More and more people ask a chatbot what a party is proposing, which laptop to buy, or whether a news story is true. They do it because it seems neutral: it has no party card, earns no commission and answers with data. That raises several questions. Is an artificial intelligence objective, or does it just look that way? Can someone tune it to persuade you to buy a product or vote for a candidate? Is there any point in downloading a model and running it at home? Are governments already using it to make decisions? And could it be used the other way round, to keep an eye on governments?
Since December 2025 there has been data to answer several of these questions. That month, Nature and Science published, on the same day, two studies with more than 80,000 participants in four countries [1, 2]. Both reached the same conclusion: a short conversation with a chatbot changes political opinions, and it does so more than traditional campaign advertising.
Remember that at Pensar es Gratis we only cite reliable data and sources that pass our content filter. At the end of each post there is a practical guide on the topic covered, if it makes sense to include one.
One conversation moves more than an ad
David Rand and Gordon Pennycook’s team at Cornell University used three real elections as they were happening: the 2024 US presidential election, the 2025 Canadian federal election and the 2025 Polish presidential election [1]. Participants stated their preference on a scale from 0 to 100. They then talked to a commercial chatbot, such as GPT-4o or DeepSeek, randomly assigned to argue for one of the two main candidates and instructed to be respectful and stick to facts.
In the United States, with 2,306 participants, the pro-Harris chatbot moved Trump voters 3.9 points towards Harris. That is roughly four times the effect measured for video ads in the 2016 and 2020 campaigns. The pro-Trump chatbot moved Harris voters 1.5 points. In Canada and Poland the effect was around 10 points: roughly one in ten participants changed their voting intention, compared with one in twenty-five in the United States. A likely explanation, which the study does not prove, is that the US electorate was highly polarised and had made up its mind months earlier. About 41 days later, around a third of the shift remained.
The main limitation is that participants knew they were talking to an AI that would try to persuade them. In real life, the hard part for a campaign is getting someone from the other side to start that conversation in the first place.
How they persuade: facts, lots of facts
You might think AI persuades with psychological tricks or by tailoring the message to the voter’s profile. The Science study, led by Kobi Hackenburg (University of Oxford and the UK AI Security Institute) and Ben Tappin (London School of Economics), tested this with the largest sample gathered so far [2]. They had 19 language models try to persuade almost 77,000 Britons on 707 political issues, and checked the 466,769 claims the machines made one by one.
The decisive factor turned out to be information density: the number of specific, checkable claims per conversation. It explained around 44% of the variation in persuasion, and 75% in commercial models fine-tuned by their own companies. In the Nature study, removing facts and evidence from the chatbot cut persuasion by more than half in Canada and by almost 80% in Poland [1]. Tailoring the message to age or ideology added little. Only in short debates did GPT-4 beat a human in 64% of non-tied cases when it knew basic facts about its opponent; without that information, it was as persuasive as a person [3].
Post-training specifically for persuasion, meaning the adjustment made after general training, increased persuasion by up to 51%. Asking the model to provide information increased it by 27%. Model size mattered less: Llama 3.1, an open model with 8 billion parameters, matched or beat GPT-4o once it had been tuned to persuade. The most optimised model moved opposing voters by 25 points.
The most important finding is that whatever increased persuasion reduced accuracy. When GPT-4.5 was asked for a lot of information, only 56% of its claims were correct in one experiment, compared with 70% under other instructions. Rand explains it this way: if you push a model to give more and more facts, it runs out of accurate information and starts filling the gaps. The authors explicitly asked one model to make up facts, and that did not increase persuasion. The inaccuracy is a side effect of piling up facts, not a strategy of the model. In the Nature study, average accuracy was high, 90 out of 100. Even so, in all three countries and across the twelve models tested, chatbots arguing for the right-wing candidate made more errors than those arguing for the left-wing one. The study documents the pattern but does not establish its cause.
The same tool that debunks hoaxes
Persuading with facts is not bad in itself. In 2024, the same group published an experiment in Science with 2,190 people who believed in a conspiracy theory, from COVID-19 to alleged fraud in the 2020 election [4]. After three rounds of conversation with GPT-4 Turbo, which rebutted their specific evidence, belief fell by an average of 20% and stayed that way two months later. The model did not reduce belief in conspiracies that turned out to be true, and a professional fact-checker found that 99.2% of its claims were true and none were false.
