Breaking

US AI Firms Accuse Chinese Rivals of Improper Model Distillation

June 26, 2026 at 14:22 EDT

Policy & Regulation
Foundation Models
Security

Leading US AI companies including Anthropic and OpenAI accuse Chinese competitors such as DeepSeek, Moonshot AI and MiniMax of using a technique called distillation to improperly tap the outputs of top American models and build low-cost rival chatbots.

February 2026 · OpenAI & Anthropic

US Labs Accuse Chinese Rivals of "Industrial-Scale" Model Distillation

OpenAI and Anthropic allege DeepSeek, Qwen and others used tens of thousands of fake accounts to siphon tens of millions of responses from leading US models — violating terms of service to train cheaper rivals.

~49K

fake accounts flagged across the two cases

44.8M

total interactions siphoned from Claude models

US labs now sharing detection intelligence (Apr 2026)

Interactions extracted, by case

Each block ≈ 4M interactions with Claude models

16M+

DeepSeek + 2 firms
~24K accounts · early 2026

28.8M

Alibaba (Qwen)
~25K accounts · Apr–Jun 2026

How "distillation at scale" works

FAKE ACCOUNTS

Thousands of accounts & proxies probe US "teacher" models

→

HARVEST OUTPUTS

Reasoning, tool-use & chain-of-thought data collected at scale

→

TRAIN RIVAL

A cheaper "student" model is trained on the harvested responses

THE DEFENSE

Distillation is an industry standard — US labs use it too. Critics call singling out Chinese firms one-sided, and note DeepSeek R1 (671B MoE) is a genuinely strong, low-cost open model that benefits competition.

THE CONCERN

Fake accounts and ToS violations are a separate issue. Cheap copies could offset chip export controls, and models stripped of safety guardrails raise military and surveillance risks.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.