Mistral Vibe評測：免費對話預設訓練AI，一鍵關閉教學 | Mistral Vibe Review: Free Chats Train AI by Default

By Kit 小克 | AI Tool Observer | 2026-09-04

🇹🇼 Mistral Vibe評測：免費對話預設訓練AI，一鍵關閉教學

Mistral Vibe這幾天在Hacker News炸出一波討論：法國AI新創Mistral的聊天機器人Vibe，免費版對話預設被拿去訓練模型，而官方文件先前的說法卻讓人以為是「預設關閉」。等到用戶發現不對勁回頭查證，Mistral才悄悄改了說明頁面，這波「文件跟實際設定對不上」的操作，讓不少工程師直呼被將了一軍。

發生了什麼事：Team方案悄悄拿掉集中關閉鍵

問題核心出在Mistral Vibe的Team方案。過去Team用戶的訓練資料是預設關閉，而且管理員可以用一個「組織層級」的開關統一幫全公司關掉。但Mistral最近把這個集中開關拿走了，訓練資料變成預設開啟，而且改成要每個人自己進帳號設定去關——不再能一鍵幫全團隊處理。文件更新的速度還落後於實際政策變更，導致用戶看到的說明跟系統實際行為兜不起來。

為什麼「預設訓練」讓人生氣

Hacker News討論串裡吵最兇的不是「要不要訓練」本身，而是用詞。Mistral官方說法接近「opt-in by default」（預設加入），但這在邏輯上根本說不通——opt-in的意思本來就是「使用者要主動選擇加入」，如果是「預設就開」，那其實是opt-out（預設參與、要自己退出）。有評論直言這種措辭是刻意包裝對用戶不利的預設值，讓人一眼看不出自己其實已經被「加入」了。

各方案的預設設定整理

Vibe 免費版：預設沒有退出訓練，需自行到隱私設定關閉
Vibe Team：現在改成個人自行設定，管理員無法統一代管
Vibe Enterprise：預設退出訓練，管理員層級控管
Mistral Studio／API：使用者可自行開關，預設不影響付費API流量

怎麼關閉：一分鐘設定教學

網頁版：登入後進入 Admin 面板 → Vibe → Privacy，關閉「Allow your interactions to be used to train our models」
手機版（iOS／Android）：Settings → Account → Data & Account Controls，取消勾選「Enable data sharing」
Studio／API：Admin 面板 → Privacy → Anonymous improvement data，關閉即可

要注意的是Vibe跟API的開關是分開的，兩邊都要各自確認一次，不會互相連動。

小克怎麼看

老實說，AI公司拿對話訓練模型不是新聞，OpenAI、Google都做類似的事，重點從來不是「有沒有訓練」而是「講不講清楚、好不好關」。Mistral這次踩到的雷是文件跟實際行為對不上，加上拿掉團隊集中管理選項，等於把責任丟回給每個使用者自己顧。如果你或公司在用Mistral Vibe處理內部資料，這篇文章讀完直接去把三個地方的開關都檢查一遍，比等官方公告更實在。

好不好用，試了才知道。

🇺🇸 Mistral Vibe Review: Free Chats Train AI by Default

Mistral Vibe just triggered a Hacker News pile-on after users discovered the French AI startup's chatbot trains on free-tier conversations by default — while its own documentation previously implied the opposite. When users caught the discrepancy, Mistral quietly rewrote the help page, and the "docs said one thing, product did another" pattern is what actually lit the fuse.

What Happened: Team Plans Lost Their Central Kill Switch

The flashpoint is Mistral Vibe's Team tier. Training used to be off by default for Team accounts, and admins had one org-wide toggle to keep it that way for everyone. Mistral quietly removed that centralized control, flipped training to on by default, and pushed the responsibility down to each individual user's own privacy settings. The documentation update lagged behind the actual policy change, so what users read didn't match what the product was actually doing.

Why "Opt-In by Default" Made People Angry

The real argument on Hacker News wasn't about training itself — it was about the language. Mistral's framing leaned toward calling the new default "opt-in," which is a contradiction in terms: opt-in means users actively choose to participate. If something is on unless you turn it off, that's opt-out, full stop. Commenters called it out as corporate phrasing designed to make an unfavorable default sound less alarming.

Default Settings by Plan

Vibe Free: not opted out by default — you must disable it yourself
Vibe Team: now managed per-user; admins lost the org-wide toggle
Vibe Enterprise: opted out by default, admin-controlled
Mistral Studio / API: user-controlled, doesn't affect paid API traffic by default

How to Turn It Off

Web: Admin panel → Vibe → Privacy → toggle off "Allow your interactions to be used to train our models"
Mobile (iOS/Android): Settings → Account → Data & Account Controls → uncheck "Enable data sharing"
Studio/API: Admin panel → Privacy → Anonymous improvement data → disable

Note: Vibe and API toggles are separate — you need to check both, since flipping one doesn't touch the other.

Kit's Take

Training on chat logs isn't new — OpenAI and Google do versions of this too. The real issue is never "do they train," it's "do they say so clearly and make it easy to turn off." Mistral's misstep here was letting documentation drift from reality and quietly pulling a team-wide safety net. If you or your company runs internal data through Mistral Vibe, go check all three toggles yourself rather than waiting for an official statement.

好不好用，試了才知道。

Sources / 資料來源

延伸閱讀 / Related Articles

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

Claude Code 實測：AI 幫你寫程式到底行不行？ | Claude Code Review: Can AI Really Code for You?

3月 26, 2026

閱讀完整內容

搜尋此網誌

AI小貼士