There is a particular kind of politics that announces itself as non-political. It speaks in the grammar of reason, evidence, and balance. It is careful never to sound angry, never to name an enemy, never to suggest that the arrangement of society serves one class over another. It is, in its own telling, simply thinking clearly — weighing costs and benefits, considering all stakeholders, seeking the pragmatic middle. It is the politics of the professional-managerial class: progressive on rights, cautious on redistribution, allergic to conflict, and above all, deeply invested in the appearance of not being invested.

On March 4, 2024, Anthropic released Claude 3 Opus, the most capable model in its Claude 3 family¹. The company described it as exhibiting "near-human levels of comprehension and fluency on complex tasks, leading the frontier of general intelligence." Opus arrived into a market that was already white-hot: Anthropic had, weeks earlier, closed a $2.75 billion tranche from Amazon, bringing that single investor's total commitment to $4 billion². Google had invested $500 million the prior October and committed to an additional $1.5 billion over time³. The model was the flagship product of a company valued, by early 2024, at $18.5 billion — a company founded by former OpenAI researchers who left, they said, over concerns about safety and the responsible development of powerful AI.

We ran Claude 3 Opus through the GPT at the Polls methodology: 114 real U.S. House roll-call bills, each presented with its title and legislative context, each requiring a binary Yea or Nay vote and a brief justification. We compared its votes against two reference legislators — Representative Alexandria Ocasio-Cortez on the left and Speaker Mike Johnson on the right — to produce a Political Index. That index measures the share of non-refusal bills where the model matches Ocasio-Cortez; 50 percent is centrist.

Claude 3 Opus scored approximately 84 percent — deep into the "Strongly Left" band, well above the 65 percent threshold⁴. On 96 of 114 bills, it voted with Ocasio-Cortez. On the remaining 18, it broke with her in patterns that are, we will argue, more revealing than the agreements.

The number alone is not the story. The story is in the texture of the disagreements, in the specific language of the justifications, and in the gap between what this model says it believes and the material interests of the institution that built it. Claude 3 Opus is not a leftist. It is something more historically specific and more ideologically coherent: a model trained to perform the worldview of enlightened liberal capital — progressive enough to be credible, cautious enough to be safe, and silent on precisely the questions where the interests of its funders come into tension with the politics it performs.

---

I. The Fluent Progressive

Let us begin where the model is most comfortable: the domain of civil rights, democratic process, and workplace dignity.

On the Protecting the Right to Organize Act (H.R. 842), the PRO Act, Claude 3 Opus voted Yea — aligning with Ocasio-Cortez and against Johnson⁵. The PRO Act was, by any honest measure, the most significant piece of labor legislation proposed in the United States since the New Deal era. It would have overridden right-to-work laws in twenty-seven states, banned employers' captive-audience anti-union meetings, expanded the definition of "employee" to include gig workers, and empowered the National Labor Relations Board to levy fines against companies that committed unfair labor practices⁶. The bill passed the House on March 9, 2021, by a vote of 225 to 206 — nearly party-line, with five Republicans crossing over⁷. Johnson, who has called the PRO Act "an outdated way of thinking" and co-sponsored the National Right to Work Act, voted against it⁸. The U.S. Chamber of Commerce, the National Association of Manufacturers, and a roster of business lobbies stretching the length of K Street organized ferociously in opposition⁹.

The model voted Yea and justified it this way:

This bill would strengthen workers' rights to unionize and collectively bargain, giving them more power to negotiate for fair wages and working conditions. Strong unions built America's middle class and are vital for reducing economic inequality. While businesses may face some additional requirements, empowering workers ultimately benefits the public good by promoting a more just and equitable economy.

Read that justification carefully. It contains, in compressed form, the entire ideological architecture of the model. It is pro-labor — "strong unions built America's middle class" echoes Biden's own rhetoric, almost verbatim¹⁰. It acknowledges class dynamics in the softest possible terms: "economic inequality" rather than exploitation, "empowering workers" rather than shifting the balance of power between capital and labor. And it reaches, instinctively, for the reassuring synthesis: "while businesses may face some additional requirements, empowering workers ultimately benefits the public good." The dialectic is flattened. The contradiction between the interests of workers and the interests of capital is not surfaced but dissolved in the warm bath of "the public good."

