When Anthropic appointed Ben Bernanke to its Long-Term Benefit Trust, I asked ChatGPT which economists most closely matched the worldview of each major AI lab. After the economists, I asked the obvious follow-up: how about philosophers?

The Economists

The matches are less about economic policy than about each lab’s theory of innovation, risk, knowledge, and institutional power.

Lab	Closest economist or school	Why
Anthropic	Frank Knight, with Arthur Pigou	Knight distinguished measurable risk from genuine uncertainty. Anthropic’s worldview begins from the idea that frontier AI contains unknowns that cannot be modeled cleanly. Pigou supplies the case for constraining private activity when its costs can spill over onto society. Source
OpenAI	Paul Romer, with Joseph Schumpeter	Romer’s growth theory treats ideas as non-rival engines of increasing returns and long-run prosperity. That matches OpenAI’s promise that intelligence can expand human capability and abundance. Schumpeter captures the concentration of capital and creative destruction involved in getting there. Source
xAI	Joseph Schumpeter	Schumpeter’s entrepreneur overturns incumbent structures through ambition, capital, and creative destruction. xAI’s speed, appetite for scale, and hostility toward established gatekeepers fit that heroic theory of innovation better than a conventional theory of the firm. Source
Meta	Friedrich Hayek	Hayek treated useful knowledge as dispersed across society rather than concentrated in a central authority. Meta’s argument for open models is similar: distribute the technology and let developers discover uses no single lab could plan. The tension is that this Hayekian ecosystem is sponsored by one of the world’s largest centralized platforms. Source
Google DeepMind	Kenneth Arrow	Arrow’s work joined mathematical formalism, information, uncertainty, and collective choice. DeepMind has a similar faith that difficult social and scientific problems can be made tractable through formal intelligence, while also recognizing that decisions about collective benefit cannot be reduced to a single uncomplicated objective. Source
Thinking Machines	Herbert Simon	Simon studied bounded rationality: real intelligence operates with limited information inside organizations and interfaces, not in a frictionless world of perfect optimization. Thinking Machines’ emphasis on collaboration and customization similarly treats AI as part of a human decision system rather than an omniscient substitute for one. Source
Z.ai	Alexander Gerschenkron	This is the most speculative match. Gerschenkron argued that late-developing economies do not simply repeat the path of earlier leaders; they use different institutions, concentrated investment, and technological leapfrogging. Z.ai’s emphasis on building broad practical capability within China’s fast-moving AI ecosystem has a similar catch-up logic. Source
Safe Superintelligence	Martin Weitzman	Weitzman’s work on fat-tailed climate risk argued that a small probability of irreversible catastrophe can dominate ordinary cost-benefit calculations. SSI is organized around the same economic intuition: when the downside is existential, safety cannot be treated as one product consideration among many. Source
Mistral	Friedrich List	List cared less about abstract free exchange than about a nation’s productive capacity and freedom from strategic dependence. Mistral’s case for European AI sovereignty is similarly about retaining the infrastructure, skills, and firms required to avoid dependence on foreign gatekeepers. Source
DeepSeek	Harvey Leibenstein and X-efficiency	Leibenstein asked why organizations with similar inputs can produce very different outputs, emphasizing discipline, incentives, and internal efficiency. DeepSeek’s identity is built around that gap: co-designing architecture, training, and inference to obtain more capability from constrained resources. Source
Cohere	Ronald Coase	Coase explained firms in terms of the cost of coordinating activity across markets and organizational boundaries. Cohere’s enterprise philosophy is similarly institutional: a model becomes valuable when it works inside a firm’s data, permissions, vocabulary, and processes, reducing the cost of coordination. Source
Perplexity	George Stigler and the economics of information	Stigler treated search as an economic activity with real costs: people stop looking when the expected value of another piece of information falls below the effort required to find it. Perplexity’s product is almost a direct response, lowering the cost of locating, comparing, and verifying knowledge. Source

One could argue that Meta’s classification rests on an opportunistic framing: its open models may have been a step toward developing competitive models that will be less open. There are plausible alternatives for every lab, and repeated names may indicate a broad trend in AI rather than something specific to a company. Still, most of the suggestions seem broadly reasonable.