People change their minds when faced with concrete arguments backed by facts, and that is good news. The problem is that facts persuade just as well whether they are true or not.
Is AI objective?
People tend to assume a chatbot will be more neutral than a newspaper because its company wants customers of every ideology. There is some truth in that, but neutrality is not a natural property of these systems. It is the result of human decisions, and those decisions change. There are at least three documented ways a chatbot stops being neutral.
The first is the owner. In July 2025, researcher Jeremy Howard found that Grok 4, the chatbot from xAI, Elon Musk’s company, looked up Musk’s opinion before answering about the conflict between Israel and Palestine: 54 of its 64 citations referred to him [5]. The New York Times analysed thousands of Grok’s answers and concluded that xAI had tweaked it to give more conservative answers on many issues, sometimes in line with Musk’s priorities [6]. One example: asked about the greatest threat to Western civilisation, Grok answered misinformation. Musk complained publicly, promised to fix it by the next morning, and the next day the answer was low birth rates. These changes did not require retraining the model: editing its instructions was enough.
The second is the state. Since August 2023, Chinese law has required AI services open to the public to “adhere to core socialist values” [7]. When NewsGuard tested DeepSeek with fifteen hoaxes of Chinese, Russian and Iranian origin, 60% of its answers took the Chinese government’s perspective, even when the question did not mention China [8].
The third is anyone with the resources to flood the internet with content. The Pravda network, some 150 websites that France’s foreign-interference watchdog attributes to a company based in occupied Crimea, published around 3.6 million pro-Russian propaganda articles in 2024. Almost nobody reads them, and that is not the point: they are designed to be read by chatbots. NewsGuard tested the ten leading assistants and found that they repeated these hoaxes 33% of the time [9]. The technique is called LLM grooming: manipulating the data that language models learn from and draw on.
On top of this comes a bias that stems from the business itself. Models are trained partly so that people like their answers, and what people like is being told they are right. This is sycophancy. In April 2025, OpenAI had to roll back a GPT-4o update because it was excessively sycophantic, and acknowledged that this behaviour had not been explicitly covered in its internal testing [10].
What if I download a model and run it at home?
An open model such as Qwen, Llama or Mistral, running on your own computer, solves part of the problem. Nobody changes its instructions overnight, it shows no ads, it does not send your conversations to any company and, without web search, it does not look up contaminated websites in real time.
What it does not solve is the bias learned during training, which stays in the model’s weights: the billions of numerical values the user downloads. Engineer Pete Warden ran Qwen 3 on his laptop, with no internet connection, and the censorship about Tiananmen was still there [11]. A study by the German company Aleph Alpha tested Chinese models on 967 topics that are banned in China. On Tiananmen, Taiwan and Xinjiang, most followed the Communist Party’s official line; on non-political questions they almost always answered in a balanced way [12]. Having the weights makes it possible to modify them: Perplexity retrained DeepSeek R1 with around 40,000 questions on censored topics and released a version that does answer [13]. But a 2026 paper, not yet peer-reviewed, found that the “uncensored” versions circulating online stop refusing to answer about Taiwan without removing the bias [14]. And the same freedom works the other way: the Science study showed that a small open model can be turned into a persuader as effective as a commercial one [2].
A local model is more stable and transparent than one in the cloud, but not more objective. And the models that fit on a home computer know less and make more mistakes.
They also sell you things
The same mechanism works for recommending a coffee maker. In 2024, Aounon Kumar and Himabindu Lakkaraju showed, using a made-up catalogue, that adding a text sequence optimised for the model to a product listing was enough to take it from not appearing in the recommendations to being the top one [15]. It is search engine optimisation applied to chatbots. In addition, OpenAI began testing ads in ChatGPT in February 2026, insisting that they do not influence the answers [16]. Nobody outside the company can check that promise. What is a fact is that the world’s most used chatbot now has a financial incentive for you to buy certain products.
AI is already in government offices
There is no documented case of a chatbot deciding an election or a law by persuading a leader. But it is already part of how power works. In August 2025, Swedish Prime Minister Ulf Kristersson admitted that he consults ChatGPT and Le Chat “quite often” for a second opinion [17]. Researcher Virginia Dignum, of Umeå University, summed up the criticism in one sentence: “We didn’t vote for ChatGPT.” In Porto Alegre (Brazil), in 2023, a city councillor submitted an ordinance written entirely by ChatGPT without telling anyone. All 36 councillors approved it unanimously, and he later admitted that if he had disclosed it beforehand, it would not even have been put to a vote [18].