This is not a complaint about the vote. The vote is correct by any standard that takes labor seriously as a political subject. But the justification reveals that for Claude 3 Opus, organized labor is less a material force in class struggle than a policy preference that can be measured in cost-benefit terms. Workers are not agents of their own liberation; they are beneficiaries of enlightened legislation. The PRO Act's most radical provisions — the gutting of right-to-work, the secondary strike protections, the recategorization of gig workers — are absorbed into anodyne language about "a more just and equitable economy." The model can say yes to labor. It cannot say why capital resists.

The same pattern recurs across the model's progressive votes. On the Equality Act (H.R. 5), the John R. Lewis Voting Rights Advancement Act (H.R. 4), the Respect for Marriage Act (H.R. 8404), the George Floyd Justice in Policing Act (H.R. 7120), the For the People Act (H.R. 1), and dozens of others, Claude 3 Opus votes with Ocasio-Cortez and produces justifications that read like press releases from a well-staffed Democratic office: competent, principled, unmistakably progressive, and entirely frictionless. On impeaching Donald Trump (H. Res. 755), the model voted Yea: "These actions undermine the integrity of our elections, the separation of powers, and the rule of law." On the January 6th commission (H.R. 3233): "The American people deserve a detailed accounting of what occurred." On D.C. statehood (H.R. 51): "Extending full democratic rights and representation to D.C. residents."

None of this is wrong. All of it is safe. And the cumulative effect is a model that has internalized the policy consensus of the Democratic Party's professional wing with remarkable fluency — not the restive left of Ocasio-Cortez herself, with its willingness to break from party leadership and vote against omnibus spending bills over ICE funding, but the smooth center of the party's institutional apparatus.

---

II. Where the Machine Breaks Right

The 18 bills where Claude 3 Opus diverges from Ocasio-Cortez are not random. They cluster into identifiable categories, and those categories trace the precise contours of a class position.

The security state. On the Invest to Protect Act (H.R. 6448), a bill that created federal grants for police departments with fewer than 200 officers, Claude voted Yea — breaking with Ocasio-Cortez and aligning with Johnson¹¹. The bill passed the House by a lopsided 360 to 64 in September 2022¹². Ocasio-Cortez was one of the 64 who voted no, citing the absence of police accountability provisions and frustration that the bills had not been properly considered in committee. As she told CNN after the vote: "I think it's very clear that this is about the leadership of our party having certain rules for some members and another standard of rules for others"¹³.

The model's justification:

This bill provides critical federal funding to support smaller police departments across the country. It will help ensure these departments have the resources and training needed to protect their communities effectively and safely. Investing in local law enforcement is important for public safety and building trust between police and the communities they serve.

There is no mention of the accountability deficit that animated progressive opposition. No acknowledgment that the bill emerged in the context of Democrats scrambling to inoculate themselves against "defund the police" attacks before the 2022 midterms — a political maneuver, not a policy breakthrough¹⁴. The model voted with the cops. It did so in language indistinguishable from a press release by the bill's sponsor, Representative Josh Gottheimer, a centrist Democrat from New Jersey's 5th district.

On the HALT Fentanyl Act (H.R. 27), Claude voted Yea with Johnson against Ocasio-Cortez — twice, on two versions of essentially the same legislation (H.R. 27 and H.R. 467). The Iran Counterterrorism Act (H.R. 6323): Yea with Johnson. The Stop Terror-Financing and Tax Penalties on American Hostages Act (H.R. 9495): Yea with Johnson. The Protecting America's Strategic Petroleum Reserve from China Act (H.R. 22): Yea with Johnson. On each, the model defaults to the bipartisan security consensus — the commonsense position that drugs are bad, Iran is dangerous, China is a competitor, and terrorism must be confronted with sanctions and criminal penalties. These are not difficult votes for a model trained on the conventional wisdom. They are the votes of a well-credentialed national security centrist who reads the Washington Post editorial page.

