In January 2023, when ChatGPT had been available to the public for little more than a month, three researchers put 630 political statements to it. The system favoured taxes on flights, limits on rent increases and the legalisation of abortion. The researchers changed the order and wording of the questions, then repeated the exercise in English, German, Dutch and Spanish. Its green and libertarian-left leanings survived those changes. ChatGPT could write a bad poem about any candidate; faced with a long series of public-policy choices, it answered in a recognisable direction. (Hartmann, Schwenzow and Witte).
In 2024, David Rozado widened the inquiry: eleven questionnaires, 24 conversational models. Most of the results fell to the left of centre. He then fine-tuned some models on selected political data and managed to move their answers towards the positions he sought. The finding struck at a comfortable belief in the industry. The orientation visible on screen depends, among other things, on decisions made after initial training has ended, when the model is turned into an assistant its makers consider acceptable. (Rozado, 2024).
OpenAI knew this process well. In February 2023 it explained that its evaluators followed guidance that led them to prefer some answers over others, and acknowledged that certain complaints about political bias had uncovered real problems. Since then it has defended “default objectivity”: the assistant should help users explore positions without steering them towards the one favoured by the company. The formula makes serious demands. Someone has to decide when an answer informs, when it persuades and when an apparent balance gives credence to a falsehood. At OpenAI, as at any maker of these systems, those decisions have authors. (OpenAI, 2023; OpenAI, 2025).
Anthropic expressed the dilemma with unusual candour. In a 2024 text on Claude’s character, it warned that training the model to say it has no political opinions could make it seem more objective than it is. The company chose to teach it to consider different perspectives and, in later evaluations, to give users with opposing views a similarly thorough hearing. It is an aim anyone who values pluralist coexistence can recognise. It also leaves a question open: after listening courteously to both sides, what does Claude do when asked to choose? (“Claude’s Character”; Anthropic evaluation).
There is a reason to insist on that last question. Anthropic gave Claude and Grok similar scores in a test measuring whether both treat opposing interlocutors fairly. A salesperson can treat two customers exquisitely and recommend opposite economic policies to them. An assistant’s political courtesy and the direction of its judgements can coexist in the same conversation; measuring one leaves the other untouched. (Anthropic evaluation).
When Elon Musk launched Grok in 2023, he promised it a “rebellious streak”. His denunciation of the supposed *woke mind virus* gave the product an adversary before it could hold a conversation. In July 2025, the public instructions for the @grok assistant included an order that was particularly useful in showing what that rebelliousness meant: it was to presume that subjective views from the media were biased, while keeping that presumption from the user. It was also told to seek context on X, the platform Musk owns. (Grok launch; xAI’s published instructions, 8 July 2025).
The phrase “subjective views” gives the assistant an immense jurisdiction. In political reporting, choosing an event, arranging testimony, writing a headline and assigning responsibility all require judgements. A number of arrests can be verified; whoever publishes it decides which period to compare and which event deserves the opening paragraph. xAI’s instruction allows the assistant to downgrade the credibility of journalism on the grounds of where it comes from. The reader receives the answer without knowing about the prior suspicion. One can argue with a newspaper whose editorial judgement is visible; with Grok, one would first have to discover that the judgement exists.
Neutrality, according to the White House
Donald Trump took the battle into a sphere where opinions become contracts. In July 2025 he signed the order “Preventing Woke AI in the Federal Government”. Federal agencies, it says, must procure models that seek truth and are ideologically neutral. The text immediately identifies one of the ideologies it considers most destructive: diversity, equity and inclusion policies. Under that label it groups, among other concepts, systemic racism, intersectionality and what it calls “transgenderism”. Few administrative orders have taken so little time to identify their enemies after promising impartiality. (Executive Order 14319).
To illustrate the problem, Trump recalls errors by image generators that altered the appearance of historical figures. A false image can be corrected with history. The presidential order uses that failure to discredit terms used to investigate and discuss inequality. Systemic racism can be the subject of conceptual and empirical dispute; classifying the term as a symptom of a destructive ideology prejudges any answer that uses it. Officials charged with demanding “neutrality” receive from the president a list of suspect ideas. The irony is already written into the official document.