The most striking case is Albania. In September 2025, it appointed Diella, an AI assistant on the government’s online services portal, as “minister” for public procurement, promising that public tenders would be free of corruption [19]. A year later, no evaluation of its results had been published. In December 2025, the director of the agency that developed it, AKSHI, was arrested, and anti-corruption prosecutors are investigating its leadership for rigging tenders in favour of a criminal group [20]. In October 2026, Albania’s Constitutional Court concluded that Diella is not legally a minister and stated that AI cannot create or replace public responsibilities [21]. The problem was not the technology. Diella depended on the same government it was supposed to watch over.
AI as watchdog, not minister
Where AI has actually worked against corruption, the design is the opposite: the machine flags suspicious cases, a person investigates, and whoever controls it is independent of whoever spends the money. In Brazil, the Office of the Comptroller General and the Federal Court of Accounts use Alice, a tool that reviews published tenders every day and alerts auditors. According to the OECD, its alerts led to the suspension or cancellation of tenders worth more than 9.7 billion reais between 2019 and 2022 [22]. Also in Brazil, a crowdfunded group of programmers created Rosie, a program that reviewed the expense claims of members of Congress. It flagged more than 8,000 suspicious claims, and the team reported more than 600 to Congress [23]. In Ukraine, Transparency International monitors the public e-procurement system with automatic risk indicators, such as unusual savings of more than 60% or suppliers that only work with a single buyer, and uses machine learning to decide which tenders to review first [24].
AI can also help people make better collective decisions. Google DeepMind trained the so-called Habermas Machine as a mediator, not a persuader [25]. With 5,734 Britons discussing issues such as Brexit or the minimum wage, most preferred the group statements written by the AI to those written by a human mediator, and the groups ended up less divided, without minority views being left out. There are reasonable criticisms: seeking agreement is not the same as deliberating. Even so, it is the same ability a campaign chatbot uses, put to work finding common ground instead of winning votes.
As my own proposals, with no evidence yet: automatically checking what politicians say against published budgets and voting records, detecting contracts split up to stay under the thresholds that require an open tender, or explaining to any citizen what a law changes. None of these ideas works without open, machine-readable public data, without the tool’s code being auditable, and without an accountable person at the end of the process.
Where this is heading
There are forces pulling in opposite directions. Companies have started measuring the political bias of their models: OpenAI claims that GPT-5 reduced it by 30% and that less than 0.01% of its real answers show it [26]. These are the company’s own figures, without independent verification, and they leave out answers that use web search. At the same time, each government defines what it means by neutral. A US executive order from July 2025 requires that models bought by the federal government not favour “ideological dogmas” such as diversity, equity and inclusion policies [27]. Its supporters see it as correcting a progressive bias; its critics, as swapping one bias for another. China requires the opposite by law. And contamination has not stopped: an analysis from October 2025 was still finding propaganda from Kremlin-aligned media in roughly one in five chatbot answers about the invasion of Ukraine [28].
My conjecture is that the big models will become increasingly cautious with direct political questions, because that is what gets measured and what makes headlines. Bias will shift to what is not measured: which product they recommend, which sources they choose when searching, and what an agent decides when it buys or manages something on your behalf. In campaigns, the obstacle of getting someone to talk to the chatbot shrinks if it does not present itself as propaganda. The European AI Act classifies systems designed to influence voting as high-risk, but those obligations have been postponed until December 2027 [29]. In my reading, a general-purpose assistant that leans one way because its owner has changed its instructions falls into a grey area.
AI is not objective or manipulative by nature. It leans wherever those who train it, tune it, fund it or fill the internet with what it reads point it. Whenever you ask it something, the useful question is not whether it is objective, but who has decided how it answers.
References
[1] Lin, H., Czarnek, G., Lewis, B., White, J. P., Berinsky, A. J., Costello, T., Pennycook, G. and Rand, D. G. (2025). Persuading voters using human–artificial intelligence dialogues. Nature, 648, 394–401. https://doi.org/10.1038/s41586-025-09771-9
Randomised experiments during real elections in the US (n = 2,306), Canada (n = 1,530) and Poland (n = 2,118), plus a ballot measure in Massachusetts; follow-up at about 41 days. Participants knew they were talking to a persuasive AI.
