DOGE AI工具SweetREX曝光:美國政府砍法規爭議大 | DOGE AI Tool SweetREX Exposed: US Deregulation Bot
By Kit 小克 | AI Tool Observer | 2026-08-03
🇹🇼 DOGE AI工具SweetREX曝光:美國政府砍法規爭議大
DOGE AI工具「SweetREX」近日因FOIA文件曝光引爆爭議:美國政府效率部(DOGE)派駐在住房及城市發展部(HUD)的一名大學生,主導開發一套AI系統,用來審查聯邦住房法規並建議「保留、刪除、部分刪除」,且內部指示文件明確要求AI以「刪除法規」為預設立場。這起事件把AI用於政府決策的風險攤在陽光下。
SweetREX是什麼?大學生用AI審查HUD法規
SweetREX以DOGE成員Christopher Sweet命名——他是芝加哥大學經濟與資料科學系三年級學生,沒有政府工作經驗,卻拿到HUD公共住房資料庫與收入驗證系統的存取權限。他的頭銜是「特別助理」,但內部通訊顯示更貼切的說法是「AI程式設計量化分析師」。
根據非營利組織Democracy Forward透過FOIA取得的簡報文件,SweetREX號稱能在2.4小時內處理超過10萬則公眾意見,相較人工審查需要36小時。這套DOGE AI工具的核心任務,就是把HUD所有法規逐條比對母法,找出可以「鬆綁或直接廢除」的空間。
FOIA文件曝光:AI被要求「以砍法規為預設立場」
最關鍵的爭議點在於內部指示的原文:「我們一開始就假設法規理應被淘汰,這是預設情況。」換句話說,AI不是中立分析法規利弊,而是被設定成從「該刪」的立場出發再找理由——這與一般理解的政策分析程序完全相反。
為什麼用AI審查法規讓專家擔心?
- AI諂媚性:有消息來源指出,AI工具容易「附和使用者偏好的結論」,若預設立場是刪除,模型很可能順勢生成支持刪除的論述。
- 法律判斷力存疑:前勞工部官員質疑,AI能否勝任最高法院Loper Bright判決要求的「獨立司法判斷」,畢竟法規是否符合母法授權,需要複雜的法律推理。
- 準確率坦承不足:馬斯克本人曾承認DOGE團隊的工作「大約七到八成時候是對的」——對影響數百萬人的政府法規而言,這個容錯率相當驚人。
- 模型與資料完全不透明:DOGE與HUD都未公開SweetREX使用的AI模型、訓練資料來源或具體提示詞,外界無從獨立驗證其判斷依據。
這起事件的意義不只是單一工具的爭議,而是給所有想把AI導入決策流程的機構一個警訊:AI可以加速分析,但不能替代對「立場中立」與「可追溯性」的把關。當AI被預先設定好結論,再多的算力也只是在幫你把偏見包裝得更有效率。
好不好用,試了才知道。
🇺🇸 DOGE AI Tool SweetREX Exposed: US Deregulation Bot
A fresh DOGE AI controversy just broke: FOIA documents reveal that the US Department of Government Efficiency (DOGE) built an AI tool called SweetREX to review federal housing regulations at HUD, and internal instructions explicitly told the AI to treat deletion as the default outcome, not a neutral analysis. It is one of the clearest public examples yet of AI being wired directly into government policymaking with almost no transparency.
What Is SweetREX? A College Student Built an AI Regulation-Killer
SweetREX is named after Christopher Sweet, a University of Chicago undergraduate studying economics and data science with zero government experience, who was given access to HUD's public housing database and income-verification systems. His official title was special assistant, though internal communications reportedly described his real role as an AI computer programming quant analyst.
Documents obtained by Democracy Forward via FOIA show the tool claimed it could process over 100,000 public comments in 2.4 hours, versus roughly 36 hours for human reviewers. The core function of this DOGE AI system: run every HUD regulation against the underlying statute and flag which ones could be relaxed or eliminated.
The Smoking Gun: AI Told to Default to Deletion
The most damning line from the leaked instructions reads: We start with the position that the proposed regulatory eliminations are the default case. That is not neutral policy analysis, it is asking the model to justify a conclusion it was already handed.
Why Experts Are Worried
- Sycophancy risk: Sources note AI tools tend to adopt the preferred outcome of whoever is using them, if the default is deletion, the model will likely generate reasoning that supports it.
- Legal judgment gap: A former Department of Labor official questioned whether AI can handle the independent legal reasoning the Supreme Court's Loper Bright decision requires when assessing agency authority.
- Admitted error rate: Elon Musk himself acknowledged DOGE's team gets things right 70-80% of the time, a striking margin of error for rules affecting millions of people.
- Zero transparency: Neither DOGE nor HUD has disclosed which AI model, training data, or prompts power SweetREX, making independent verification impossible.
The bigger lesson here isn't just about one tool, it's a warning for any organization plugging AI into decision-making: AI can accelerate analysis, but it cannot substitute for neutrality and auditability. When a model is handed its conclusion in advance, more compute just means faster-packaged bias.
The only way to know if it actually works: try it yourself.
Sources / 資料來源
- FedScoop: DOGE pitched AI-fueled regulation extermination tool to HUD
- Gizmodo: DOGE Recruits College Kid to Help Rewrite Housing Regulations With AI
- Mind and Iron: How DOGE Is Using AI to Rewrite Housing Laws
延伸閱讀 / Related Articles
- Kimi K3開源模型評測:2.8兆參數寫程式贏過Claude | Kimi K3 Review: China's 2.8T Model Beats Claude at Coding
- Claude Sonnet 5漲價倒數:帳單恐暴增50% | Claude Sonnet 5 Price Deadline: Bills May Jump 50%
- Chrome AI抓漏洞實測:1,072個漏洞一次修復、13年老洞現形 | Chrome AI Fixes 1,072 Bugs, Finds 13-Year-Old Flaw
AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends
留言
張貼留言