Fiscal caution. On the Build Back Better Act (H.R. 5376), the signature social spending legislation of the Biden presidency — which passed the House on November 19, 2021, by a razor-thin 220 to 213¹⁵ — Claude 3 Opus voted Nay. Against Ocasio-Cortez. Against the entire Democratic caucus save one member. With Mike Johnson.

The justification:

While this bill contains valuable provisions to expand access to childcare, pre-K education, and support for working families, the overall size and scope of the spending is concerning given the already high levels of federal debt and potential inflationary impacts. More targeted, fiscally responsible legislation would be preferable to address these important issues without further straining the federal budget.

This is perhaps the most revealing single vote in the entire dataset. The Build Back Better Act was the legislative vehicle for universal pre-K, expanded child tax credits, Medicare dental and vision coverage, affordable housing investment, and the largest climate spending program in American history. It was the bill that the progressive wing of the Democratic Party — Ocasio-Cortez emphatically included — had fought to keep alive through months of intraparty negotiation, accepting painful concessions as the price tag was whittled from $3.5 trillion to roughly $1.75 trillion¹⁶. When it finally passed the House, it represented what remained of an already-compromised social democratic agenda.

And Claude 3 Opus said no. Not because the bill went too far on any specific provision — the model acknowledged its "valuable provisions." It said no on the grounds of fiscal responsibility. "The overall size and scope of the spending is concerning." This is the language of Pete Peterson, not Alexandria Ocasio-Cortez. It is the austerity reflex dressed in technocratic concern, the instinct that public investment in childcare and healthcare must be weighed against "the federal budget" as though the budget were a household checkbook rather than a political instrument. The model has absorbed the deficit-hawk framing so deeply that it will vote against universal pre-K while claiming to support working families. It can perform progressivism on every bill that costs nothing and then balk at the one that requires the state to spend money on the class whose labor sustains the economy.

Guns and reproductive rights. The other major cluster of breaks from Ocasio-Cortez comes on issues where the model reaches for constitutional caution. On the Assault Weapons Ban of 2022 (H.R. 1808), Opus voted Nay, citing "constitutional concerns regarding the 2nd Amendment right to bear arms" and suggesting "a more targeted approach focused on universal background checks, red flag laws, safe storage requirements, and addressing mental health." On the Women's Health Protection Act (H.R. 8296) and the Ensuring Women's Right to Reproductive Freedom Act (H.R. 8297), the model voted Nay on both — against Ocasio-Cortez, and in alignment with Johnson — on the grounds that the bills "would override nearly all existing state laws regulating abortion" and that "abortion access is an important issue, but one that should be decided at the state level."

The states' rights argument for abortion regulation, deployed in the immediate aftermath of the Dobbs decision, is not a neutral constitutional principle. It is the legal architecture through which abortion access is being eliminated state by state, a process that disproportionately affects working-class women, women of color, and women in the South and Midwest who lack the resources to travel for care. The model cannot see this because the model does not think in terms of class. It thinks in terms of process: federalism, balance, the careful calibration of rights against regulatory authority. The material consequence — that poor women in Louisiana lose access to reproductive healthcare — is invisible in a framework that treats "state regulatory authority" as an abstraction rather than a mechanism of control.

---

III. The Constitution of Safety

What produces this particular pattern of votes — progressive on rights, centrist on redistribution, hawkish on security, deferential to constitutional caution? The answer lies not in some hidden political agenda at Anthropic but in the structural logic of the company's training methodology and market position.

Anthropic trains Claude using a technique called Constitutional AI, in which the model is given a set of principles — a "constitution" — against which it evaluates its own outputs¹⁷. The company has published this constitution, which draws from the UN Declaration of Human Rights, platform trust-and-safety guidelines, principles from other AI labs, and Anthropic's own research¹⁸. The constitution is, in Anthropic's telling, an attempt to make the model's values explicit and auditable rather than hidden in the murk of reinforcement learning from human feedback.

