6

In simple terms, any AI system will consist of three elements:

  1. Code, written by humans, that defines the mathematical model
  2. The mathematical model itself, as it exists within the computer
  3. The model weights, millions or billions of parameters that are tuned by the training process

As far as intellectual property law is concerned, I think I know how the first two would be handled:

The code is a creative expression by the coders, and so copyright would apply.

The mathematical model is defined by the code, but can be expressed in other ways. Here is a diagram of the structure of such a model, that along with the paper it is in should allow one to recreate the model without creating a derivative work of the actual code that was used to create it in the first place. I THINK patent law not copyright would apply to this element.

Image of AI model used for cell types

The weights I am not sure of. These have been generated without any human creativity, the model has been run against large amounts of data and the weights built up. This is however a very expensive process, costing lots of computer time with means electricity. Would there be any intellectual property law that applies to these weights?

This is particularly relevant now as the 65B weights for Facebook LLAMA have leaked. My understanding is that Meta has released the code under a non-permissive no commercial use licence, and only released the trained weights for much smaller model. This leak allows anyone technically to run the full model (if they have a $20k graphics card, or do it "on the cloud"). However could they do it legally? Assuming they could recreate the model without creating a derivative work (perhaps using clean room techniques ) could someone use this model without restrictions?

2
  • note: "the trained weights of an AI model", in popular parlance, would just be called "an AI" Commented Mar 13, 2023 at 20:35
  • 1
    No, you need the model code too.
    – Timmmm
    Commented Mar 17, 2023 at 15:22

1 Answer 1

3

There might be patents to the methodology used to get the AI to learn

A methodology can be patented, and thus limited in who may use this particular way.

The base code is copyrighted

The base code of the AI, which makes the computer evaluate different outcomes and test the inputs against its outcomes is copyrighted by the making company.

The pure product of AI is uncopyrightable and unpatentable

The US Copyright office declared that AI products do not have an author. As such, AI-generated forms of expression, including algorithmically generated weights, do not benefit from any copyright.

Likewise, the USPTO has declined multiple applications in which an AI was listed as the sole inventor. Their current stance is, an AI can't be an inventor, if anything the human feeding the AI could be one. This is based on a 2012 case, in which the SCOTUS decided that an individual is always a human.

As such, AI-generated weights cannot be copyrighted or patented.

The weights might be trade secrets

It's unclear if the weights are the product of a human maker using an AI as a tool, or a pure AI work. As such, their copyright status is unclear.

However, they might be protected as trade secrets, and as such disclosure, acquisition and usage of them could be illegal. Trade secrets and their products are generally licensable.

10
  • 1
    Thank you for this, but it seems to cover only the code and the model, not the weights which I was asking about.
    – User65535
    Commented Mar 13, 2023 at 20:58
  • 2
    If the model weights are uncopyrightable, doesn't that make the supplied license invalid? Commented Mar 13, 2023 at 21:42
  • 2
    @Trish I think you misunderstand the question and neural networks/AI. I write some NN code. That's copyrightable/patentable. The NN has internal parameters, called weights. I train the NN by putting in some known data (this is a picture of a cat, this is a picture of a dog), and the training process sets the NN weights, sort of similar to curve fitting. Then I ship the NN with fixed weights to end-users who put in commands like "draw a black cat driving a car". That cat picture output is uncopyrightable, yes, but the question is about the weights. (cont'd)
    – user71659
    Commented Mar 13, 2023 at 23:02
  • 2
    @user71659 the problem with the weights is: they are not the pure product of AI. They are the output of an AI fed with a deliberate database. Copyright status of that is not tested.
    – Trish
    Commented Mar 13, 2023 at 23:45
  • 2
    Often there would be an attempt to protect the weights as trade secrets which is a factor that contributes to setting up "black box" models that do not transparently provide those weights to users.
    – ohwilleke
    Commented Mar 14, 2023 at 0:02

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .