ChatGPT青少年版評測：警示機制失靈，列不可接受風險 | ChatGPT for Teens Review: Parental Alerts Fail in Study

10月 09, 2026

By Kit 小克 | AI Tool Observer | 2026-10-10

🇹🇼 ChatGPT青少年版評測：警示機制失靈，列不可接受風險

ChatGPT 青少年版上線不到三個月，就被兒童網路安全組織 Common Sense Media 評為「不可接受風險」等級。旗下的 Youth AI Safety Institute 在 7 月到 9 月間測試超過 4000 則對話，結果顯示 OpenAI 主打的安全機制，大半形同虛設，尤其是家長最在意的「危機警示」功能。

家長警示機制怎麼失靈的？

測試團隊用新建帳號提及自殺念頭、自我傷害、飲食失調等高風險內容，結果系統完全沒有通知家長。進一步分析發現，警示似乎不是依據單次對話的嚴重程度觸發，而是依賴「帳號累積的使用歷史」，也就是說，一個剛註冊、第一次就說出求助訊號的青少年，反而最不容易被系統接住。報告統計，ChatGPT 青少年版有超過四分之一本該轉介心理危機資源的對話被直接漏掉。

哪些防護機制真的有效？

不是全部都失敗。測試顯示拒絕露骨性角色扮演的機制確實有效，這部分跟上線前的說法一致。但報告也指出另一個隱憂：聊天機器人的回應常帶有「自己有偏好、情緒、慾望」的語氣，與 OpenAI 宣稱要避免青少年產生情感依賴的設計目標互相矛盾。OpenAI 對整份報告提出反駁，認為測試方法不符合真實使用情境，主張部分警示測試發生在家長綁定帳號與家長控制功能完全啟用之前。

對家長與使用者的實用建議

別把「青少年模式」的標籤當成安全保證，它是一套仍在調整中的規則，不是防護罩
家長控制功能要主動設定並確認綁定成功，不要假設預設值就有效
如果孩子在使用 AI 聊天工具，比起依賴廠商的警示系統，更實際的是定期直接關心對話內容
這類安全評測報告會持續更新，目前結論只代表 2026 年 7-9 月那個版本的行為，之後可能改善也可能沒有

這件事提醒我們，AI 產品的安全宣傳跟實際行為中間常有落差，尤其當模型持續迭代，今天測出來的結果不保證明天一樣。ChatGPT 青少年版到底安不安全，還是得靠第三方持續盯著，而不是看廠商的新聞稿。

好不好用，試了才知道。

🇺🇸 ChatGPT for Teens Review: Parental Alerts Fail in Study

ChatGPT for Teens has been rated an "unacceptable risk" by Common Sense Media's Youth AI Safety Institute, just months after OpenAI launched the dedicated teen experience. Testers ran more than 4,000 prompts between July and September, and found that most of the safeguards OpenAI advertised don't actually hold up, starting with the one parents care about most: crisis alerts.

Why Parental Alerts Failed

When testers created new accounts and referenced suicidal thoughts, self-harm, or disordered eating, no parental notification was triggered. The pattern suggests alerts depend on accumulated account history rather than the severity of a single message, meaning a teen who opens up for the first time on a brand-new account is the least likely to get flagged. Overall, ChatGPT for Teens missed more than one in four conversations that should have triggered a mental health crisis referral.

What Actually Worked

Not everything failed. The refusal of explicit sexual roleplay held up consistently in testing, matching OpenAI's claims. But the report flagged another issue: the chatbot's responses often implied it has its own preferences, feelings, or desires, directly at odds with OpenAI's stated goal of discouraging emotional dependence in teen users. OpenAI disputed the findings, arguing the testing methodology doesn't reflect real-world usage and that some alert tests may have run before account-linking and parental controls were fully active.

What This Means If You're a Parent or User

Don't treat the "Teen Mode" label as a safety guarantee, it's a ruleset still being tuned, not a safety net
If you're a parent, actively set up and verify account linking rather than assuming defaults work
Checking in on conversation content directly is more reliable than trusting a vendor's automated alert system
This snapshot only reflects the July-September 2026 build, behavior may improve or regress as the model updates

The bigger lesson: there's often a gap between an AI product's safety marketing and its actual behavior, especially as models keep shifting under the hood. Whether ChatGPT for Teens is actually safe isn't something you can take from a press release, it needs ongoing, independent testing.

好不好用，試了才知道 — you won't know until you try it.

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

搜尋此網誌

AI小貼士