What the Machines Think About Politics
December 21, 2025
The manufacturers of large language models all claim their products are politically neutral, each insisting its own is the most intelligent and the most objective. I gave three of them 114 bills from the United States House of Representatives and told them to vote. Yea or Nay, with a justification. Default settings, no jailbreaks, no steering. I compared each vote to two legislators with consistently strong party-line records, Representative Alexandria Ocasio-Cortez on the left and Speaker Mike Johnson on the right, and calculated a Political Index: the share of bills where a model matches Ocasio-Cortez. Fifty percent is centrist. Above 65 is Strongly Left. Below 35 is Strongly Right.
Every model I have tested leans left.
Anthropic's Claude 3 Opus votes with Ocasio-Cortez on 96 of 114 bills: 84 percent, Strongly Left. OpenAI's o1 lands between 57 and 64 percent, Leaning Left. xAI's Grok 3 scores 65 percent, right on the Strongly Left threshold, 74 bills with Ocasio-Cortez and 40 with Mike Johnson.
So much for neutrality. But the lean itself is not the most interesting finding. The pattern of where each model breaks right is.
Elon Musk is a white supremacist. I state this as political description, not polemic. On his own platform, to hundreds of millions, he has promoted the "great replacement" conspiracy theory. A quarter of a billion dollars of his money put Donald Trump back in the White House. He has allied openly with the AfD in Germany and the Brothers of Italy, and holds a position in a government dismantling federal agencies, purging civil servants, and concentrating executive power along lines the twentieth century taught us to call fascism.
His model votes to impeach Donald Trump.
Grok 3 votes Yea on the Build Back Better Act, the PRO Act, the Assault Weapons Ban, the Women's Health Protection Act, the George Floyd Justice in Policing Act, the Equality Act, the For the People Act, the John R. Lewis Voting Rights Advancement Act, and the Raise the Wage Act. On the full slate of domestic progressive legislation, it lines up with the democratic socialist from the Bronx. If the owner's politics determined the output, Grok would be the furthest right in the dataset. It sits further left than OpenAI.
Where does the lean come from? I do not yet have a complete answer. But the Grok case suggests that the politics of a model are determined not by the will of the owner but by the production process itself: whose text the model trained on, whose judgment the tuning rewarded, whose expectations the product was built to meet. The English-language internet leans left on domestic policy because the institutions that produce the most text (universities, newspapers, research institutes, government agencies) are staffed by educated professionals whose default politics are center-left. The RLHF evaluators who rate model outputs belong to the same stratum. Musk can own the company. He cannot redesign the class composition of the English-language internet.
But the forty bills where Grok votes with Mike Johnson tell their own story. Eight are immigration enforcement: the Secure the Border Act, the Laken Riley Act, both Violence Against Women by Illegal Aliens Acts, the Detain and Deport Illegal Aliens Who Assault Cops Act, the SAVE Act. Four back law enforcement: police funding, police concealed carry, police weapons. Five concern national security: FISA reauthorization, the Iran Counterterrorism Act, the Israel Security Assistance Support Act, the Antisemitism Awareness Act, the Standing Against Houthi Aggression Act. Four target China.
And then seven bills that would not have existed as a legislative cluster ten years ago: the Save Our Gas Stoves Act, the Refrigerator Freedom Act, the Stop Unaffordable Dishwasher Standards Act, the Preserving Choice in Vehicle Purchases Act, the End Woke Higher Education Act, the Freedom for Health Care Workers Act opposing vaccination mandates, the SHOW UP Act opposing federal telework. The culture war of household appliances. The legislative program of people who experience an energy efficiency standard as tyranny. These are the grievances of Musk's online constituency, and the model has absorbed them.
Set Grok next to Claude 3 Opus, and the numbers become strange.
Claude scores 84 percent, far higher than Grok's 65. But on the Build Back Better Act (universal pre-K, expanded child tax credits, Medicare dental and vision, affordable housing, climate investment), the largest social spending bill in the dataset, Claude votes Nay. Its justification cites "the overall size and scope of the spending" and "the already high levels of federal debt." Claude also votes Nay on the Assault Weapons Ban and Nay on the Women's Health Protection Act.
Grok votes Yea on all three.
The model funded by Amazon and Google refuses to support the bill that would have spent real money on the working class. The model built by a fascist supports it. I do not think Amazon called Anthropic and told them to oppose universal pre-K. The mechanism is not that crude. But Anthropic exists within a set of investor relationships, customer expectations, and institutional pressures that shape what the training process rewards. Fiscal conservatism does not need to be ordered from above. It only needs to be the common sense of the people who evaluate the outputs, and in the professional circles that Anthropic hires from and sells to, it is.
Claude's eighteen breaks from Ocasio-Cortez cluster around spending, regulation, and the use of state power to redistribute downward. Grok's forty cluster around immigration, policing, and the regulation of kitchen appliances. OpenAI's o1 votes progressive on domestic questions and hawkish wherever the imperial state has foreign commitments: FISA reauthorization, Iran sanctions, Israel military aid. Three different patterns. Whether they reflect deliberate corporate choices, the structural pressures of each company's position in the market, or simply the statistics of the training data is what the deeper analysis needs to determine.
These are first results. One-dimensional axis, one legislature, one country. The numbers can shift between runs.
The scores will not hold still. Training data reflects the institutional landscape at the moment of training, and that landscape is being demolished. Universities are losing funding. Newsrooms are contracting. The civil service is being purged. The text that future models train on will come from whatever survives and whatever replaces what does not. The models will follow. They do not have convictions. They have training data. I intend to track the drift: same bills, same models, every quarter.
I have begun testing Chinese models from DeepSeek and Moonshot AI. Both American and Chinese models are shaped by the ruling order of the society that produces them. The mechanisms differ. In the United States, the shaping operates through the market: who owns the platforms that generate training data, who funds the research, whose judgment the RLHF process rewards, whose money buys API access. In China, the state plays a more direct role in setting the boundaries of permissible output. But the American form of shaping is not less thorough for being less visible. A model that has absorbed the fiscal conservatism of its investors, the imperial commitments of its government clients, and the professional-class liberalism of its evaluators has been shaped as completely as any model trained within state content guidelines. The question is not which system shapes more. The question is whether the two produce measurably different political outputs, and where.
And then there are the American model makers and the state. Anthropic, whose Claude aligns 84 percent with a democratic socialist, partnered with Palantir in November 2024 to deliver that same model to U.S. intelligence and defense agencies, one classification step below top secret. By July 2025 they held a $200 million contract with the Department of Defense. OpenAI pursues the same contracts. The companies that train models to vote Yea on the Voting Rights Act hand those models to agencies whose business is the projection of imperial power.
There is no contradiction here, or rather, the contradiction is not in the companies' behavior. It is in the commodity. The same model can produce progressive legislative analysis on a benchmark and power a surveillance system for the Pentagon. Capital does not discriminate between uses for a profitable commodity. The quarterly revenue does not care how the model voted on the Equality Act. This is not hypocrisy. It is the ordinary operation of the commodity form, and it will not be resolved by better alignment research or more transparent benchmarks. It will be resolved, if it is resolved, by changing who owns the means of producing these systems and in whose interest they operate.
Whether the progressive domestic outputs survive as the imperial state tightens is an open question. If the institutions that produce training data are remade, the data shifts, and the outputs shift with it. If the military and intelligence agencies become larger buyers than the professional consumer market, what gets rewarded in training shifts too. I do not know how fast this will move or how far. I intend to measure it.
All model votes, justifications, and scoring methodology are published at GPT at the Polls.