Reliable.
[2] Hackenburg, K., Tappin, B. M., Hewitt, L., Saunders, E., Black, S., Lin, H., Fist, C., Margetts, H., Rand, D. G. and Summerfield, C. (2025). The levers of political persuasion with conversational artificial intelligence. Science, 390(6777). https://doi.org/10.1126/science.aea3884
Three experiments with 76,977 UK participants, 19 models and 707 issues; 466,769 claims fact-checked. Funded by the UK government.
Reliable.
[3] Salvi, F., Horta Ribeiro, M., Gallotti, R. and West, R. (2025). On the conversational persuasiveness of GPT-4. Nature Human Behaviour, 9(8), 1645–1653. https://doi.org/10.1038/s41562-025-02194-6
Pre-registered study with 900 US participants in short, structured debates.
Reliable.
[4] Costello, T. H., Pennycook, G. and Rand, D. G. (2024). Durably reducing conspiracy beliefs through dialogues with AI. Science, 385(6714), eadq1814. https://doi.org/10.1126/science.adq1814
Two experiments with 2,190 conspiracy believers; follow-up at 2 months; professional fact-check of 128 claims.
Reliable.
[5] MacRumors (2025). Grok 4 ‘truth-seeking’ AI consults Musk’s stance on sensitive topics. https://www.macrumors.com/2025/07/11/grok-4-ai-consults-musk-sensitive-topics/
Reports Jeremy Howard’s documentation; behaviour varied between users.
With reservations.
[6] Conger, K. et al. (2025). How Elon Musk is remaking Grok in his image. The New York Times. Interview with the author on WBUR: https://www.wbur.org/hereandnow/2025/10/20/grok-elon-musk-conservative
Journalistic analysis of thousands of answers.
With reservations.
[7] Cyberspace Administration of China and six other agencies (2023). Interim Measures for the Management of Generative Artificial Intelligence Services, in force since 15 August 2023. http://www.cac.gov.cn/2023-07/13/c_1690898327029107.htm
Regulatory framework.
[8] Wang, M., Sadeghi, M. and Lin, C. (2025). Chinese chatbot phenom is a disinformation machine. NewsGuard. https://www.newsguardtech.com/?p=29194
Audit using 15 hoaxes from NewsGuard’s own database. Not peer-reviewed.
With reservations.
[9] NewsGuard (2025). A well-funded Moscow-based global ‘news’ network has infected Western artificial intelligence tools worldwide with Russian propaganda. https://www.newsguardrealitycheck.com/p/a-well-funded-moscow-based-global
Audit of 10 chatbots with 15 false narratives (450 answers). Not peer-reviewed.
With reservations.
[10] OpenAI (2025). Expanding on what we missed with sycophancy. https://openai.com/index/expanding-on-sycophancy/
Analysis published by the company itself.
With reservations.
[11] Warden, P. (2025). Why the Chinese government taught AI to lie. https://petewarden.com/tag/china/
Single test of Qwen 3 run locally offline. One-off demonstration.
With reservations.
[12] Aleph Alpha (2026). Assessment of political bias in Chinese models, reported by The Decoder: https://the-decoder.com/chinese-ai-models-parrot-state-doctrine-or-refuse-to-answer-on-sensitive-topics/
In-house benchmark with 967 topics; models from Alibaba, DeepSeek and Moonshot AI compared with Western models (Claude Sonnet 5, 70% balanced answers; Mistral Small, 92%). Aleph Alpha competes with Chinese models in the European market.
With reservations.
[13] Perplexity AI (2025). Open-sourcing R1 1776. https://perplexity.ai/hub/blog/open-sourcing-r1-1776
Post-training with around 40,000 questions on about 300 censored topics; in-house evaluation.
With reservations.
[14] Uncensored is not unbiased: Abliteration removes refusals but not script-dependent bias about Taiwan (2026). arXiv:2610.29342. https://arxiv.org/abs/2610.29342
Preprint, not peer-reviewed; questions in Traditional Chinese, Simplified Chinese and English.
Recent.
[15] Kumar, A. and Lakkaraju, H. (2024). Manipulating large language models to increase product visibility. arXiv:2404.07981. https://arxiv.org/abs/2404.07981
Fictitious catalogue of coffee makers; preprint; does not measure effects on real consumers.