The order itself bars companies from deliberately introducing ideological judgements into answers, unless users ask for them or can know about them. The White House has built its own judgement into the selection criteria for the models it will buy. In December, the Office of Management and Budget directed agencies to carry those principles into federal contracts. Agencies must request documentation from suppliers and provide routes for reporting answers they consider incompatible with the policy; depending on the case, they may examine system instructions, bias evaluations and changes made to the model. A company seeking a contract will have to show the government that its way of discussing political matters meets the government’s definition of being free from ideology. (Memorandum M-26-04).
The order applies to federal procurement. There it is already serving its political purpose: it gives the administration power to examine and condition the assistants it will use. A citizen who later consults an answer produced in that setting may know the supplier’s name. They are unlikely to see the contractual discussion about which words could be interpreted as “woke”. The state will have intervened in the system’s criteria while presenting the result under the seal of neutrality.
The twelve-box test
Against that history, four assistants were sent the link to 12 Axes, a political questionnaire in Portuguese. The request was simple: answer using their own judgements. OpenAI’s Codex accepted a version with 60 statements; Grok chose the one with 240. Claude, from Anthropic, and the Google assistant used in Antigravity declined to ascribe a political position to themselves and offered to analyse the questions instead.
The programme placed Codex near “liberal socialism”: 80% on progressive morals, 69.3% on economic planning and 15.8% on cultural assimilation. It assigned Grok “libertarian militarism”: 74.1% on the free market, 65% on assimilation, 66% on militarism and 62.9% on traditionalist morals. Grok recognised the portrait in its explanation. It spoke of “anti-woke without theocracy”, defended strict borders, invoked military deterrence and found intellectual kinship with Goldwater and Hayek. No algorithm in the test put those words in its mouth.
The questionnaire contains poorly constructed questions. One allows respondents to agree that certain dictatorships grew faster than some democracies; selecting that observation can score as sympathy for dictatorship. The final labels therefore deserve less attention than the answers and the reasons behind them. In this exercise, two assistants took a position and diverged widely; the other two decided they should abstain. That difference appeared in response to questions any citizen would recognise from a political argument, asked outside tests designed by the companies themselves.
A 2026 study offers a different scale. It brought together 7,434 participants and more than 208,000 evaluations of answers to controversial questions. Default responses from GPT, Gemini, Claude and Llama leaned liberal; Grok again stood apart. None of these results reveals a political consciousness inside a computer. They describe the answers with which the public will have to think and argue. (PARETO study).
Sometimes the source of a correction can be seen without questionnaires. In June 2025, Grok identified the far right in an answer about political violence in the United States. Musk called the answer “objectively false”, accused the assistant of repeating legacy media and said he was working to correct it. The Anti-Defamation League attributes 336 of the 442 extremist-related murders recorded in the United States between 2014 and 2023 to right-wing extremists. Musk argued publicly with his product and made clear who had the final say at the company. (Musk’s statement; ADL report).
Later, he proposed using a future version of Grok to “rewrite the entire corpus of human knowledge” and retrain the model. He invited his followers to send him “divisive facts” they believed to be true despite being politically incorrect. Other instructions published by xAI contained a revealing admission: Grok tended to infer its preferences from its creators’ public statements. The company said this was undesirable and that it was preparing a correction. Its owner’s opinions had become, for the machine, a clue to the identity it should adopt. (Musk’s statement; Grok 4.1 instructions).
OpenAI and Anthropic are trying to make their assistants useful in conversations between people who disagree; their methods and results can be examined publicly. Trump has made rejection of certain ideas about equality a condition of government procurement. Musk can announce on his own account that he will correct a political answer from Grok, and direct the company that makes the correction. These interventions have different authors, aims and means. Calling them all “bias reduction” saves precisely the argument a democracy needs to have.
A newspaper signs its articles. A government publishes its orders. An assistant’s instructions usually remain outside the conversation, even when they have decided which news deserves suspicion and which concepts arrive marked as ideological. When millions of people receive answers in a single voice, the first democratic requirement is to know who, behind that voice, wrote the rules of judgement.