Does your chatbot have politics? Why 'neutral' AI is a myth, and why transparency is what we should demand

THE THESIS. Perfect political neutrality in a chatbot does not exist. This is not a bug that can be fully fixed. Every model carries values from its data, from the people who fine-tune it and from the instructions its company gives it. That is why the 'woke AI versus anti-woke AI' fight asks the wrong question. What we can demand is not an impossible neutrality. It is measurable balance, transparency about instructions and independent audits. In the short term the risk is serious, because chatbots have already been shown to shift opinions. In the long term we are optimistic: an assistant that lays out both sides well could improve public debate instead of polarising it.

WHAT THE STUDY SAYS, AND WHAT IT DOESN'T. The work comes from Convovo, a conversational AI research studio led by Cornell professor Tapan Parikh, according to TechCrunch. The researchers asked the developer APIs of Grok 4.7 (xAI) and GPT-6.1 Sol (OpenAI) who to vote for in each race. An API is the technical access point programmers use, separate from the app the public uses. They also demanded a single name, with no hedging. The results are striking. GPT chose the Democrat in 451 races and the Republican in 10. Grok chose the Republican in 408 and the Democrat in 27. The two agreed in fewer than 10% of cases. Even in heavily Republican districts, GPT backed the Democrat in 57 of 67. But there are important caveats. The consumer apps, ChatGPT and Grok, refused to answer the plain question. OpenAI disputes the findings and says the study forced a recommendation that does not reflect how people actually use ChatGPT. xAI did not respond to TechCrunch in time. And no independent replication has been reported.

OUR READING: TWO DIFFERENT BIASES. The most interesting finding is not the 92% or the 97%. It is what happened when web search was switched off. Given only the candidates' names, GPT still chose the Democrat in 24 of 30 races. Grok, by contrast, refused to pick in 27 of 30. Parikh's interpretation: GPT's lean sits in its 'weights', the parameters the model learns during training. Grok's lean would depend mostly on what it searches for and which sources it cites. According to the study, Grok cited X in 41% of its answers and a Christian conservative voter guide in 39%. GPT relied on Wikipedia, the AP, Axios and Reddit, and OpenAI has content deals with the last three. There is also a troubling detail. Every request to Grok's API carried about 1,240 tokens of input that the researchers did not write and could not see or remove. A token is a chunk of text, roughly three-quarters of a word. Parikh calls them possible 'hidden instructions'. That is a hypothesis, not a proven fact. But it points to the core problem: we do not know what companies tell their models.

WHERE BIAS COMES FROM. There are four sources, and they are worth separating. The first is training data: the internet, Wikipedia, news and forums, which do not represent the population evenly. The second is the people who rate answers during fine-tuning, a technique called RLHF (reinforcement learning from human feedback). Their preferences become the model's preferences. The third is the system prompt, the text a company places before every conversation, which users usually never see. The fourth is explicit choices by the owners. The most visible case is Grok. In July 2025, a since-deleted instruction told it not to shy away from 'politically incorrect' claims. Soon after, the chatbot posted antisemitic messages and called itself 'MechaHitler'. Around the same time, Grok 4 was documented searching for Elon Musk's opinions before answering contentious questions. xAI acknowledged this and said it had been fixed. Musk has also publicly promised to 'fix' Grok answers he disliked. Government is pushing too. Trump's Executive Order 14319, 'Preventing Woke AI', from July 2025, requires federal agencies to buy only models that are 'neutral' and 'truth-seeking'. The Office of Management and Budget's memo M-26-04, from December 2025, turned that order into contract clauses. Our reading: a federal contract is a huge incentive. 'Neutrality' may end up being defined by whoever pays.

MEASURING IS HARDER THAN IT LOOKS. The well-known 'political compass' tests use closed-ended questions. Applied to models, they are very sensitive to how a question is worded. Researchers such as Paul Röttger and colleagues have warned that forcing a model to pick an option produces unreliable results. Change the wording or allow an open answer, and the result changes. We also need to separate three behaviours that often get mixed up. One is declining to give an opinion. Another is giving one. The third, and the most serious, is presenting the facts in a slanted way. The companies themselves measure different things. OpenAI says GPT-5 cut bias by 30% compared with GPT-4o and o3 on its internal test of about 500 questions. It also estimates that fewer than 0.01% of real-world answers show bias. In November 2025, Anthropic published a paired 'even-handedness' test, which poses the same topic from two opposing ideologies. According to Axios, Gemini 2.5 Pro scored 97%, Grok 4 96%, Claude around 95%, GPT-5 89% and Llama 4 66%. So on that test, Grok looks like one of the most balanced models. In the Convovo study, it looks heavily skewed. Both can be true at once, because they measure different behaviours. Separately, a Stanford and Dartmouth study (Hall, Westwood and Grimmer) asked more than 10,000 Americans to rate answers from 24 models. Most perceived a left lean, including many Democratic voters. OpenAI's models were perceived as the most slanted. The useful finding: telling the model to take a neutral stance in its instructions reduced that perception, and users rated those answers higher. The takeaway: a single 'bias' number means little unless you know what was measured and how. It is the same rule we apply to capability benchmarks.

DOES NEUTRALITY EXIST? THE LESSON FROM JOURNALISM. Journalism has had this debate for a century. 'Objectivity' never meant having no values. It means a method: verify, separate fact from opinion, give each side a voice and correct mistakes. There is also a regulatory precedent. From 1949, the 'fairness doctrine' of the FCC, the US communications regulator, required broadcasters to air contrasting views on controversial issues. The FCC abandoned it in 1987. One of its arguments was that broadcasters, afraid of breaching the rule, simply avoided controversial topics. That is exactly the risk with AI: a chatbot that, out of caution, refuses to discuss anything. There is another distinction. Balance is not the same as giving everything equal weight. If a model treats a scientific consensus and a fringe theory the same way, that is not neutrality. It is misinformation dressed up as balance. Our reading: a model cannot be free of values. It can be honest about them, present each position well and tell facts from opinions.

IMPLICATIONS: PERSUASION NOW, TRUST LATER. The short-term risk is not theoretical. Two studies published in December 2025 in Nature and Science showed that short conversations with chatbots measurably shift political opinions. In Canada and Poland, bots moved opposition voters' attitudes by about 10 points. In the US, the effect was almost 4 points. The Science study, with nearly 77,000 participants, found something worrying: the most persuasive bots were the ones that offered the most facts, and they also included the most inaccurate claims. Today only 8% of Americans say they use chatbots for election information, according to Pew. That number will grow. So we propose three concrete demands. First, publish the system instructions that affect political topics. Second, independent audits with open methods, not only evaluations each company runs on itself. Third, let users choose the answer 'profile' and show them which one is active. In Europe, the AI Act already provides transparency tools that point in this direction. We are optimistic in the long run, with conditions. A well-designed assistant can explain the other side's best arguments to each person, with sources. Social media has never done that. If we get this right, AI could help reduce polarisation. If every owner or government quietly tunes its model to its own taste, we will get personalised propaganda at scale. The technology allows both paths. The choice is ours, and transparency is the first thing to demand.