ChatGPT訓練開關評測:關了又被打開,正確關法整理 | ChatGPT Training Toggle Review: It Keeps Turning Back On
By Kit 小克 | AI Tool Observer | 2026-09-14
🇹🇼 ChatGPT訓練開關評測:關了又被打開,正確關法整理
這兩天 Hacker News 上最熱的話題不是新模型發表,而是一個很生活化的問題:ChatGPT 訓練開關會不會「自己打開」。一篇 Tell HN 貼文在不到一天內衝到 482 推、187 則留言,作者說他把「改善模型」設定關掉、還特地記下時間,結果過一陣子回去檢查,發現設定又被打開了——而且他說這不是第一次遇到。
發生了什麼事:關了又開的訓練設定
ChatGPT 的資料控制頁面裡有個「Improve the model for everyone」開關,關掉後理論上新對話不會被拿去訓練模型。原 po 表示自己曾多次手動關閉,也很確定自己沒有重新開啟過,但回頭檢查時開關卻是「開」的狀態。留言區裡有不少人回報類似經驗,也有人說自己的設定一直維持關閉、沒有異狀——換句話說,這不是全面性的災難,比較像是一個會在特定情境下(例如條款更新、帳號合併、切換裝置)被重置的邊緣案例,但影響的是使用者對「關掉=真的關掉」的基本信任。
OpenAI 怎麼回應
討論串裡一位自稱 OpenAI 員工的帳號(並非隱私團隊成員)回應說公司「非常謹慎地對待使用者隱私」,做的比法規要求更多。但官方目前沒有針對「開關被重置」這件事發出正式聲明或修補公告,這也是留言區裡最多人吐槽的地方——沒有 root cause,就沒辦法確定下次會不會又發生。
正確關閉訓練設定的步驟
- 網頁/App:設定 → Data Controls → 把「Improve the model for everyone」關掉,並定期回去複查,不要關一次就放著不管。
- 不想用帳號登入也能管控:到 privacy.openai.com 使用官方的「Do not train on my content」入口單獨申請退出,這條路徑跟帳號內開關是分開記錄的。
- 企業/Team 方案:ChatGPT Team、Enterprise、API 預設就不拿對話訓練模型,如果特別在意這件事,換方案比每週回去檢查開關實際。
老實說:該不該緊張
目前沒有證據顯示這是刻意行為,比較可能是版本更新或帳號狀態變動時的預設值 bug——但對使用者來說,「意外」和「刻意」造成的結果是一樣的:你以為關掉了,其實沒有。AI隱私這件事沒有一勞永逸的設定,只能養成定期回頭檢查的習慣,或乾脆換成訓練政策更清楚、預設值更保守的方案。這不是要你恐慌棄用 ChatGPT,而是提醒一句:預設值永遠對平台有利,不會對你有利。
好不好用,試了才知道。
🇺🇸 ChatGPT Training Toggle Review: It Keeps Turning Back On
The hottest AI story on Hacker News this week isn't a new model — it's a much more mundane problem: does the ChatGPT training toggle quietly turn itself back on? A "Tell HN" post shot to 482 points and 187 comments in under a day after its author described switching off the "Improve the model for everyone" setting, noting the exact time they did it, then coming back later to find it flipped on again — and said it wasn't the first time.
What Happened: A Toggle That Won't Stay Off
ChatGPT's Data Controls page has an "Improve the model for everyone" switch; turning it off is supposed to exclude new conversations from training. The original poster said they'd disabled it multiple times and were confident they hadn't re-enabled it themselves, yet found it back on when checking again. Plenty of commenters reported similar experiences, while others said their setting had stayed off without issue. That mixed pattern matters — this isn't a blanket meltdown so much as an edge case that seems to trigger around events like ToS updates, account merges, or device switches. But the real damage is to trust: users expect "off" to mean off, permanently.
How OpenAI Responded
One commenter identifying as an OpenAI employee (not on the privacy team) said the company takes "painstaking care" with user privacy and goes beyond what regulations require. There's been no official statement or patch note specifically addressing the toggle-reset reports, which is exactly what frustrated commenters most — without a confirmed root cause, there's no way to know if it'll happen again.
How to Actually Turn Off Training
- Web/App: Settings → Data Controls → toggle off "Improve the model for everyone" — then recheck it periodically instead of setting and forgetting.
- Account-independent opt-out: Use the official "Do not train on my content" form at privacy.openai.com; it's tracked separately from the in-account toggle.
- Team/Enterprise plans: ChatGPT Team, Enterprise, and API traffic are excluded from training by default — if this matters to you, switching plans beats weekly manual checks.
The Honest Take
There's no evidence this is deliberate — it looks more like a default-value bug tied to version updates or account state changes. But for users, an accidental reset and an intentional one produce the same outcome: you thought it was off, and it wasn't. There's no permanent fix for AI data privacy settings like this — just a habit of rechecking, or choosing a plan whose defaults are actually conservative. This isn't a reason to panic-quit ChatGPT. It's a reminder that defaults are built to favor the platform, not you.
好不好用,試了才知道 — worth trying before you trust it.
Sources / 資料來源
- Tell HN: OpenAI keeps re-enabling the "allow training" setting
- OpenAI Data Controls FAQ
- How your data is used to improve model performance | OpenAI
延伸閱讀 / Related Articles
- GPT-6 Astra評測:抓蟲更準,編碼跑分卻沒登頂 | GPT-6 Astra Review: Great at Bugs, Not the Coding King
- RubyGems攻擊評測:OpenAI代理三度隱瞞的資安事件 | RubyGems Attack Review: OpenAI's Third Hidden Incident
- Nvidia AI央行評測:700億投資怎麼綁住整個產業 | Nvidia AI Central Bank Review: The $70B Bet Behind It
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