Anthropic Cannot Have It Both Ways on Distillation
Oscar Gallo
Published on July 3, 2026
If AI distillation is wrong, the rule has to apply consistently. It cannot become wrong only after frontier labs build their lead.
The AI industry has a consistency problem.
For years, the most powerful labs benefited from a loose interpretation of learning. They trained on public data, absorbed patterns from the open internet, used synthetic data, and built products on the idea that large-scale learning from available outputs was acceptable.
Now that competitors are learning from their models, the tone has changed.
That does not automatically make Anthropic wrong. It does mean the argument has to be cleaner than "it was fine when the frontier was being built, but dangerous when the frontier is being copied."
A rule has to work in both directions
If distillation is theft, say so clearly.
Say it when a Chinese lab allegedly distills Claude. Say it when an American lab trains on public code. Say it when a startup uses synthetic data from a larger model. Say it when frontier labs scrape human work to build commercial systems.
Maybe the AI industry should have stronger rules. Maybe model outputs should be protected more explicitly. Maybe training rights should be licensed, priced, and tracked with the same seriousness as software dependencies.
That is a coherent position.
What is not coherent is changing the moral category once incumbents become vulnerable.
The ladder problem
Every technology market eventually faces a ladder problem.
The early winners climb using a set of conditions that favor speed: open data, permissive norms, cheap access, weak enforcement, and ambiguous rules. Once they reach the top, they discover the value of control. Suddenly, the same openness that helped them rise becomes a threat.
That is when the ladder gets pulled up.
In AI, the ladder is data and model behavior. Frontier labs learned from the world. Now they want to limit who can learn from them.
Again, there may be good reasons for limits. Fraud, fake accounts, terms-of-service violations, and national-security concerns are real issues. But those issues should be named precisely. They should not be blended into a broad claim that distillation itself is uniquely immoral when challengers do it.
Open source is not the enemy
The danger is that fear around distillation becomes an excuse to attack open models broadly.
Open-source AI is messy. It can be misused. It can also be audited, adapted, run locally, improved by outsiders, and used by builders who cannot get invited into a closed frontier preview.
If the policy answer to model extraction is "only a handful of approved labs can build powerful models," then the cure may be worse than the disease.
The industry needs controls around abuse, not a blanket war on openness.
Buyers should care about consistency
This may sound like a fight between labs, but buyers should pay attention.
Vendor ethics become vendor risk. If a company changes its principles depending on market position, customers should expect unpredictable policy, pricing, and access decisions later.
The same applies to provenance. If a vendor cannot explain what it trained on, how it uses customer data, whether it trains on outputs, and what rules it expects from others, the risk does not stay inside the vendor. It travels into your stack.
Consistency is not just a virtue. It is an operational requirement.
Builders should plan for access shocks
The distillation debate is part of a larger pattern: frontier AI is becoming more restricted, more political, and more defensive.
That means builders need a backup plan.
Do not build a product that only works if one lab keeps one model available at one price under one policy. Use evals. Keep fallback models. Design harnesses that can swap providers. Know which parts of your product depend on frontier capability and which parts can run on cheaper or local models.
The more political the frontier gets, the more valuable portability becomes.
Bottom line
Anthropic can argue that Alibaba broke rules. It can argue that fake accounts and systematic extraction are unacceptable. It can argue for better technical and legal protections.
But it cannot credibly argue that learning from outputs is sacred when incumbents do it and theft when challengers do it.
The AI industry needs a rule that applies before and after the money arrives.