Inside ByteDance's Ban on AI Distillation: Protecting TikTok, Evading Sanctions, and Building From Scratch

The ban is not entirely new. Some Chinese media report that Zhang had issued similar directives in prior internal meetings, and at least one report suggests the policy traces back to 2023, predating the current US-China tensions .

However, what changed in 2026 is the enforcement posture and the public nature of the directive. ByteDance has not publicly confirmed the instruction .

Why the Ban Exists: Three Reasons

1. Protecting TikTok from a US Ban

The most immediate driver is TikTok's fraught relationship with the US government . ByteDance has been under intense US scrutiny since the Protecting Americans from Foreign Adversary Controlled Applications Act was upheld by the Supreme Court in January 2025 . The company eventually finalized a deal to split TikTok's US operations from its global business in early 2026 .

People familiar with the company's thinking say the anti-distillation policy is "partly influenced by ByteDance's complex relationship with the US government" over TikTok . Aggressive use of American AI models for distillation could trigger regulatory retaliation from Washington, creating additional legal and political risk for TikTok's US operations .

ByteDance explicitly banned employees from distilling US AI models specifically — not distillation in general. This reflects concern that such practices could give the US government grounds for sanctions or a forced TikTok sale/ban .

2. Long-Termism and Technical Independence

Zhang argued that the Seed lab must develop frontier intelligence from first principles, building its Doubao models on original research rather than copying competitors . He framed it as a matter of "delayed gratification" and long-term technical depth .

Some Chinese media (e.g., Guixinren) report the policy actually dates back to 2023 and was driven by Zhang's belief in independent innovation, not US regulatory concerns .

3. Avoiding IP Theft Accusations

By publicly renouncing distillation, ByteDance distances itself from mounting US accusations that Chinese AI labs are stealing American intellectual property, reducing the risk of being targeted by Washington .

The US-China Distillation War

The ByteDance policy cannot be understood outside the broader context of a sharp escalation in US-China conflict over AI model distillation in July 2026.

US sanctions threats. Treasury Secretary Scott Bessent said the US would investigate whether Chinese AI models were distilled from American models and warned: "If we see that overseas models are stealing from our great companies, we have the ability to sanction them" . Bessent later said sanctions "remain on the table" .

Accusations against Moonshot AI. White House Science and Technology Adviser Michael Kratsios publicly accused China's Moonshot AI of a "large scale" effort to steal capabilities from Anthropic's Fable model through distillation . Kratsios said on X that the US government has information that Moonshot AI "distilled" capabilities from Anthropic's Fable AI for the development of its K3 model .

Broader allegations. US officials have also leveled accusations at DeepSeek and Alibaba, claiming their models show evidence of adversarial distillation — extracting vast volumes of responses from American frontier models . Anthropic sent a letter to the US Senate Committee on Banking, Housing, and Urban Affairs accusing Alibaba of carrying out "the largest known distillation attack on Anthropic to date" through 28.8 million exchanges using thousands of fraudulent accounts .

China's response. China's Ministry of Commerce flatly denied the accusations, arguing that distillation is a widespread practice in both countries and accusing the US of "AI hegemonism," threatening countermeasures . The ministry said Washington was threatening Chinese companies with punishment based on allegations alone, despite what it called a lack of factual or legal grounds .

Impact on safety dialogue. Analysts warn US threats are unraveling efforts to establish bilateral AI safety talks .

Enforcement Challenges

• No public confirmation. ByteDance has not publicly confirmed the instruction .
• Internal enforcement challenges. The ban applies to the Seed team, but policing whether individual researchers use distillation from open-source models in practice is extremely difficult — distillation traces are hard to detect.
• The ban covers US models specifically. Some reports indicate the primary prohibition is on distilling American models, not all distillation . Distilling from other Chinese open-source models may be a gray area.
• Prior directives had limited effect. Zhang had issued similar instructions before, yet ByteDance's models continued to show competitive performance, suggesting previous bans were not fully enforced .

How ByteDance Competes Without Distillation

Rather than taking shortcuts, ByteDance's Seed team is pursuing first-principles AI research. Early evidence suggests the strategy is paying off commercially.

Seedance 2.0 — ByteDance's video-generation model — has become the company's first truly profitable AI product. Key metrics:

• $147 million+ in monthly revenue (over RMB 1 billion) .S
• $2 billion annualized revenue run rate, up approximately 10x from the prior year .S
• 70–90% gross margins. Reports vary: LatePost reported 70% gross margin, while AI practitioners told 36Kr it reaches as high as 90% .S Volcano Engine head Tan Daize told 36Kr that the outside world has paid too much attention to Seedance's profits, and that the figure is not actually that high S.
• The model achieves this through an efficient Mixture-of-Experts (MoE) architecture with 200 billion parameters and relatively low inference compute costs .S
• Over 90% of Seedance 2.0's revenue comes from enterprise customers via Volcano Engine's MaaS (Model-as-a-Service) platform .S
• Volcano Engine raised its 2026 MaaS revenue target to RMB 15 billion ($2.2 billion), a tenfold increase from the approximately RMB 1.5 billion actually booked in 2025 .S

The core idea: by building original architecture from scratch rather than distilling, ByteDance avoids legal risk, develops proprietary technical moats, and — in the case of Seedance — achieves margins that rival or exceed those of US cloud AI services.

However, the long-term bet carries risk. Zhang acknowledged that the ban may cause ByteDance's Seed model to lag behind Chinese competitors such as DeepSeek, Kimi, and Q1 in benchmark rankings . Whether the build-from-scratch approach can scale to frontier language models — which are harder to monetize than video generation — remains an open question.