US open source AI regulation

Modified Chinese Open-Weight Models Need Provenance Disclosure, Not a Ban on Editing

Startups now strip censorship from Qwen and sell the result to US firms. Disclosure and testing fit that better than blanket bans or pre-release gates.

Evaluating PRC Open-Weight Models People of Internet Research · US 12x Agent hijacking vs US models R1-0528 agents followed malicious … ~8 months DeepSeek V4 lag behind frontier CAISI's estimate for the most capa… ~12 US firms using ReAligned Qwen Customer count cited by Lazarus's … peopleofinternet.com
Evaluating PRC Open-Weight Models People of Internet Research · US 12x Agent hijacking vs US models ~8 months DeepSeek V4 lag behind frontier ~12 US firms using ReAligned Qwen peopleofinternet.com

Key Takeaways

Rest of World reported on 2 November 2026 that Western startups are rewriting Chinese open-weight models for Western customers. Boston's Lazarus AI removed the mandatory censorship from Alibaba's Qwen. It drew the banned topics from leaked censorship guidelines and used answers from Google's Gemini Flash to realign the models. Lazarus's former chief scientist Eric Hartford says about a dozen US companies run these "ReAligned" versions. Hirundo edits Qwen and other Chinese open model weights to change what Rest of World calls their "values, geopolitical perceptions, censorship, and state-aligned bias." Some government agencies, Hartford says, remain reluctant to adopt Chinese models at all.

The case is small, but it exposes a gap in US policy. A model edited and redistributed by a third party is neither a PRC product nor a US frontier release. Existing rules do not say what it is.

The strongest case for restriction

Restriction has a serious argument behind it. NIST's Center for AI Standards and Innovation (CAISI) evaluated DeepSeek R1, R1-0528 and V3.1 in September 2025. It found R1-0528 agents were 12 times more likely than US frontier models to follow malicious instructions in a simulated environment. The models also echoed four times as many inaccurate CCP narratives as US reference models. If weights carry hidden risks, a startup's fine-tune can remove one visible flaw, censorship, while leaving others in place. A buyer who sees a clean "Western" label may assume the model is safe when it has only been made agreeable. Rest of World itself cites a CrowdStrike analysis finding that politically sensitive prompts made DeepSeek R1 up to 50% more likely to produce code with severe security vulnerabilities. Editing the refusals does not necessarily fix that.

The worry is legitimate. The tool proposed to address it is the wrong one.

Why a ban or a licensing gate misfires

The 2024 NTIA report on open model weights reached a conclusion that still holds. It found that "current evidence is not sufficient to definitively determine either that restrictions on such open weight models are warranted, or that restrictions will never be appropriate in the future." Its answer was to monitor a portfolio of risks and stay ready to act. It did not impose restrictions up front. Nothing in the 2026 evidence reverses that logic. A dozen enterprise customers running edited models is not a risk event.

A ban on Chinese-origin weights would not work in practice, for three reasons.

A mandatory pre-release gate would also fail. CAISI's review, as the news hook describes it, covers closed frontier models. Extending it to every open-weight derivative would stall small developers and reward the largest firms, which can afford compliance. It would also mismatch the review to the problem. CAISI's public PRC-model evaluations come after release, and that is the better pattern.

What proportionate rules would look like

CAISI's own record shows the capability gap is narrowing but real. Its May 2026 evaluation of DeepSeek V4 Pro called it the most capable PRC model it had evaluated and put it roughly eight months behind the US frontier. That is close enough that buyers have a reason to want these models. It is far enough that the claim of parity needs checking. A realistic policy would rest on three measures.

  1. Provenance disclosure. Anyone redistributing a modified model should state the base model, the method used (fine-tuning, unlearning, weight editing) and the data source for any realignment. A buyer should not need to guess whether "ReAligned Qwen" still carries Qwen's other behaviors.
  2. Procurement-level testing, not market-level gates. Federal buyers can require published results on agent hijacking, jailbreak resistance and code security before adopting any model, including derivatives. The agencies that are wary today would then have an objective standard in place of a blanket reluctance. This reuses CAISI's existing benchmarks and does not create a licensing regime.
  3. Evidence on self-reported claims. Vendors of edited models advertise large reductions in bias scores. These are self-reported, and CAISI has already found that developer-reported results can run more favorable than its own. Independent replication should be a condition of government use, and open evaluation suites make it cheap.

There is a free-expression point as well. A model that refuses to discuss Tiananmen or Xinjiang degrades the information environment for anyone who uses it. Hirundo's analysis found Qwen models do not engage with those subjects. Lawfully removing that behavior is speech-positive. Policy should not penalize it just because the weights started in Hangzhou.

The open-source dependency

Open-weight policy also sits inside a wider question about who maintains open infrastructure. EFF argued in October 2026 that corporations that depend on volunteer-maintained projects such as PHP should fund them. The same applies here. If US firms rely on open weights, public money for open evaluation tools buys more safety than restrictions on distribution.

Bottom line

The edited-Qwen market is early, and the evidence of harm is thin. Congress and the agencies should respond with disclosure rules and measured procurement standards. Lawful modification of open weights should stay legal, and any restriction should wait for evidence of specific risk, as NTIA advised in 2024.

Sources & Citations

  1. Rest of World: Startups recalibrate Chinese AI for the West
  2. NIST CAISI: Evaluation of DeepSeek AI Models (Sept 2025)
  3. NIST CAISI: Evaluation of DeepSeek V4 Pro (May 2026)
  4. NTIA: Dual-Use Foundation Models with Widely Available Model Weights (2024)
  5. EFF: The Vulnerability Apocalypse and Why We Should Fund Open Source