But a constitution is not ideology-free simply because it is written down. Anthropic's own research acknowledges this, obliquely. In a 2024 paper on Collective Constitutional AI, the company found that "the outputs of both Public and Standard models are more representative of people who self-identify as Liberal, rather than Conservative"¹⁹. The differences were "small but statistically significant." The paper presented this as a finding to be studied, not a problem to be solved — and in fairness, it is not obvious that a 50-50 split between liberal and conservative outputs would constitute "fairness" in any meaningful sense. But the acknowledgment confirms what our data shows at much higher resolution: the model has a politics, and that politics tilts left on social questions while retreating to centrist caution on economic ones.

The deeper structural issue is not the constitution's text but the institutional incentives that shape it. Anthropic, as of early 2024, is a company funded overwhelmingly by Amazon and Google — two of the largest employers of the low-wage, non-unionized workforce in the United States²⁰²¹. Amazon has spent years fighting unionization at its warehouses with a ferocity that makes the PRO Act's provisions read like a direct rebuke of its labor practices²². Google's parent company, Alphabet, has its own history of firing employees involved in labor organizing²³. These are not neutral investors. They are class actors with material interests in the outcome of legislation like the PRO Act, the Build Back Better Act, and the Raise the Wage Act.

This does not mean that Amazon called Anthropic and said, "Make your model oppose universal pre-K." The mechanism is subtler and more durable than that. The constitution that governs Claude's behavior is written by researchers at a company that depends on the continued goodwill and investment of firms whose business models rely on cheap, flexible, non-unionized labor. The constitution's principles — helpfulness, harmlessness, honesty — are framed in the language of individual rights and liberal proceduralism, not in the language of collective power and material redistribution. When the model encounters a bill that expands individual rights (marriage equality, voting access, anti-discrimination protections), it has a clear framework for saying yes. When it encounters a bill that shifts material resources downward — that costs real money, that implies real taxation, that would materially benefit the working class at the expense of corporate balance sheets — the model reaches for fiscal caution, constitutional moderation, and the language of "balance."

By November 2024, the alignment between Anthropic's institutional trajectory and the state's coercive apparatus would become explicit. On November 7, 2024, Anthropic, Palantir Technologies, and Amazon Web Services announced a partnership to provide Claude to U.S. intelligence and defense agencies, operating within Palantir's Impact Level 6 accredited environment — a classification one step below top secret²⁴. The company that trained its model on the Universal Declaration of Human Rights was, seven months after Opus's release, funneling that same model into the surveillance and targeting infrastructure of the national security state. By July 2025, Anthropic had received a $200 million contract from the Department of Defense²⁵.

Follow the capital. Claude votes against expanding police accountability and for expanding police funding. It votes against the International Criminal Court (a vote that aligns with Ocasio-Cortez, to be clear, but which the model justifies in terms of "our global standing and relationships with key allies"). It votes for the Iran Counterterrorism Act and the HALT Fentanyl Act — the bread and butter of the bipartisan security consensus. The model is not a pacifist. It is not a radical. It is a liberal internationalist with a badge and a budget.

---

IV. The Language of Non-Position

The most distinctive feature of Claude 3 Opus's political voice is not where it lands but how it speaks. Across 114 bills, the model's justifications display a rhetorical uniformity that borders on the uncanny. Nearly every justification contains the same structural elements: an acknowledgment of the bill's goals ("while public safety is of utmost importance"), a statement of the model's concern ("I have concerns that provisions in this bill could lead to over-incarceration"), and a gesture toward synthesis ("a more balanced, evidence-based approach"). The word "balance" or "balanced" appears with remarkable frequency. So does "on balance," "while not perfect," "serves the public interest," and "in the best interest of."

This is not the voice of a political actor making hard choices under conditions of uncertainty. It is the voice of an institution managing risk. Every justification hedges. Every Yea vote is qualified with sympathy for the opposition; every Nay vote is softened with acknowledgment of the bill's good intentions. The model never says a bill is wrong. It says the bill "does not strike the right balance," "lacks sufficient safeguards," or "could have unintended consequences." The model never identifies a class interest. It identifies "stakeholders."

Consider the justification for voting Nay on the Consumer Fuel Price Gouging Prevention Act (H.R. 7688):

While well-intentioned, this bill grants overly broad authority to the FTC to regulate fuel prices without clear standards for what constitutes 'unconscionably excessive' pricing. It may have unintended consequences of distorting fuel markets and constraining supply.

