Clearview AI 誤判：無辜婦女坐牢 108 天，人臉辨識的真實風險 | Clearview AI False Match: Innocent Woman Jailed 108 Days
By Kit 小克 | AI Tool Observer | 2026-03-30
🇹🇼 Clearview AI 誤判：無辜婦女坐牢 108 天，人臉辨識的真實風險
2026 年 3 月，CNN 披露了一則令人心寒的案例：美國田納西州 50 歲的祖母 Angela Lipps，因 Clearview AI 人臉辨識系統的一次錯誤比對，在未犯任何罪的情況下，被迫在監獄中度過了整整 108 天。這不是科幻小說——這是 AI 誤判在現實世界的真實代價。
事件經過：一張假 ID 害了一個無辜的人
2025 年 4 至 5 月間，北達科他州法哥市（Fargo）發生一系列銀行詐騙案，罪犯持偽造美軍身分證開設帳戶。執法人員透過 Clearview AI 系統掃描嫌犯面孔，結果系統將 Lipps 的照片與犯罪現場照片判定為同一人。
問題核心在於：警方沒有做任何交叉驗證。Lipps 的銀行記錄與社會安全存款收據明確顯示，案發期間她人在田納西州。這些證據本可在數天內整理完畢，但她已先在田納西被捕、引渡至北達科他。整個案件直到 2025 年聖誕夜才終於撤銷。
這 108 天裡，她失去了家、車子，以及她的狗。
Clearview AI 是什麼？誤判率有多高？
Clearview AI 是一家美國公司，其資料庫據稱含有從網路公開來源抓取的數百億張人臉照片，供執法機關進行 AI 人臉辨識比對。但準確率爭議從未停止——多項研究顯示，該技術對膚色較深或年長女性族群的誤判率明顯偏高，而 Lipps 正好屬於這個族群。
Tom's Hardware 的調查指出，Lipps 的案例並非特例：類似的 AI 人臉辨識冤案在全美已有數十起記錄，而執法機構仍持續使用這項技術，問責機制幾乎付之闕如。
技術問題，還是制度問題？
這個案件揭露的核心困境不只是 AI 準確率——更是執法文化對 AI 辨識結果的盲目信任。當系統輸出一個「比對結果」，人類應扮演最後防線，而非橡皮圖章。
- 美國目前沒有聯邦層級的人臉辨識規範法律
- 歐盟《AI 法案》執行日期已延後至 2027 年
- 白宮最新 AI 框架未明確限制執法機關的人臉辨識使用
對普通公民來說，這意味著：你的臉可能被 Clearview AI 誤判，而當前幾乎沒有任何法律能快速還你清白。
開發者與產品人該注意什麼？
如果你正在構建任何含有 AI 辨識功能的系統，這個案例是最直白的警告：高召回率（recall）不等於高精確率（precision），任何影響人身自由的判斷，都不應讓演算法做最終決定。Human-in-the-loop 不是可選項，是基本道德義務。
好不好用，試了才知道——但代價是別人的 108 天自由，那就不只是功能問題了。
🇺🇸 Clearview AI False Match: Innocent Woman Jailed 108 Days
In late March 2026, CNN exposed a chilling case that puts AI facial recognition in the harshest possible light: Angela Lipps, a 50-year-old grandmother from Tennessee, spent 108 days in jail for crimes she never committed — all because Clearview AI matched her face to a fraudster's fake military ID. This is not a cautionary tale. It already happened.
What Happened: A Fake ID, a False Match, 108 Days Gone
Between April and May 2025, a series of bank fraud cases hit Fargo, North Dakota. The perpetrator used a forged U.S. Army ID to open accounts and withdraw funds. Police fed the suspect photo into the Clearview AI system and got a hit: Angela Lipps. No further verification was performed before the arrest warrant was issued.
Her bank records and Social Security deposit receipts placed her in Tennessee throughout the entire period. Her public defender gathered this evidence within days — but by then, Lipps had already been arrested in Tennessee and extradited to North Dakota. The charges weren't dismissed until Christmas Eve 2025. In those 108 days, she lost her home, her car, and her dog.
The Clearview AI Problem in Numbers
Clearview AI operates a database of tens of billions of facial images scraped from public web sources. Law enforcement agencies pay for access to run AI facial recognition matches against suspects. The company claims high accuracy — but independent research consistently shows elevated error rates for older women and people with darker skin tones, the exact demographic Angela Lipps belongs to.
Tom's Hardware's investigation makes clear this is not isolated: documented wrongful arrests from AI facial recognition across the U.S. now number in the dozens, and police departments continue deploying these tools with minimal oversight or accountability mechanisms in place.
Technology Failure or Institutional Failure?
The deeper issue is not accuracy alone. It is the institutional blind trust in AI output. When a system returns a facial match, humans are supposed to be the last line of defense — not a rubber stamp. In Lipps' case, no officer cross-checked basic alibi evidence before she was arrested and extradited across state lines.
- The U.S. has no federal law regulating law enforcement use of facial recognition
- The EU AI Act enforcement has been delayed until 2027
- The White House's new AI framework does not restrict police use of Clearview AI or similar tools
For ordinary citizens, this means: your face could be misidentified by an AI system, and there is currently almost no legal mechanism to quickly restore your freedom.
What Every Builder Should Take From This
If you are building any system with AI recognition capabilities, this case is the clearest warning available: high recall is not the same as high precision, and any decision that affects human liberty must not be delegated to an algorithm alone. Human-in-the-loop design is not an optional feature — it is a baseline ethical requirement, not an upsell.
You won't know until you try it — but when the cost of a mistake is someone else's 108 days of freedom, it stops being just a product problem.
Sources / 資料來源
- CNN: Police used AI facial recognition to arrest a Tennessee woman for crimes she says she never committed
- Tom's Hardware: Facial recognition is jailing the wrong people, but police keep using it anyway
- Boing Boing: Grandmother jailed for months on AI facial recognition false match
延伸閱讀 / Related Articles
- ChatGPT 讓 Cloudflare 偷讀你的資料：打字前已收集 55 項資訊 | ChatGPT Lets Cloudflare Collect 55 Data Points Before You Type
- Claude Code Bug：背景自動執行 git reset --hard 靜默刪除你的程式碼 | Claude Code Bug: Background git reset --hard Silently Deletes Your Code
- ARC-AGI-3 發布：頂尖 AI 全部得分不到 1% | ARC-AGI-3: Every Top AI Model Scored Under 1%
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