ANPD Builds Internal AI Centre and Prepares Second Sandbox
Brazil’s National Data Protection Authority is standing up an internal artificial-intelligence centre inside its Technological Innovation Superintendence and preparing a second regulatory sandbox focused on Digital ECA challenges, Convergência Digital reported from the agency’s September privacy seminar. Superintendent Lucas Anjos said the centre aims to build infrastructure and specialist staff so ANPD can analyse generative AI, automated decisions and high-volume personal-data systems without waiting solely for post-hoc rules, while the first sandbox—testing LGPD Article 20 transparency and explainability with medical, transport and financial-security firms—nears end-October results.
Filed under Policy and dated September 30, 2026, this AI4Brazil briefing treats the ANPD centre and sandbox track as Brazilian AI-governance capacity news distinct from yesterday’s Itaú–Google Batalha de Agentes. The agency already monitors generative training on personal data and non-consensual sexualised image models, has commissioned UNDP studies on manipulative design, child profiling, parental controls, AI agents and platform labour data, and wants 2027 sandbox evidence to feed both rulemaking and enforcement—alongside Sedigi’s earlier companion-app risk referral.
Why it matters: Brazilian firms already ship chatbots and scoring tools under LGPD. An authority that can test systems in sandboxes can raise evidence quality—but only if participation rules, publication of findings and human appeal paths stay clear.
What it means in practice
Brazilian privacy, product and counsel leads should inventory which automated-decision tools may fit Article 20 sandbox lessons; demand named owners for end-October first-sandbox readouts; assign an owner for Digital ECA sandbox eligibility if child-facing; run time-boxed explainability reviews on credit or hiring models; and prefer designs that keep humans on significant decisions. Connect the build-out to Sedigi’s AI-companion referral to ANPD and TSE’s chatbot behaviour note.
Caveats come first. A centre in formation is not a finished AI Act-style regulator; October results can disappoint; and UNDP studies are not sanctions. AI4Brazil therefore presents the ANPD plan as directional oversight context until published sandbox reports and centre staffing appear.
What to watch next: first-sandbox public findings; second-sandbox criteria; and whether Congress AI bills name ANPD technical roles. Readers can continue on the AI4Brazil homepage, or browse the Newsroom for additional briefings.
Bottom line: treat this update as orientation, not instruction. Brazilian AI oversight is building in-house technical muscle and remains early. Organizations that benefit most will prepare explainability packs now, keep humans on high-impact decisions, and refuse to confuse a seminar reveal with finished enforcement.