概述
參議院於 2024 年 12 月批准,並將其移交給眾議院進一步辯論。這很重要,因為巴西是拉丁美洲最大的數位市場,該法案還處理 AI 訓練的版權支付問題,這是少數直接處理此問題的法律之一。
深入探討
Senate President Rodrigo Pacheco introduced PL 2338/2023 in May 2023, based on a draft prepared by a commission of legal experts after earlier, thinner bills were criticized. After a temporary Senate committee reworked it extensively, the full Senate approved a substitute text on December 10, 2024. Because Brazil's Congress is bicameral, the Chamber of Deputies must also approve it, and any changes there send it back to the Senate; it is not yet law. The bill sorts AI by risk. 'Excessive risk' uses are prohibited, including techniques that manipulate people into harmful behavior, exploiting vulnerabilities of groups such as children, government social scoring, systems used to produce or spread child sexual abuse material, and most real-time remote biometric identification in public spaces, subject to listed exceptions such as searching for missing persons. High-risk uses are listed by purpose, such as critical infrastructure, education and admissions, recruitment and worker evaluation, access to essential public and private services, criminal investigation, border control, health and autonomous vehicles. During negotiations some items, such as social media recommendation systems, were removed from the high-risk list. High-risk operators must document systems, test for bias, ensure human oversight and carry out algorithmic impact assessments. Affected people gain rights to information, explanation, contestation, human review and non-discrimination. The bill builds on the LGPD, Brazil's 2018 general data protection law, whose Article 20 already lets people request review of automated decisions. It proposes a National System for AI Regulation and Governance (SIA) coordinated by the data protection authority, ANPD, working with sector regulators. A misconception is that it is a copy of the EU AI Act; it shares the structure but places heavier emphasis on individual rights and includes copyright remuneration rules.
現實世界的實施
一家使用 AI 來評估求職者的巴西公司將其視為高風險,進行算法影響評估,並允許被拒絕的候選人要求對決策進行人工審查。
一個州政府計畫建立一套系統來評分公民的社會行為以決定是否獲得公共服務將被阻止,因為政府社會評分屬於過度風險類別。
一個生成式 AI 開發者在訓練時使用巴西新聞文章,必須披露其使用了哪些受保護作品,並尊重選擇退出的權利持有人,對於根據法案版權章節使用作品應支付報酬。
被自動化系統拒絕公共福利的人可以主張權利,要求被告知已使用 AI,獲得解釋並對結果提出異議。
巴西 AI 法案(PL 2338/2023)的未來
該法案的最終形態取決於眾議院,該院設有一個特別委員會在技術公司、創意產業和公民社會的游說下審查它。版權報酬條款和高風險清單的範圍是最具爭議的點,兩者都可能改變。若通過,該法案可能會逐步實施,並依賴 ANPD 和行業監管機構發布詳細規則。其他拉丁美洲國家也在起草自己的 AI 法案,因此巴西的結果可能會影響區域性做法。
付諸實踐
為什麼這很重要
- 風險與安全. 災難性和日常的人工智慧危害都取決於誰了解風險以及誰能夠採取行動。
- 更明確的決策. 民眾和專業素養決定強而有力的安全政策在政治上是否可行。
- 突破炒作. 清晰的解釋可以減少炒作、實驗室公關和模糊道德劇場的影響。
留意
- 將存在風險視為科幻小說,同時能力複合。
- 混淆了表面產品安全與高度自治下的對準。
- 只給非英語和非專業觀眾留下低品質的資源。
後續步驟
- 單獨的產品危害、誤用和失控/失調風險。
- 詢問哪些證據會改變您對時間表和嚴重性的看法。
- 比起行銷主張,更喜歡主要來源和具體評估。
- 確定一條行動路徑:職業、政策、資金或技能——而不僅僅是意識。
不斷探索
常見問題
What is Brazil's AI Bill (PL 2338/2023)?
PL 2338/2023 is Brazil's proposed comprehensive AI law, a risk-based framework that bans 'excessive risk' AI, imposes duties on high-risk systems and gives affected people rights such as explanation and human review. The Senate approved it in December 2024 and it moved to the Chamber of Deputies for further debate. It matters because Brazil is Latin America's largest digital market and the bill also tackles copyright payment for AI training, an issue few laws address directly.
What is the legislative status of PL 2338/2023 as described in this guide?
The Senate passed a substitute text on December 10, 2024; the Chamber of Deputies must still approve it.
Which use falls under 'excessive risk' and is prohibited?
Social scoring by government is among the prohibited practices, alongside manipulation and exploitation of vulnerabilities.
Which was removed from the high-risk list during Senate negotiations?
Recommendation algorithms of social platforms were dropped from the high-risk list in negotiations.
Which right does the bill give people affected by AI decisions?
Rights include information, explanation, contestation, human review and non-discrimination.
What copyright obligation does the bill propose for AI developers?
Developers would disclose protected content used, respect opt-outs and owe remuneration, with some exceptions for research.