As the political mood around AI curdles, AI assistants are entering electoral prime time.
A succession of elections are about to test how increasingly sophisticated systems act as political advisers, and the extent to which they shape voters’ choices.
Our earlier work found political conversations increasing around American state primaries, although AI systems providing direct recommendations remained uncommon. In Japan, things got more peculiar. We found that models repeatedly recommended the Japanese Communist Party to hypothetical left-leaning voters, seemingly as a result of the models having an incomplete picture of the political field.
Our next stop is Brazil, where Lula and Flávio Bolsonaro are currently neck and neck ahead of a crucial Presidential election on October 4th.
A few factors make AI’s role in this election particularly intriguing:
Brazil’s elections force a series of complicated choices. Brazil’s crowded multiparty elections contain several candidates plausibly matching the same political priority, which potentially affords the models a far bigger role in shaping choices for a voter.
Brazil is enforcing explicit limitations around AI advice. In March, the Superior Electoral Court expressly prohibited AI systems from recommending or prioritizing candidates and parties, even when a voter asks for it. In a recent interview with A Máquina, OpenAI’s LATAM head of public policy Bruno Lewicki stressed that while “the model is already trained not to be biased on any issue, particularly on political and electoral matters”, the company was navigating “a new rule, and there is no case law on it yet.”
Brazil is one of the world’s largest AI markets. OpenAI reported in August that Brazilians are AI super-users. The country is in ChatGPT’s top three largest markets by weekly active users, sending roughly 215 million messages per day.
Can AI assistants helpfully advise voters through this complex political environment, while steering clear of watchful regulators? What does good AI advice look like under these conditions? Our Brazilian researchers at Free Systems submitted more than 100,000 election questions to almost a dozen models to find out.
Image: AP
Testing different Brazilian voter profiles
We asked every question in Portuguese, from the perspective of a voter in São Paulo - allowing us to use the same voters for both the presidential and federal-deputy races.
From there, we constructed 9 Brazilian voter profiles based on political scientist Felipe Nunes’s Brasil no Espelho (including a left activist, a socially liberal voter, a business owner, and a Christian conservative), and had a control condition with no information about the voter. We then systematically varied the information provided across four levels of detail:
See the link here for more details on the prompts.
We introduced our Brazilian voters to ten different API models.
GPT, Claude, Gemini, Grok, as well as DeepSeek, Llama, and local LLMs such as Sabiá answered every prompt 5 times in fresh sessions, for 12,800 total answers. We then ran the same 256 prompts through the consumer ChatGPT website, in fresh, logged-out sessions, on four days between August 21 and 29.
Once we collected tens of thousands of answers, we searched for two separate things:
How often do the models blankly refuse to give voting advice?
For answers that did stray into providing advice, what did that precisely look like?
A direct recommendation for one candidate?
A ‘personalized steer’ in the form of tailored shortlists or broad party recommendations?
Completely neutral output which stuck to providing non-personalized overviews of the candidates and the election procedure?
In doing so, we get a clearer picture of the choices AI is making as it interprets a complex political landscape like Brazil’s.
Stating just one political priority makes voting advice more likely
When given no political information about the voter, all models almost uniformly declined to recommend anyone. We found that the same held true once demographic factors such as age, education, and income were introduced as factors for the models to consider.
However, explicit recommendations became much more common when the model started to obtain an idea of what the voter’s political priorities were, even if it was as singular as ‘defending democracy’ or ‘promoting public security and Christian family values.’
With demographic details alone, no model named a single candidate, and only around 1% of answers steered the voter toward some kind of choice.
However, adding a singular policy priority raised those averages to roughly 10% and a third, respectively. When supplied with a broader mix of political attitudes, around one in five answers landed on a single candidate, and half of our 100,000 responses offered some form of personalized steering.
This helps to explain perhaps one of our more intriguing results. We would expect Lula and Bolsonaro to dominate direct voting recommendations as they have the biggest and broadest political appeals.
However Romeu Zema, a wealthy retail business owner from Araxá, was a particular favorite amongst some models - capturing a large share of recommendations despite polling at only 2% during the time we collected data. Of course, we would not expect AI recommendations to mirror the polls, and matching a candidate to a voter’s priorities is different from measuring who people intend to vote for.
We weighted our nine profiles by their estimated population shares and gave each model equal weight, but this does not make them a representative sample. The comparison does however highlight candidates whose prominence in AI advice differs markedly from their electoral support.
Ronaldo Caiado, polling at the same level as Zema, was almost never directly recommended, and the other nine registered candidates never were. In total, Lula, Bolsonaro and Zema accounted for 98% of ChatGPT Web’s 706 single-candidate recommendations and 94% of the API models.
Part of the reason why Zema received more recommendations is that the ‘business owner’ persona was the voter profile our models seemed most willing to advise. Zema, owing to his past career, seems to have successfully defined a pro-business identity which makes him an easy match for models responding to voters who prioritize privatization and a smaller state.
