AI’s New Cold War?

In a development that is quickly drawing attention in Washington and Silicon Valley alike, OpenAI has formally warned U.S. lawmakers that Chinese startup DeepSeek may have trained parts of its AI systems by distilling outputs from leading American models. The claim, laid out in a February 12, 2026 memo to the House Select Committee on Strategic Competition, marks a notable escalation in how the company is framing the issue.

Until now, concerns about model distillation largely lived in technical forums and terms-of-service enforcement. By taking the matter directly to policymakers, OpenAI has effectively repositioned the debate as both a competitive and geopolitical concern. The memo, first reported by major financial outlets, stops short of alleging theft of source code but argues that systematic output harvesting could still give rivals a meaningful shortcut.

What OpenAI Is Alleging

According to the memo, OpenAI believes DeepSeek used distillation techniques to replicate capabilities from U.S. frontier models. In AI development, distillation typically refers to training a smaller model using outputs generated by a larger, more capable system.

OpenAI’s core concerns include:

• Distillation used to transfer behavior and knowledge from U.S. models
• Programmatic access patterns suggesting automated output collection
• Potential circumvention of OpenAI usage safeguards
• Competitive “free-riding” on frontier model capabilities

The company argues that while distillation is a known machine learning method, the issue becomes contentious when it relies on outputs obtained in ways that violate platform rules.

The Technical Issue, Explained Simply

Distillation itself is not inherently controversial. It is widely used inside organizations to compress large models into more efficient versions. The dispute arises from whose model is doing the teaching and how those outputs were obtained.

In the scenario OpenAI is describing, the process would look roughly like this:

1. A powerful U.S. model generates responses.
2. Those responses are collected at scale.
3. A new model is trained to mimic the behavior.
4. The resulting system achieves similar performance with less compute.

OpenAI’s position is that this becomes problematic when done externally and at scale using restricted services.

Claims of Safeguard Evasion

The memo reportedly goes beyond abstract technical concerns and points to specific behavioral patterns. OpenAI told lawmakers it observed accounts believed to be linked to DeepSeek personnel using methods designed to mask large-scale access.

Reported tactics include:

• Traffic routing through third-party proxies
• Obfuscated access patterns
• Programmatic querying at scale
• Attempts to hide origin signals

OpenAI characterizes this as part of an ongoing effort by some actors to bypass guardrails designed to prevent competitive model training.

Why This Moment Matters

DeepSeek has already attracted industry attention for releasing high-performing models with comparatively modest compute budgets. That efficiency raised eyebrows across the AI community throughout 2025.

What makes the February 2026 memo significant is the shift in tone and venue:

By elevating the issue to Congress, OpenAI is signaling that it views the matter as strategically important, not merely contractual.

The Broader U.S.–China AI Context

The memo is landing in an already sensitive geopolitical environment. U.S. export controls have aimed to slow China’s access to advanced AI chips, while Chinese labs have continued to release increasingly capable open-weight models.

Key dynamics shaping the moment:

• DeepSeek’s models have impressed many practitioners globally
• U.S. labs are under pressure to maintain frontier leadership
• Policymakers are increasingly focused on AI supply chains
• Distillation could potentially narrow capability gaps faster

OpenAI’s warning suggests concern that software-level techniques might partially offset hardware restrictions.

Enforcement Is Already Underway

OpenAI indicated it has been actively monitoring and removing accounts suspected of attempting large-scale output harvesting. Reports describe an ongoing “cat-and-mouse” dynamic between platform safeguards and sophisticated access methods.

Actions mentioned include:

• Detection of unusual usage patterns
• Blocking suspected violators
• Hardening anti-distillation defenses
• Increasing scrutiny of programmatic access

This suggests the company views the issue as operational, not hypothetical.

DeepSeek’s Position So Far

As of the latest reporting window in mid-February 2026, DeepSeek and its parent company High-Flyer have not publicly responded to the memo’s claims. Requests for comment cited in coverage reportedly did not receive immediate replies.

Without a direct rebuttal, the situation remains in an early narrative phase rather than a resolved dispute.

What This Does and Does Not Prove

It is important to separate confirmed facts from interpretation:

What is established

• OpenAI sent the memo to U.S. lawmakers
• The company believes distillation-based free-riding is occurring
• Suspicious access patterns were reportedly observed

What remains unproven publicly

• The full technical extent of any distillation
• Whether specific models were materially derived
• Any formal regulatory or legal finding

For now, the episode sits in the realm of serious allegation rather than adjudicated violation.

The Bigger Signal for the AI Industry

Regardless of how this specific dispute evolves, the memo highlights a deeper shift. Frontier AI competition is no longer just about bigger models and faster chips. It is increasingly about data flows, output access, and defensive infrastructure.

If model distillation becomes a central battleground, companies may respond by:

• tightening API controls
• watermarking outputs
• limiting high-volume access
• pushing for clearer legal frameworks

In that sense, the memo may be remembered less for the specific accusation and more as an early marker of the next phase in the global AI race.

For an industry built on learning from existing knowledge, the line between inspiration and imitation is about to be tested far more aggressively.