AI蒸餾風波評測：六家中國AI廠遭美方點名抄襲模型 | AI Distillation Review: US Names 6 Chinese AI Labs

9月 09, 2026

By Kit 小克 | AI Tool Observer | 2026-09-10

🇹🇼 AI蒸餾風波評測：六家中國AI廠遭美方點名抄襲模型

AI蒸餾是什麼？NSA、CISA、FBI 為何聯手出手

美國國家安全局（NSA）、網路安全暨基礎設施安全署（CISA）與聯邦調查局（FBI）9 月 9 日發布聯合公告，直指六家中國AI公司——DeepSeek、Moonshot AI、阿里巴巴（Alibaba）、MiniMax、StepFun、Z.AI——自 2024 年底起，對美國前沿AI模型進行「工業規模」的AI蒸餾，並認定這種行為已違反美國廠商的使用條款，威脅美國的技術領先地位。

六家廠商各自的「取材」對象

公告內容相當具體，點名了每家公司蒸餾的目標模型：

DeepSeek：2024 年底到 2025 年中，被指蒸餾 Claude 3.7、Claude Sonnet 4／4.5、Claude Opus 4.1、Gemini 2.5 Pro／Flash Preview、GPT-4、GPT-4o、GPT-4 Mini、GPT-4 Nano、GPT-5、Grok 4，用來訓練 R1 與 V3 模型。
Moonshot AI：至少從 2025 年中起，大量擷取 Claude Fable 5 的輸出訓練 Kimi-K3，並用 GPT-4o 的輸出訓練 Kimi-K2。
阿里巴巴：2025 年底蒸餾 Claude 與 GPT-5，用來強化 Qwen 系列模型。

公告用詞很重：蒸餾不是這些公司的「輔助手段」，而是「核心」的模型建構方式。

蒸餾到底是不是「抄襲」？

技術上，AI蒸餾指的是大量呼叫別家模型的 API、收集輸出結果，再拿這些資料訓練自己的（通常更小、更便宜的）模型——這在學術界是行之有年的正常技巧，OpenAI、Google 自己也用類似方法壓縮模型。爭議點在於「規模」與「條款」：Claude、GPT、Gemini 的使用條款都明文禁止用輸出訓練競品模型，這次公告等於用國安層級的措辭，把違反 TOS 的行為定調成國安威脅。但要注意，這是一份「公告」（advisory），不是起訴書，目前沒有任何法律行動，執行力其實有限。

對開發者與用戶的實際影響

如果你在用 Claude、GPT-5 或 Gemini 的 API，這份公告解釋了一個現象：為什麼便宜的中國模型（尤其 DeepSeek、Qwen）常常能在幾個月內追上前沿模型的跑分——很大一部分靠的就是蒸餾。對獨立開發者的實際提醒是：檢查自己專案有沒有用模型輸出訓練別的模型，這踩的是同一條 TOS 紅線，帳號被封的風險是真實存在的，不需要國安等級的爭議也會發生。

好不好用，試了才知道。

🇺🇸 AI Distillation Review: US Names 6 Chinese AI Labs

What Is AI Distillation? Why NSA, CISA, and the FBI Got Involved

The NSA, CISA, and FBI published a joint advisory on September 9 naming six Chinese AI companies — DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI — for running "industrial-scale" AI distillation campaigns against U.S. frontier models since late 2024, arguing the practice violates U.S. companies' terms of use and threatens American technological leadership.

Who Distilled What

The advisory gets specific about each company's targets:

DeepSeek: Between late 2024 and mid-2025, allegedly distilled Claude 3.7, Claude Sonnet 4/4.5, Claude Opus 4.1, Gemini 2.5 Pro/Flash Preview, GPT-4, GPT-4o, GPT-4 Mini, GPT-4 Nano, GPT-5, and Grok 4 to train R1 and V3.
Moonshot AI: Since at least mid-2025, extracted large volumes of Claude Fable 5 output to train Kimi-K3, and GPT-4o output to train Kimi-K2.
Alibaba: Distilled Claude and GPT-5 in late 2025 to strengthen its Qwen family.

The language is blunt: distillation isn't a side technique for these firms — it's described as core to how they build models at all.

Is Distillation Actually "Theft"?

Technically, AI distillation means repeatedly calling another company's model API, collecting its outputs, and training a (usually smaller, cheaper) model on that data — a well-established technique that OpenAI and Google themselves use to compress their own models. The real dispute is scale and contract terms: Claude, GPT, and Gemini's terms of service explicitly ban using outputs to train competing models. This advisory frames a TOS violation in national-security language. Worth noting: it's an advisory, not an indictment — no legal action has followed, and enforcement teeth are limited.

What This Actually Means for Developers

If you use Claude, GPT-5, or Gemini APIs, this advisory explains a pattern you've probably noticed: why cheap Chinese models — especially DeepSeek and Qwen — keep closing the benchmark gap within months. Distillation is a big part of the answer. The practical takeaway for indie developers: check whether your own project trains another model on API outputs — that's the same TOS line, and account bans happen without needing a national-security-level dispute.

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