Brazil's ANPD halts Meta's use of personal data to train generative AI (2024)

The problem

On 26 June 2024 an updated Meta privacy policy took effect that let the company use publicly available information and content shared by users across its platforms (Facebook, Messenger and Instagram) to train and improve its generative-AI systems. Brazil’s data-protection regulator, the Autoridade Nacional de Proteção de Dados (ANPD), opened an inspection on its own initiative and moved against the policy under Brazil’s General Data Protection Law (LGPD) (source).

What was built

No system was built here: this is a regulator-origin AI-governance honest-negative, and the “provider” is the acting authority, the ANPD, not a vendor. The processing under review was Meta’s use of Brazilian users’ public data to train its generative-AI systems, enabled by the 26 June 2024 privacy-policy update applied across Facebook, Messenger and Instagram (source).

The outcome

On 2 July 2024 the ANPD ordered a halt: in its own words it determined the “imediata suspensão, no Brasil, da vigência da nova política de privacidade da empresa Meta” that authorised the use of personal data “para fins de treinamento de sistemas de inteligência artificial (IA)” (source). It set a “multa diária de R$ 50 mil por descumprimento”, a daily fine of R$ 50,000 for non-compliance (source). The Authority acted in the face of the risk of “dano grave e de difícil reparação aos usuários” (grave and hard-to-repair harm to users), and based the measure on preliminary findings of “tratamento de dados pessoais com base em hipótese legal inadequada, falta de transparência, limitação aos direitos dos titulares e riscos para crianças e adolescentes” (source).

Two independent newsrooms carried the order firsthand. Infosecurity Magazine reported that the ANPD “has issued a preventive measure halting Meta’s processing of personal data for the training of artificial intelligence (AI) systems,” a decision “enforceable by a daily fine of R$ 50,000 (roughly US$9,000) for non-compliance,” which “underscores worries about inadequate legal bases for data processing, lack of transparency” (source). TechCrunch reported independently that “the ANPD set a daily fine of 50,000 reais in case of non-compliance,” and that the measure rested on “the imminent risk of serious harm and irreparable or difficult-to-repair damage to the fundamental rights of guardians” (source).

As a practical effect, Meta pulled back: TechCrunch reported that “Meta has suspended the use of its AI assistant after Brazil’s National Data Protection Authority (ANPD) banned the company from training its AI models on personal data from Brazilians,” a decision a Meta spokesperson confirmed (source).

Weakest load-bearing source, named where you meet it: the Preventive Measure and the R$ 50,000/day fine originate with the ANPD’s own release (Tier 1), whose live page is now access-restricted (401), so the primary text is bound not to a live fetch but to the same-day Wayback capture (20240702155331) and byte-tied; the two corroborating reports (Infosecurity, TechCrunch) are Tier-2 independent newsrooms. This is a Brazilian LGPD action by the ANPD, distinct from the separate June-2024 pause of Meta’s EU AI training after the Irish DPC, which is not relied on here. The later disposition of the Preventive Measure is not asserted here as a cited fact; it is carried as a confirmation item under Path to green below.

Path to green

The honest-negative event, the 2 July 2024 Preventive Measure ordering Meta to halt generative-AI training on Brazilians’ data under a R$ 50,000/day fine, rests on the ANPD’s own published release (Tier 1) plus two independent newsrooms. This is a regulator-origin record: green never depends on Meta or the ANPD confirming anything, and the bar rises only by demanding more of the independent public record. The remaining items are human, post-order confirmations held off the page: a recorded confirmation from the ANPD (or Meta) of the order’s disposition, including whether, and on what terms, the Preventive Measure was lifted (independent press reported a late-August 2024 lifting after the ANPD approved a Meta compliance plan; not byte-tied here), and of the scope and dates over which Meta’s generative-AI training/tools were suspended in Brazil, captured on the record. No green badge is sought; verified is human-only.

How this was verified

• Method: Every quoted finding was read verbatim from the captures on file in sources/captures/ and re-checked against them this session (whitespace-normalised match, decode utf-8/replace for the Portuguese primary). No number or holding is paraphrased, rounded, or recalled.
• Primary anchor: The ANPD’s own release, “ANPD determina suspensão cautelar do tratamento de dados pessoais para treinamento da IA da Meta” (2 July 2024, Tier 1, Portuguese). The live gov.br page is now access-restricted (401); the primary text is bound to the same-day Wayback capture 20240702155331 and byte-tied.
• Standard: Regulator-origin. The outcome is the adjudicated Preventive Measure, so the public record, not any subject’s confirmation, is the proof.
• Independence: The Preventive Measure and the R$ 50,000/day fine are each carried firsthand by two mutually independent Tier-2 newsrooms (Infosecurity Magazine, 4 July 2024; TechCrunch, 18 July 2024), neither relaying the other and neither an ANPD reprint or Meta PR; the ANPD’s release is the origin and is not counted toward corroboration.
• Conflation guard: Severed from the separate June-2024 Irish-DPC EU AI-training pause (distinct regulator, distinct EU/GDPR action). The ~102M-users figure is ANPD scale context, not a measured harm.
• Open item: The final disposition of the Preventive Measure (whether/when/on what terms it was lifted) and the exact scope/dates of Meta’s suspension in Brazil.
• Verified on: 2026-08-10 (presentation rebuilt 2026-08-19).

Related case files

• Meta pauses EU AI training on public data after the Irish DPC (2024): the direct sibling, the same company pulling the same generative-AI training practice, but under a different regulator (Ireland’s DPC / GDPR) weeks earlier, so the two read together as one global retreat forced by two data-protection authorities.
• Spain’s AEPD orders Worldcoin (Tools for Humanity) to halt biometric iris scanning: another DPA using a preventive/precautionary halt against an AI-data practice, a close procedural analogue to the ANPD’s own-initiative Preventive Measure.
• The Dutch DPA fines the Tax Authority €2.75M for unlawful algorithmic risk profiling: a DPA enforcing against an automated/algorithmic data practice, the enforcement end of the same regulatory posture the ANPD opens here.
• Hungary’s NAIH fines Budapest Bank HUF 250M over unlawful AI emotion/voice analysis: a European DPA penalising an AI system on the same LGPD-style grounds (no adequate legal basis, lack of transparency) the ANPD cites against Meta.

Sources

1. [Tier 1] Autoridade Nacional de Proteção de Dados (ANPD) · “ANPD determina suspensão cautelar do tratamento de dados pessoais para treinamento da IA da Meta” · 2 July 2024 · https://www.gov.br/anpd/pt-br/assuntos/noticias/anpd-determina-suspensao-cautelar-do-tratamento-de-dados-pessoais-para-treinamento-da-ia-da-meta (live page access-restricted; bound to Wayback capture https://web.archive.org/web/20240702155331/ )
2. [Tier 2] Infosecurity Magazine · “Meta Faces Suspension of AI Data Training in Brazil” · 4 July 2024 · https://www.infosecurity-magazine.com/news/meta-suspension-ai-data-training/
3. [Tier 2] TechCrunch · “Meta puts a halt to training its generative AI tools in Brazil” · 18 July 2024 · https://techcrunch.com/2024/07/18/meta-suspends-generative-ai-tools-in-brazil/