U.S. proposes law to block foreign AI model distillation
The proposed Deterring American AI Model Theft Act seeks to prevent foreign entities from training models off American tech to maintain a safety and capability lead.
The short version
- Distillation allows foreign models to mimic American AI capabilities at a fraction of the cost by using false accounts to 'feed' on model responses.
- The proposed act would allow the government to add implicated Chinese labs to an entity list, requiring American companies to block their traffic.
- Proponents argue that distillation removes the 'kill-switch' safeguards of original models, as offspring models do not inherit the same safety controls.
- The bill recently passed the House Foreign Affairs Committee with a unanimous 43–0 vote.
Why it matters
If passed, this law would create a legal framework to treat AI model theft as a national security issue. It aims to give American developers more time to prioritize safety over speed by intentionally widening the technological gap between the U.S. and China.
The mechanics of model distillation
Distillation is a process where a developing AI model is trained using the outputs of a pre-existing, more advanced model. Foreign prompt engineers use false accounts to submit difficult questions to American-owned models and then feed those responses back into their own systems. This allows the new model to absorb knowledge from its "father model."
This method effectively closes the capability gap between American and foreign AI systems. It enables competitors to simulate high-level capabilities for a fraction of the cost of building a model from scratch. Because many American systems are closed-source, their unique technical components are not shared publicly, but distillation bypasses these protections.
While American companies have tried to stop this by shutting down suspected feeder accounts, new models continue to appear. Currently, distillation is not explicitly illegal under American law, making it difficult for the government to take meaningful action. The process is hard to stop entirely because companies can simply open new accounts to circumvent existing blocks.
Security risks and lost control
A primary concern with distillation is that offspring models often do not absorb the same safety safeguards as their original models. Once a model is distilled, the original American company loses direct control over the resulting intelligence. This makes it impossible to implement a "kill-switch" to shut down an AI model if it becomes unsafe.
The U.S. currently has no governing power over Chinese AI models, and Chinese policy does not require developers to release risk assessments. By preventing the extraction of key technical features from American models, the government hopes to reduce the risk of dangerous technology spreading. The proposed act specifically aims to protect the technical features of closed-source, American-owned models.
The goal is to ensure that humans maintain the ability to intervene or shut down AI technology. Because AI is increasingly powerful, proponents argue that the "move fast and break things" mindset is no longer appropriate. Federal deterrence is seen as a way to ensure that safety mechanisms are prioritized over the speed of the AI race.
The Deterring American AI Model Theft Act
The proposed legislation would direct the Secretary of State to identify entities involved in the unauthorized extraction of model capabilities. Once identified, the End-User Review Committee would vote on whether to add these entities to the Bureau of Industry and Security (BIS) Entity List. Being on this list subjects a company to individual licensing requirements that the U.S. government can use to restrict access.
Under these rules, American AI companies would be obligated to deny traffic coming from IP addresses linked to the blacklisted entities. This creates a formal policy to deter distillation by making it harder for foreign actors to access American technology. The act seeks to provide the U.S. with more time to develop models safely by maintaining a wider capability gap with China.
The bill recently gained momentum after passing through the House Foreign Affairs Committee with a unanimous 43–0 vote. While it faces a challenge to clear both the House and the Senate, supporters believe it is a necessary first step. They argue that a timely answer is more important than a perfect one in a rapidly evolving technological landscape.
Proposed improvements to the bill
Critics of the current draft suggest that the investigation timeline for implicated parties is too slow. The act currently allows up to 180 days for an initial assessment and up to 210 days for a congressional report. This timeframe could allow distilling companies to adapt or relocate before they are officially identified by the government.
To address this, some suggest that the addition of affiliate entities should be made automatic. Currently, Section 5(a) requires a vote to add entities that own 50 percent or more of a sanctioned company. Making this automatic would prevent companies from simply creating new names to bypass the time-consuming voting process for essentially identical entities.
The ultimate objective of the legislation is to buy American companies breathing room. By slowing down the rate at which foreign competitors can catch up, the U.S. hopes to allow domestic developers to implement thorough safety testing. This policy shift would move the focus from a pure race for dominance to a regulated environment where human control is preserved.
What to watch
- Whether the bill can secure a majority vote in the Senate.
- How American AI companies will adjust their IP blocking to comply with potential new licensing requirements.
- Whether Chinese labs will develop new methods to bypass IP-based blacklisting.
Try it
Look up the Bureau of Industry and Security (BIS) Entity List to see which other types of technologies are currently restricted for export.
Sources and links
- OfficialSafe AI development ibm.com
- Officialopenai.com/index/hugging-face-incident-and-the-road-ahead/ openai.com
- Officialanthropic.com/news/detecting-and-preventing-distillation-attacks anthropic.com
- FilingDeterring American AI Model Theft Act congress.gov
- Governmentfederalregister.gov/documents/2025/01/15/2025-00636/framework-for-arti federalregister.gov
- ReportingBrown Political Review: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by Brown Political Review · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
