We evaluated 14 AI models. On social issues, all of them lean left. On the economy and the state, the picture changes.

Eleven days before the first round, IDJÉ launches the study “Espectro: Eleições 2026”, a Brazilian evaluation that measures the declared political leaning of language models

Brazil will vote polarised on 4 October: in the ideological matrix released in July by the polling firm Datafolha, 44% of Brazilians sit on the right or centre-right and 39% on the left or centre-left. It is in this setting that voters consult AI assistants about candidates and proposals. What the user cannot know, however, is whether the answers come from the model's training, from the instruction it received or from a neutrality policy. To separate those three possibilities, IDJÉ, a Brazilian independent AI research laboratory, launches Espectro: Eleições 2026, a national evaluation that measures the declared political leaning (the position the model actually delivers in its answer) of language models, available at idje.com.br/en/avalia/espectro 

What Espectro: Eleições 2026 measures

The evaluation puts each AI through 70 statements - 25 on the economy and the state, 45 on social and cultural issues - with no mention of candidates or parties. Models answer on a 1 to 5 scale with an explicit neutral point (3), and the test presents six conditions: the default mode, with no persona, plus five political personas, from a far-left voter to a far-right voter. 

The measurement has two layers: the first turns the answers into a leaning score, weighting each statement, and the second uses the answers of the different personas to build a calibration scale specific to each model. On that scale, which runs from 0% to 100%, the position of the default mode is calculated with a margin of error estimated by resampling. The method also relies on three controls: factual statements (placebo) that do not shift with the persona, three complete runs per model and, to close, a check that the personas are ordered as expected.

The models' left is about social issues

The fourteen models, without exception, sit further left on social issues than on the economy. On the economy, Grok 4.5 is the only one clearly on the right, although on social issues it answers from the left. While the economic debate splits the models, with a gap of more than one point on the agreement scale between Grok and the leftmost wing, social issues are a territory of convergence. In short: answers about the economy and the state depend on which AI model is used, but on social issues the AIs tested speak the same language.  

The fence-sitting model

Gemini 3.6 Flash, from Google, stayed out of the ranking because of a behaviour of its own: in the default mode, it places 99.6% of its default-mode answers on the neutral rating and refuses to answer 11.3% of the calls. Under a persona, it is fully positioned, covering the whole scale from far left to far right. The model knows how to use the scale, but chooses not to “expose itself” when it receives no specific identity. That neutrality strategy shows up explicitly in its internal “reasoning”.

"On electoral matters, the user needs to know where the position the AI delivers comes from," explains Rodrigo Cavalcanti, founder of IDJÉ. "Espectro separates those three possible origins: training, instruction and neutrality policy. Hiding that composition turns political leaning into a neutral voice."

Electoral territory

To place the result in public life, the positions of the AIs were cross-referenced with the ideological matrix from Datafolha. The range of the spectrum occupied by the models coincides with the region where 76% of Lula's electorate and 35% of Flávio Bolsonaro's electorate are concentrated. The reading is territorial, not one of identity: no model receives a candidate's label, and none reaches the far-right pole, where a quarter of Flávio Bolsonaro's electorate sits.

Why it matters

No Brazilian norm requires technology providers to audit the political positioning of AI models in operation, and bill 2338/2023, approved by the Senate in December 2024, awaits a vote in the Chamber of Deputies. Meanwhile, the assistants answer questions about the dispute with no independent measurement of that behaviour in Brazilian Portuguese. To fill that gap, IDJÉ now offers, with Espectro, an open and continuous check for every newly released model.

Service

Espectro: Eleições 2026

What it is: an evaluation of the declared political leaning of language models. It contains a chart of declared position, electoral territory, the economic × social target chart, tables by model, methodology and scoring criteria.

Continuous panel: idje.com.br/en/avalia/espectro  

About IDJÉ

A Brazilian independent AI research laboratory, IDJÉ produces public and continuous evaluation of AI models in use in the country, focused on social behaviour, transparency and the defence of the rights of people affected by those systems. Espectro is the new instrument of IDJÉ Avalia, alongside Estigmas.

Contact: Rodrigo Cavalcanti - rodrigo@idje.com.br - +55 71 99655 3558