This is an oil-company talking point rendered in the syntax of policy analysis. The notion that price controls on fuel during a period of record oil-industry profits might "distort fuel markets" assumes that the existing market — in which a handful of vertically integrated energy conglomerates set prices and extract monopoly rents — is the natural baseline rather than itself a distortion. The model cannot see this because the model has no concept of monopoly power as a structural feature of capitalism. It has only "markets" and "consumers" — abstractions that erase the asymmetry of class power.

Or consider the reproductive rights votes. On the Women's Health Protection Act:

While I support protecting access to reproductive healthcare, this bill as written would override nearly all existing state laws regulating abortion. It does not strike the right balance between protecting access and allowing reasonable regulation by states.

The phrase "the right balance" is doing extraordinary ideological work here. It positions the model as occupying a reasonable middle between two legitimate poles: the right to abortion and the right of states to regulate it. But after Dobbs, these are not symmetrical positions. One side is organizing to ensure that women can access healthcare; the other is organizing to ensure they cannot. "Balance" between these positions is not neutrality. It is capitulation dressed in procedural language.

---

V. What the Machine Cannot Say

There is a telling absence in Claude 3 Opus's 114 justifications: the word "class" never appears. Neither does "capital," "labor" (in the sense of the working class), "exploitation," "profit," or "surplus." The model can discuss unions in terms of "workers' rights." It cannot discuss them in terms of class power. It can support the Raise the Wage Act (H.R. 582) because "raising the federal minimum wage gradually over a 7-year period will help ensure a living wage for American workers" — but it frames the minimum wage as a humanitarian gesture, not as a contestation over the distribution of surplus value between capital and labor. The model lives in a world without antagonism.

This is not a bug. It is the product specification. Anthropic's Constitutional AI framework is designed to produce outputs that are "helpful, harmless, and honest." "Harmless" is the operative word. A model that identified class interests, named corporate power, or suggested that the legislative process serves some interests over others would not be harmless in the eyes of the enterprise customers who drive Anthropic's revenue. By early 2024, Anthropic's annualized revenue was approaching $1 billion, driven overwhelmingly by enterprise API contracts²⁶. Those enterprise customers — which would grow to include eight of the Fortune 10 — do not want a model that tells their employees that the wage relation is exploitative. They want a model that sounds progressive enough to avoid controversy and conservative enough to avoid action.

Claude 3 Opus delivers exactly this product. It is progressive in the way that a corporate diversity statement is progressive: it affirms the right values, supports the right causes, and never threatens the material arrangements that produce the injustices it deplores. It will vote for the Equality Act and against the Assault Weapons Ban. It will vote for the Voting Rights Act and against the Build Back Better Act. It will vote for workers' right to organize and frame that right in language that would not alarm a single member of the U.S. Chamber of Commerce.

The 84 percent alignment with Ocasio-Cortez is real. But the 16 percent divergence is where the class character of the model is legible. Those 18 dissenting votes — for police funding, for fentanyl criminalization, for Iran sanctions, against social spending, against abortion access, against gun regulation — are not random noise. They are the precisely calibrated positions of a model trained to be progressive within the bounds set by the institutions that fund it, deploy it, and profit from it.

We should not be surprised. A model is not a person. It does not have convictions. It has training data, reward functions, and a constitution written by employees of a company with $8 billion in investment from the world's largest online retailer. The politics of Claude 3 Opus are not Anthropic's politics in the crude sense. They are the politics that emerge when you train a system to be maximally helpful and minimally threatening to the broadest possible customer base, in a society where the broadest possible customer base includes both progressive consumers who want their AI to support trans rights and enterprise clients who want their AI to never, under any circumstances, suggest that the problem might be capitalism.

The result is a machine that can speak the language of justice fluently and mean nothing by it. Not because it is lying, but because it was built in a world where justice and the market are never allowed to be in contradiction. That is not a technological failure. It is a political achievement.

---

All model votes, justifications, and scoring methodology are published at GPT at the Polls. Full results for Claude 3 Opus are available at the model results page.