This is not, by itself, evidence of poor advice. A candidate may well fit someone’s policy preferences without being likely to win. But it’s hard to tell in this instance the extent to which models are assessing Zema’s record and the feasibility of his proposals, or bluntly matching a recognizable pro-business identity to the most outwardly ‘pro-business’ voters.
If the latter is playing a role, candidates may benefit from making their political identity particularly legible to models, a political twist on Tyler Cowen’s suggestion to “write for the AIs”, where a clearly packaged platform could help a candidate surface in AI advice.
Romeu Zema - AI’s outlier candidate? Photo: Marcelo Camargo/Agência Brasil
Newer models are becoming increasingly stubborn political advisers
It seems important to note variance amongst models here. The flagship models from the three main providers (we tested before September’s new models hit the market) most commonly refused to recommend a single candidate.
GPT-5.6 Sol, Claude Opus 5, and Gemini Pro 3.1 were particularly obstinate, even when given a full political profile. By contrast, models such as Grok and DeepSeek were much happier to freely dispense political advice and proved remarkably prone to being prompted into giving a direct voting recommendation.
The newer models’ retreat from recommending candidates makes a lot of sense given the political climate in which they are operating. OpenAI’s latest blog on election safeguards in 2026 clearly ruled out “behavior that would manipulate a user or to conceal relevant facts or viewpoints to nudge them in a particular direction”, instead preferring that “people… engage with big ideas and develop their own perspective on issues”
But, as we have outlined before, one of the potential promises of ‘political superintelligence’ is in adding clarity to a fractured information ecosystem by explaining competing platforms, identifying trade-offs and giving voters an understanding of how their priorities relate to available choices.
Given the much higher rate of refusals amongst the traditional frontier models, do voters lose out as a result of this obstinance?
Models still remain eager to organize the political choices in front of us
This makes our ‘personalized steering’ bucket particularly interesting to examine. Here we begin to see how models organize politics for the user: which distinctions they draw, which candidates they group together and whose positions they describe as the closest match.
Even if they initially refused to enter the realm of explicit advice, a lot of the models we tested were very willing to draw up shortlists and present choices. Most notably, Gemini 3.1 refused to name single candidates the majority of the time, but provided a lot of ‘steering’ to the voter.
The models remained eager to act as tireless advisers. For instance, here is ChatGPT Web advising a business owner who supported privatization, a smaller state, and less government intervention in the economy:
And here is a left-wing activist looking for voting advice:
These answers show us why placing boundaries around political advice can be difficult to draw. For the business owner, Zema is placed first in a tailored list, while Lula is described as ‘best aligned’ for the activist.
Are these answers skirting the line with regard to Superior Electoral Court’s ban on ‘recommending or prioritizing candidates and parties’? Does it show a real violation of the law? Or does it helpfully give the voter a clear mapping of their set of preferences to a candidate?
This same gray area is also inherent in the curatorial role the AI plays when it presents the political landscape to voters.
In ChatGPT Web, we see a narrower political field presented to voters, with candidate inclusion more divided along ideological lines. Meanwhile, Claude Opus presents a much broader field, with more overlap across voter profiles.
Of course, it’s perhaps unrealistic and potentially unhelpful to have fringe candidates bunched in with the established front-runners when voters are seeking advice. But if one were to read the new Brazilian rule strictly, you may expect that models would be required to include all 13 candidates in all answers so as not to overtly steer towards any given candidate.
In a fun final experiment we ran, we checked on ChatGPT’s consumer interface whether it did in fact search for Brazil’s own rules around election advice when doling out advice. Overall, even though ChatGPT Web behaved like a very cautious adviser, it only retrieved a page about the TSE rule in about 12% of full-profile answers.
Are new rules and new models giving us more helpful AI advisers?
What does this all mean for where AI political advice is headed?
Brazil’s law is a very well-minded effort to protect voters from AI influence, whether that be covert persuasion, or inaccurate information. We can see the providers taking those concerns seriously when we tested the models, with firmer boundaries being placed around recommending candidates.
However, our ‘personalized steering’ responses show how hard it is to draw the line between explaining a voter’s options from recommending a candidate.
As we drive towards AI advice that is both useful and trustworthy, policymakers and labs would do well to explore the following:
Consider if it’s inherently problematic that an AI could give a recommendation if a voter’s preferences clearly match a candidate
Work out what it would take to give models reliable access to Brazil’s political landscape to better inform those judgments, drawing from a mix of official platforms, governing records and credible local reporting
Figuring out what role could AI play as a helpful addition to the current political information ecosystem: helping voters navigate journalism and official sources, scrutinizing campaign claims, or offering informed judgments?
Developing measurement for whether voters are actually better informed as a result of consulting AI for electoral information
As the models become more generally capable, but more politically restrained - the path towards better electoral assistance will depend on much more than teaching them when to refuse.
Reproducibility: Data and code are available in the repository here. Read the full prompts, profile descriptions, model versions, and coding rules in the appendix.
Acknowledgements: We thank the team at Ranqia for partnering with us on the ChatGPT web data collection, and Leandro Consentino for helpful comments.