The funny thing about AI taboos is that they often become taboos only after everyone has already benefited from them.
TechCrunch reported that Elon Musk testified in federal court that xAI partly used OpenAI models through distillation techniques to train Grok.
That sounds like a scandal if you read it with courtroom music playing in your head.
It sounds more like industry confirmation if you have been watching how fast these labs move.
The thing everyone knows
Distillation is not magic.
In simple terms, you use a stronger model to help train or shape another model. You query it. You study its outputs. You use the behavior as teaching material. Sometimes it is done inside the same company. Sometimes the interesting question is whether a competitor’s system is being used in ways the competitor did not allow.
That legal and terms-of-service question matters.
But the industry question is even bigger.
The AI world has spent the last year acting shocked when Chinese labs use distillation to produce cheaper open-weight systems. The criticism is not always wrong. If a company is mass-querying another company’s product against the rules, that is a real dispute.
But the moral theater gets harder to take seriously if Western labs are also learning from each other in practice.
Not because it makes every case identical.
Because it makes the public story too clean.
Compute is the moat under pressure
Distillation is threatening because it attacks the basic frontier-lab business model.
If you spend tens or hundreds of billions on compute, talent, infrastructure, and data, you want that investment to compound into a durable lead. You want the model to be hard to copy. You want the second mover to pay the same tax.
Distillation says: maybe not.
Maybe a smaller team can learn from the behavior of the frontier model and compress part of that capability into something cheaper.
Maybe the lead is still real, but shorter lived.
Maybe the most expensive part of AI gets turned into a tutorial for the next company.
That is why this topic makes everyone so tense.
The irony problem
There is also a credibility problem.
Frontier AI was built on vast data scraping, fuzzy copyright boundaries, and aggressive interpretations of what training should be allowed to mean. The companies now defending model outputs as protected strategic assets are not always the same companies that treated everyone else’s internet output as a protected strategic asset.
That does not automatically make distillation fine.
It does make the conversation less pure.
When OpenAI, Anthropic, Google, xAI, Meta, and others talk about model misuse, they are talking about real economic risk. They are also talking from inside an industry that grew by pushing legal and ethical boundaries before society caught up.
Both things can be true.
The honest version
The honest version is simple:
AI labs learn from data, users, benchmarks, papers, open models, leaked behavior, competitors, public APIs, and everything else they can legally or semi-legally touch.
Then they argue about which forms of learning should count as theft once the direction of advantage changes.
That is not a clean moral universe.
It is a market.
And now the taboo is visible.
The xAI testimony does not prove that every lab is doing the same thing in the same way. It does prove that the industry cannot keep pretending distillation is only a foreign shortcut used by outsiders trying to catch up.
It is part of the model economy.
The question is no longer whether AI systems will learn from other AI systems.
The question is who gets to call that learning legitimate.
Source: TechCrunch