Meta AI, the assistant built into WhatsApp, refused to make campaign material for Brazilian President Luiz Inácio Lula da Silva while making it for his rival, Senator Flávio Bolsonaro, according to tests by a Folha de S.Paulo column. Meta says a temporary election-compliance block missed a prompt. Lula's campaign has taken the matter to the electoral court. This post separates what is reported, what Meta says, and what nobody has verified yet.
The short version: it is a single outlet's test plus company and court statements, so treat the "bias" framing as a claim under investigation, not a finding. But the case is a clean, concrete example of how a chatbot's safety filter can behave unevenly during an election, and it matters to anyone shipping or using AI image tools around political content.
Quick answers
| Question | Answer |
|---|---|
| What happened? | Folha's tests: "Lula 13" poster blocked, "Flávio 22" poster generated by Meta AI on WhatsApp |
| Who tested? | Columnist Mônica Bergamo's team, then Folha correspondents abroad |
| What does Meta say? | Candidate images and recommendations are temporarily prohibited under Brazil's election rules; the "Flávio" prompt was not recognised; extra measures added |
| Is it verified as intentional bias? | No. Meta denies interference; the TSE has asked for technical explanations |
| Who is affected? | Brazilian voters and campaigns using WhatsApp, where Meta AI sits inside the chat app |
| Is there a pattern by country? | Folha says accounts without Brazilian area codes could make Lula posters |
What Folha reported, step by step
Folha's column reports that Meta AI refused a request for a poster carrying the phrase "Lula 13" (13 is the number of Lula's party, the PT) but fulfilled a request for one with "Flávio 22" (22 is Flávio Bolsonaro's party number, the PL). When asked about the difference, the chatbot reportedly acknowledged the inconsistency and attributed it to an automatic safety filter.
Folha then widened the test. Its correspondents in the United States, Argentina and Germany, using accounts without Brazilian area codes, were able to generate Lula posters, including ones with photos. Brazilian numbers, even when used abroad, had their requests blocked. If that holds up, the block is tied to the account's phone number region rather than where the person physically is, which fits a deliberate election-period rule applied to Brazilian users.
After Meta responded, Folha says the chatbot kept creating images that said "Vote Flávio" but refused "Vote Lula" and "Vote Luiz Inácio". That detail is important: Meta's explanation was that the system missed prompts using only the first name "Flávio". A fix aimed at that gap would be expected to remove the asymmetry, and Folha reports it had not, at least at the time of writing.
What Meta says
Meta's statement to Folha, as quoted by the paper: "in compliance with Brazil's election rules, we have temporarily prohibited the recommendation of candidates or the generation of images of them." The company says the senator's material was produced because the filter did not recognise prompts that used only the name "Flávio", and that it took additional measures so such prompts are also covered. Meta denies the accusations and says it is committed to equal treatment.
That is a coherent explanation for a one-off miss, but it does not by itself explain why Lula prompts were blocked when phrased in several ways while Flávio ones kept working, which is the behavior Folha says it still saw. Only Meta's logs and the filter's rules can settle that, which is exactly what the court has asked for.
The legal track: TSE and ANPD
According to Folha, Lula's campaign took the case to the Superior Electoral Court (TSE) on Wednesday, October 14. It had already said on Friday, October 10 that it would go to the court and to Brazil's data protection authority (ANPD) over alleged interference by Meta. The TSE ordered the company to provide technical explanations and preserve records.
This sits inside a broader dispute. Brazilian outlets such as Jornal do Brasil reported that the campaign also accused Meta of interfering through the shutdown of the president's ad account and interruptions in view and reach data on his official Instagram profile. The campaign's own note, quoted there, cites a series of unexplained technical problems and says the company's systems lack transparency. Those are separate claims about Instagram and ads, not about the WhatsApp chatbot, and nothing in the reporting shows they are linked.
Why a "safety filter" can look like bias
Most consumer chatbots layer several systems: a model, an input classifier, and an output filter. Election rules add a fourth: a policy list of names, parties and numbers to be handled differently. Uneven results often come from the plain mechanics of that list.