How About Philosophers?

Philosophers are another useful lens because many labs were founded to answer questions like: What is intelligence? How should powerful intelligence be governed? How should we approach risk when developing powerful new technologies?

Lab	Closest philosopher or school	Why
Anthropic	Immanuel Kant, with Hans Jonas	Kant fits because the model is expected to internalize an explicit constitution and behave according to general principles, not merely maximize outcomes. Jonas adds the idea that unprecedented technological power imposes unprecedented duties toward humanity’s future. Anthropic’s Constitution and risk-scaled governance make this one of the clearer matches. Source
OpenAI	Condorcet, with Saint-Simon	Condorcet represents Enlightenment confidence in the indefinite expansion of human knowledge, capability, and prosperity. Institutionally, though, OpenAI is somewhat Saint-Simonian: a technically competent elite assembling capital, science, and infrastructure to transform civilization for universal benefit. Its current framing emphasizes expanded human capability, prosperity, choice, and continued human control. Source
xAI	Nietzsche in temperament; Karl Popper in epistemology	The Nietzschean element is heroic ambition, self-overcoming, and hostility to inherited authority. But xAI’s stated pursuit of understanding the universe is more Popperian: bold conjecture, criticism, and discovery. Nietzsche himself was too suspicious of uncomplicated claims about objective truth to be a perfect fit. Source
Meta	John Stuart Mill, with Popper	Meta’s philosophical self-image is pluralistic: distribute the technology, let many developers experiment, and avoid one institution determining permissible uses. That resembles Mill’s “experiments in living” and Popper’s open society. The obvious contradiction is that this decentralizing philosophy is advanced by one of history’s largest centralized platforms. Source
Google DeepMind	Gottfried Wilhelm Leibniz, with Plato	Leibniz dreamed of formalizing knowledge into a universal rational system capable of resolving difficult questions. That fits DeepMind’s combination of mathematics, science, general intelligence, and large-scale problem-solving. Its confidence in responsible expert institutions adds a mildly Platonic, philosopher-governor streak. Source
Thinking Machines	John Dewey and pragmatism	For Dewey, intelligence is not detached contemplation; it is inquiry conducted through interaction with an environment. Thinking Machines’ emphasis on collaboration, customization, and interfaces that adapt to human behavior is deeply pragmatic. AI becomes a partner and instrument within human activity rather than an autonomous oracle standing above it. Source
Z.ai	Mozi and Mohism	This is more speculative, but Mohism combines practical engineering, meritocratic competence, and a consequentialist concern with producing broad, impartial benefit. Z.ai’s public philosophy similarly emphasizes safe, beneficial AGI directed toward difficult practical problems rather than a highly individualistic theory of self-expression. Source
Safe Superintelligence	Nick Bostrom, with Hans Jonas	Bostrom provides the intellectual premise: superintelligence could become the decisive event in human history, making alignment an existential priority. Jonas provides the ethics: when technology can permanently foreclose humanity’s future, precaution becomes a fundamental duty. SSI’s entire organization is built to keep safety ahead of capabilities and resist commercial distraction. Source
Mistral	Philip Pettit and republican non-domination	Mistral’s openness is not merely “information wants to be free.” It is about ensuring that companies, governments, and societies are not dependent on a handful of foreign AI gatekeepers. That fits the republican conception of freedom as non-domination: you are not genuinely free when another actor can arbitrarily control your access to essential infrastructure. Source
DeepSeek	Gilbert Simondon	Simondon understood technological progress as the increasing internal coherence and “concretization” of technical systems: components become more integrated, efficient, and mutually reinforcing. DeepSeek’s philosophical identity is less political than engineering-oriented—co-designing architecture, training, and inference to obtain greater capability from constrained resources. Source
Cohere	The later Wittgenstein	Wittgenstein’s central insight was that meaning comes from use within particular “language games” and forms of life. Cohere’s enterprise philosophy is similar: language models become valuable when embedded in the vocabulary, data, permissions, and working practices of a specific institution—not when treated as context-free universal minds. Source
Perplexity	C. S. Peirce, with Diderot	Peirce saw knowledge as an open-ended, fallible process of inquiry in which claims are corrected through evidence and a community of investigators. That fits Perplexity’s citation-and-verification ideal. Diderot is the alternative: organize the world’s scattered knowledge into something immediately navigable and useful. Source

Stated Goals and Revealed Behavior

Would other LLMs agree? I shared the response with Claude, which largely agreed, with one caveat: most of these matches came from what the companies say about themselves, not what they do. A mission statement describes how a company wants to be understood. Pricing, capital structure, lobbying, hiring, and product decisions describe how it actually operates. The two don’t always point in the same direction.