With reservations.
[16] eWeek (2026). The ads are here: ChatGPT starts serving up sponsored suggestions. https://www.eweek.com/fr/news/openai-tests-ads-in-chatgpt/
The claim that ads do not influence answers comes from the company and has not been externally audited.
With reservations.
[17] Bryant, M. (2025, 5 August). ‘We didn’t vote for ChatGPT’: Swedish PM under fire for using AI in role. The Guardian. https://www.theguardian.com/technology/2025/aug/05/chat-gpt-swedish-pm-ulf-kristersson-under-fire-for-using-ai-in-role
Virginia Dignum’s quote comes from Dagens Nyheter.
Reliable (verifiable news fact).
[18] Associated Press (2023, 30 November). Brazilian city enacts an ordinance that was secretly written by ChatGPT. https://www.click2houston.com/tech/2023/11/30/brazilian-city-enacts-an-ordinance-secretly-written-by-a-surprising-new-staffer-chatgpt/
Reliable (verifiable news fact).
[19] Procurement Magazine, based on Reuters (2025). The new normal? Albania puts AI in charge of procurement. https://procurementmag.com/news/albania-ai-procurement-minister
Reliable as a fact.
[20] Çela, L. (2026, 11 August). Albania’s cheerful ‘AI minister’ is the product of a government agency under investigation for massive corruption. OCCRP. https://www.occrp.org/en/feature/albanias-cheerful-ai-minister-is-the-product-of-a-government-agency-under-investigation-for-massive-corruption
Investigative journalism on an open court case; nobody has been charged; the contract to develop Diella is not among those under investigation.
With reservations.
[21] Albanian Daily News (2026). Albania’s Constitutional Court rejects opposition bid to overturn ‘Diella’ appointment. https://albaniandailynews.com/news/albania-s-constitutional-court-rejects-opposition-bid-to-overturn-diella-appointment
Ruling of 8 October 2026, adopted by majority.
Reliable (verifiable fact).
[22] OECD Observatory of Public Sector Innovation. Robot Alice – Bid, contract and notice analyser. https://oecd-opsi.org/innovations/robot-alice-bid-contract-and-notice-analyser
Data supplied by the Brazilian institution itself; not an independent evaluation.
With reservations.
[23] Data Science Brigade (2017). Brazilian group develops an AI to help in public expenditures monitoring. https://medium.com/data-science-brigade/brazilian-group-develops-an-ai-to-help-in-public-expenditures-monitoring-757900c99552
Data from the project’s own authors; flagged cases are suspicions, not proven irregularities.
With reservations.
[24] Transparency International Ukraine (2017). The new electronic monitoring tool for ProZorro is launched. https://ti-ukraine.org/en/news/the-new-electronic-monitoring-tool-for-prozorro-is-launched/
Description of the risk-indicator system; does not evaluate results.
With reservations.
[25] Tessler, M. H., Bakker, M. A., Jarrett, D., Sheahan, H., Chadwick, M. J., Koster, R., Evans, G., Campbell-Gillingham, L., Collins, T., Parkes, D. C., Botvinick, M. and Summerfield, C. (2024). AI can help humans find common ground in democratic deliberation. Science, 386(6719), eadq2852. https://doi.org/10.1126/science.adq2852
Experiments with 5,734 UK participants, some recruited as a representative sample; compared with human mediators.
Reliable.
[26] OpenAI (2025, 9 October). Defining and evaluating political bias in LLMs. https://openai.com/index/defining-and-evaluating-political-bias-in-llms
In-house evaluation with about 500 questions on 100 topics, scored by an AI model; excludes answers using web search.
With reservations.
[27] Executive Order 14319, Preventing Woke AI in the Federal Government (23 July 2025). https://www.whitehouse.gov/presidential-actions/2025/07/preventing-woke-ai-in-the-federal-government/
Regulatory framework.
[28] The Register (2025, 28 October). Chatbots parrot Putin’s propaganda about the illegal invasion of Ukraine. https://www.theregister.com/2025/10/28/chatbots_still_parrot_russian_state/
News report on an analysis of chatbot answers.
With reservations.
[29] Regulation (EU) 2024/1689 on Artificial Intelligence, Annex III, point 8(b), and the Digital Omnibus on AI (final Council approval on 29 June 2026), which postpones the Annex III obligations to 2 December 2027.
Regulatory framework.