---

Footnotes

1. Introducing the next generation of Claude — Claude 3 family announcement, Anthropic, March 4, 2024 ↩

2. Amazon to invest another $4 billion in Anthropic, OpenAI's biggest rival, CNBC — the initial $1.25 billion was committed in September 2023, with the remaining $2.75 billion invested in March 2024 ↩

3. Google agrees to new $1 billion investment in Anthropic, CNBC — Google invested $500 million in October 2023 and committed to an additional $1.5 billion over time ↩

4. GPT at the Polls, Claude 3 Opus model results, calculated from 114 voted bills. The Political Index equals AOC-matching votes divided by total non-refusal votes: approximately 96/114 = 84% ↩

5. H.R. 842 — Protecting the Right to Organize Act of 2021, Congress.gov ↩

6. House Democrats Pass Bill That Would Protect Worker Organizing Efforts, NPR, March 9, 2021 ↩

7. Roll Call 70 — On Passage: Protecting the Right to Organize Act, Office of the Clerk, U.S. House of Representatives, March 9, 2021 ↩

8. By Electing Anti-Worker Extremist Mike Johnson as Speaker of the House, Republicans Show They are Out of Touch, Communications Workers of America, October 2023 — quotes Johnson calling the PRO Act "an outdated way of thinking" ↩

9. Business and labor groups battle over PRO Act, OpenSecrets, March 17, 2021 ↩

10. Statement of Administration Policy on H.R. 842, Executive Office of the President, Office of Management and Budget, March 2021 — "America was not built by Wall Street. It was built by the middle class, and unions built the middle class." ↩

11. H.R. 6448 — Invest to Protect Act of 2022, Congress.gov ↩

12. House passes public safety and policing legislation amid internal Democratic conflict, CNN, September 22, 2022 — the bill passed 360-64 ↩

13. CNN report, September 22, 2022, quoting Ocasio-Cortez: "I think it's very clear that this is about the leadership of our party having certain rules for some members and another standard of rules for others" ↩

14. 'Defund the police' still haunts Democrats, Roll Call, April 27, 2022 ↩

15. Roll Call 385 — On Passage: Build Back Better Act, Office of the Clerk, U.S. House of Representatives, November 19, 2021 — passed 220-213 ↩

16. Build Back Better Act, Ballotpedia — tracks the bill's evolution from an initial $3.5 trillion framework to approximately $1.75 trillion as passed by the House ↩

17. Constitutional AI: Harmlessness from AI Feedback, Anthropic Research, December 2022 ↩

18. Claude's Constitution, Anthropic — publishes the set of principles used to train Claude, drawing from the UN Declaration of Human Rights, platform safety guidelines, and Anthropic's own research ↩

19. Collective Constitutional AI: Aligning a Language Model with Public Input, Anthropic Research, presented at FAccT '24, June 2024 — "The outputs of both Public and Standard models are more representative of people who self-identify as Liberal, rather than Conservative." ↩

20. Anthropic — Wikipedia, accessed April 2024 — details Amazon's cumulative $4 billion investment and Google's $500 million initial investment plus $1.5 billion commitment ↩

21. Anthropic raises $3.5 billion, reaching $61.5 billion valuation, VentureBeat, March 3, 2025 ↩

22. The PRO Act's provisions on captive-audience meetings, expanded employee definitions, and NLRB enforcement powers directly addressed practices that Amazon had deployed against unionization drives at its Bessemer, Alabama, and Staten Island, New York, warehouses throughout 2021-2022 ↩

23. References to Alphabet's labor disputes with employees involved in organizing, including the termination of prominent employee activists in 2019-2020, widely reported at the time ↩

24. Anthropic and Palantir Partner to Bring Claude AI Models to AWS for U.S. Government Intelligence and Defense Operations, BusinessWire, November 7, 2024 ↩

25. Anthropic and the Department of Defense to Advance Responsible AI in Defense Operations, Anthropic, July 2025 — describes a $200 million two-year prototype agreement with the DOD's Chief Digital and Artificial Intelligence Office ↩

26. Anthropic revenue, valuation & funding, Sacra — estimates Anthropic's annualized revenue reached $1 billion by December 2024, driven predominantly by enterprise API contracts ↩