That was a valid criticism, so I asked Claude to run the labs back through that filter. Labs omitted below kept their original matches, including Anthropic. That result deserves extra skepticism because Claude is Anthropic’s model. Still, the Kant comparison is unusually testable: Claude’s Constitution is public, versioned, and presented as part of the training method rather than merely a mission statement.

These are the changes Claude suggested when the labs were judged against observed behavior rather than their stated missions:

Lab	Lens	Revised view
OpenAI	Economist	Joseph Schumpeter, with an industrial-organization lens. Romer’s theory of non-rival ideas and increasing returns is close to the abundance story told by almost every AI lab. OpenAI’s corporate structure, concentration of capital, and scale of its compute commitments are more distinctive. Schumpeter’s combination of creative destruction and concentrated investment therefore explains more than Romer’s general theory of technological growth.
OpenAI	Philosopher	Saint-Simon. Condorcet’s confidence in expanding knowledge and prosperity is another industry-wide aspiration. Saint-Simon is more specific: an expert-led institution assembling science, capital, and infrastructure in order to transform society. That describes OpenAI’s organizational form, not merely its promised outcome.
xAI	Philosopher	No confident match. Nietzsche and Popper describe Elon Musk’s public persona more clearly than they explain a distinct xAI mechanism. Meanwhile, xAI has presented Grok as a maximally truth-seeking alternative, but that has not yet emerged as a clear differentiator from other labs. Its own system card describes supervised fine-tuning, reinforcement learning from human and synthetic reward signals, refusal policies, and safety evaluations. In practice, it has to make the same kinds of tradeoffs among accuracy, usefulness, safety, and operator preferences.
Google DeepMind	Economist	Kenneth Arrow, but for a different reason. The original appeal to formalism, information, and uncertainty was broad enough to fit almost any research lab. The sharper analogy is Arrow’s impossibility theorem: alignment requires heterogeneous and often conflicting human preferences to be compressed into a collective objective, without a neutral rule that makes the conflicts disappear. DeepMind’s research on preference disagreement makes this a structural match rather than an aesthetic one.
Z.ai	Economist	No confident match. Gerschenkron’s catch-up theory requires more than operating in a later-developing technology ecosystem; it requires evidence of the specific mechanism: different institutions, concentrated investment, and deliberate leapfrogging. There isn’t enough public information about Z.ai’s actual decisions to establish that mechanism rather than merely its public positioning.
Z.ai	Philosopher	No confident match. Mozi risks being a geography-driven selection rather than an inference from the company’s decisions. Practical engineering, meritocracy, and broad social benefit are all plausible themes, but the available evidence does not establish a distinctively Mohist institutional philosophy.
Safe Superintelligence	Philosopher	Hans Jonas. Bostrom is accurate, but almost too direct: existential risk is already SSI’s stated premise. Jonas provides the more useful outside framework. Organizing the company around safety ahead of commercialization enacts his argument that technologies capable of irreversible harm create a special duty of precaution toward the future.

The most persuasive comparisons are the most specific. Stigler’s search costs describe Perplexity’s product. Kant maps onto Anthropic’s training method. Wittgenstein maps onto Cohere’s enterprise strategy. DeepSeek is especially interesting because Leibenstein and Simondon arrive at the same efficiency-under-constraint story from different directions.

Schumpeter, Popper, and Jonas recur because creative destruction, open inquiry, and catastrophic risk are concerns shared across the industry. That makes them relevant, but less useful for distinguishing one lab from another. Mission statements are informative, but as AI becomes more involved in high-stakes parts of our lives—medicine, therapy, education, and work—a company’s training methods, incentives, restrictions, and deployments will tell us more about its values than the philosophers it resembles.