fewer farmed animals.
To me, in some ways, EA often presents itself as Scott’s "Elua", the rational coordination mechanism that channels resources toward what actually matters. For global health, animal suffering, and CBRN/GCRs, EA credibly is, IMO. In AI safety, the entanglement between the optimiser and the optimised-for creates structural tension. EA-aligned people build the AI, EA-aligned organisations write the evaluations, and EA-aligned donors fund the evaluators. But METR's independence and this very discussion are evidence the system isn't fully closed. At least not yet. Still, the entanglement at multiple levels of the feedback loop is concerning, and the independent orgs may face resource disadvantages.
The question we need to sit with is not "are we individually acting in good faith?" - of course we are - but "does our collective structure have the capacity to produce findings and fund organisations that could materially threaten the financial interests of our most powerful members?" If the answer is yes, the structure is healthy. If the answer is "it wouldn't need to, because our most powerful members are aligned with good outcomes", that's the answer a captured system would produce.
If I'm wrong, I've wasted your time with an overwrought analogy. If I'm even directionally right, the community best equipped to diagnose structural traps may be sitting inside one. Given the stakes, the second risk is worth the cost of the first.