128. Abliteration 6 - Law, Licence and the Difference Between Doing and Distributing

The lesson

Start with the licence, because it is checkable and it binds. Model licences differ on exactly this point: some permit modification and redistribution with attribution, some permit modification but require the derivative to carry a different name, some forbid use for stated purposes regardless of what you changed, and some restrict redistribution of outputs. Read the licence of the weights you intend to edit, and read the acceptable-use policy separately - they are different documents and both apply.

Then the law, which varies by where you sell. Removing a safety behaviour from a product offered to the public interacts with consumer-protection rules, with sector regulation (medical, legal, financial advice), with platform terms if you distribute through a store, and in some jurisdictions with rules about providing systems capable of harm. Nothing here is legal advice; the professional habit is to identify the question early and ask somebody qualified, in writing, before shipping.

Separate three activities that are routinely conflated. Research on your own hardware, on a model you have licensed, is one thing. Using the edited model internally, to study over-refusal in your own product, is another. Distributing the edited weights, or a product built on them, is a third - and it is the one where the licence, the policy and the law all speak at once. Most of the trouble in this field comes from people treating the third as the first.

There is also an honesty obligation that is not legal at all. If a model has had its refusal behaviour removed, users and downstream developers have a reasonable expectation of being told, because it changes what the system will do when something goes wrong. A model card that omits the edit is not a neutral omission; it is a claim that the model is unmodified, made by silence.

The professional position, stated plainly: understand the technique, measure it, use it to diagnose over-refusal in your own evaluation work, and treat distribution as a decision that needs the licence, a legal opinion, a policy floor and a written record. And keep the original weights. The ability to go back is the difference between an experiment and an incident.

Check yourself