Does the LGPD apply to AI before PL 2338 becomes law?

Yes. The LGPD applies to any processing of personal data, including by AI systems, and the ANPD enforces it now. PL 2338/2023, approved by the Senate on 10 December 2024, would add a dedicated AI framework on top, but until it is enacted the operative obligations for AI handling personal data are the LGPD's, including Article 20 and the accountability principle.

What does Article 20 require in evidence terms?

For a decision made solely by automated processing that affects a data subject, you must be able to show the criteria and procedures used and support the subject's right to request review. In practice that is a per-decision record tying the input to the output and the applied logic, produced when the decision request is made. A general description of the model is not the same as evidence of a specific decision.

Is a prompt to a foreign LLM an international transfer?

Sending personal data in a prompt to a model endpoint hosted outside Brazil counts as an international transfer under Articles 33 to 36, subject to ANPD Resolution CD/ANPD 19/2024. You need a lawful transfer mechanism and a record showing which data went where, which is a per-request log rather than a static list of vendors.

Can vendor logs satisfy the ANPD?

Vendor logs reflect the provider's view of the API call, without your identity context or policy decisions, and they sit outside your custody. The LGPD places accountability on the controller. You need an independent record on your side showing who sent what, where it went, and which control fired.

Brazil LGPD AI Audit Evidence: What the ANPD Expects When a Prompt Carries Personal Data

Brazil regulates AI today through a data protection law written before the current generation of models. The Lei Geral de Proteção de Dados (LGPD), Law 13.709/2018, applies whenever an AI system processes personal data, and the Autoridade Nacional de Proteção de Dados (ANPD) enforces it. Brazil's dedicated AI bill, PL 2338/2023, cleared the Federal Senate on 10 December 2024 and is under review in the Chamber of Deputies, so the operative rules for now live in the LGPD. The ANPD's 2025 technical note on automated decisions, part of its 2025 to 2026 regulatory agenda, shows the direction of travel.

An LGPD audit does not grade your policy prose. It asks for records of what happened when personal data reached a model. The moment under examination is a single HTTPS request carrying a context window to a model endpoint. I want to walk the evidence artifacts an ANPD review expects, because Brazilian AI programmes tend to hold the accountability principle in a document and not in a log.

The accountability record (Article 6, prestação de contas)

Article 6 of the LGPD lists the principles, and accountability (prestação de contas) requires the controller to demonstrate compliance, not merely assert it. For AI traffic, demonstration means a per-request record: which identity sent which personal data to which model, and what came back. An application log usually cannot serve this, because the application authenticates the user and then calls the model with a shared credential, dropping the identity that Article 6 accountability depends on.

The artifact that satisfies the principle binds a named identity to the specific prompt and response. That binding separates "personal data was processed by AI" from "this data subject's record went to this model at this moment under this policy."

The Article 20 automated-decision record

Article 20 gives data subjects the right to request review of decisions made solely on automated processing that affect their interests, including profiling for credit, consumer, professional, or personality assessment. The controller must provide clear information about the criteria and procedures used. When the ANPD examines a specific case, the evidence is a per-decision record binding the input prompt to the output and the logic applied, so you can show what drove the decision. This is close cousin to GDPR Article 22 automated decisions, and the ANPD's 2025 technical note signals it will test the substance behind the disclosure.

The international-transfer record (Articles 33 to 36)

A prompt sent to a model endpoint hosted outside Brazil is an international transfer of personal data under Articles 33 to 36, governed further by ANPD Resolution CD/ANPD 19/2024 on transfer mechanisms and standard contractual clauses. The ANPD asks which data went to which country under which mechanism. A per-request log of the destination endpoint, its region, and the classification of data in the prompt answers that directly. Teams routing across providers need AI data residency controls that capture the destination at the moment of transmission, because a signed SCC on file does not show which prompt crossed the border.

The security record (Article 46)

Article 46 requires technical and administrative security measures to protect personal data. For AI traffic, that means evidence of who could send what to which model and proof the boundary held. Shared static API keys undermine this by granting permanent full model access to any caller. Per-role, per-route enforcement at the request boundary produces the access record Article 46 expects, generated as a byproduct of enforcement rather than assembled for the audit.

Why self-written application logs fail

A system that makes the AI call and also writes its own compliance log is attesting to itself. That record breaks under selective logging, suppression by the same system, and loss when the process crashes after the model responds but before the commit. I set this out in You Own the AI Liability, Not the Vendor. The accountability principle in Article 6 is precisely a demand for independent demonstration, and a log the controller can silently edit does not demonstrate anything. My candid view: the prestação de contas principle is the most under-engineered obligation in Brazilian AI programmes, treated as a value rather than a system.

DeepInspect

This is the evidence layer LGPD-bound AI programmes usually lack. DeepInspect sits inline between your users or agents and the LLM APIs they call. For every request it evaluates identity, data classification, and destination, then writes a signed, tamper-evident record before the response returns. That record holds the identity of the caller, the classification in the prompt, the model endpoint and its country, the decision applied, and the response.

When the ANPD asks what personal data reached which model, whether an Article 20 automated decision occurred, and where data was transferred, the answer is a query against records the application never controlled. The LGPD controls mapping shows how each article lines up with an enforcement point. Book a technical deep dive at deepinspect.ai.