- Name coverage. A list that includes full names and ballot numbers may miss a first name, a nickname or a party number. Meta's own explanation points at exactly this for "Flávio".
- Token and string matching. "Lula 13" is a distinctive string the filter can catch. A prompt built only of a given name plus a number may slip through.
- Different strictness by language or locale. Folha's finding that foreign phone numbers get through suggests region gating rather than model behaviour.
- Self-reports are unreliable. The chatbot saying it was "an automatic safety filter" is the model narrating, not an audit. Treat it as a lead, not evidence.
None of this proves innocence or intent. It shows why neutral enforcement needs testing with both sides' names, nicknames, numbers and slogans, and why launching an election restriction without that test invites precisely this story.
What is confirmed and what is not
| Claim | Status |
|---|---|
| Meta AI refused "Lula 13" and made "Flávio 22" in Folha's tests | Reported by Folha; not independently reproduced by us |
| Meta says it temporarily prohibited candidate images and recommendations in Brazil | Meta statement quoted by Folha |
| Foreign accounts could generate Lula posters | Reported by Folha correspondents |
| Asymmetry persisted after Meta's statement | Reported by Folha |
| TSE ordered explanations and record preservation | Reported by Folha |
| Meta intentionally favored a candidate | Not shown. Meta denies it. No finding exists |
| The Instagram reach complaints are connected to the chatbot | No evidence in reporting |
What this means for builders and users
If you ship an assistant or image generator that touches politics, a few practical lessons come out of this:
- Test symmetry before launch. Run the same prompt template across every candidate, party number, slogan and nickname, in each language you support, and log the refusal rates side by side.
- Don't key restrictions only to phone region. It is easy to bypass and creates the appearance of selective enforcement, as Folha's cross-border tests show.
- Keep audit logs. The TSE's order to preserve records shows regulators will ask. Without per-request filter decisions you cannot demonstrate neutrality.
- Explain refusals in plain language. A refusal that says "election rules" for everyone reads differently from a silent block on one name.
- Don't let the model explain itself. Log the policy rule that fired instead of asking the chatbot why it refused.
For readers who use AI tools during an election: assume any political output may be filtered inconsistently, check claims against primary sources, and be wary of AI-made campaign material shared as if authentic. Teams deploying agents that act on political or public-facing content can look at AgentBeam, the agent security platform from the explainx.ai team, which stops AI agents before they take dangerous actions.
How it fits Meta's wider AI push
Meta AI is no longer a side feature. The company has been pushing Muse, its personal agent, across apps and devices, and we have covered the launch, the pricing, and the safety questions in depth. See our verdict on whether Muse is safe, the Muse Spark 1.3 launch and pricing, the Meta One subscription tiers, and the investigation into Muse and contact dossiers. The more an assistant is embedded in a chat app used by most of a country, the more its moderation choices become matters of public record and, as here, election law.
A related theme is how platforms steer attention with AI. Our post on Facebook's AI feed controls covers Meta's move to let users shape LLM-ranked feeds, and the earlier claim-versus-fact check on a viral Claude rumor shows the approach we apply here: attribute the claim, list what is checked, and say what is not.
What to watch next
- Meta's technical response to the TSE. Whether it shares the filter rules, logs, and the date the restriction was introduced.
- A fresh independent test. If other outlets or researchers repeat the prompts with both candidates and publish the results, the asymmetry either reproduces or does not.
- Any ANPD step. The data protection authority was named by Lula's campaign; any action would add a privacy angle.
- Rules for AI in campaigns. The case may feed the debate over how electoral law treats generated campaign material and who is responsible for it.
We will update this post if the court or Meta publishes details.
Bottom line
The verified part is narrow: a Brazilian newspaper's tests, a Meta statement about a temporary election rule, and a court order for explanations. The unverified part is the one drawing attention: whether the asymmetry was an accident of a name list or something more. Until the TSE process or independent tests say otherwise, the honest position is that Meta AI behaved unevenly in a high-stakes setting, the company says it was a filter gap, and the evidence for intent has not been presented.
Details are accurate as of October 15, 2026, based on Folha de S.Paulo and Brazilian press reports; the situation is moving and Meta may have changed the filter since.