Is it legal to license and sell the output of a neural network that was trained on data that you don't own the license to? For example, suppose you trained WaveNet on a collection of popular music. Could you then sell the audio that the WaveNet produces? There are copyright restrictions on using samples to produce music, but the output of a generative neural network might not include any exact replicas from the training data, so it's not clear to me whether those laws apply.

  • $\begingroup$ I think you'll need to ask a lawyer. Intellectual copyright issues for music are far from trivial, and sometimes get decided counter to intuition. $\endgroup$ Commented Sep 26, 2020 at 17:09
  • $\begingroup$ Thanks for the pointer @RobbyGoetschalckx. I'm new to this space, and I'd assumed there would a lot of questions and common knowledge around this topic, but I can see why it might be nuanced. $\endgroup$
    – Robz
    Commented Sep 27, 2020 at 4:10
  • $\begingroup$ Great question!!! I'm glad you asked it here because it is fairly specialized. I've posted a comprehensive basic answer, and will likely come back to amend and clarify, so feel free to post any followups in comments. $\endgroup$
    – DukeZhou
    Commented Oct 3, 2020 at 0:07

1 Answer 1


Disclaimer: I am not an attorney and this does not constitute formal legal advice.

  • If the output is novel the copyright resides with the creator

In this case, almost certainly the human who utilizes the algorithm†. There was a recent US patent case "Dabus" [U.S. Patent Application No.: 16/524,350] where the human programmers tried to claim an AI as an inventor, which was rejected by USPTO. This is largely because an inventor is defined as a "natural person". But it's an interesting challenge to the notion of inventorship and authorship.

An argument regarding output specifically as a salable commodity would be that most human creative endeavor is result of previous work that forms the basis for the novel arrangement of elements in the form of an original work.

Music is more specific because melodies are mathematical, and what are traditionally protected, although Blurred Lines case blurred this line, in that the ruling there was the production aspect, the "feel", was what was plagiarized.

And, in fact, algorithmic music generation, which is heavily utilized in pop music, depends on combinatorial functions producing novel sequences, not protected by prior copyright.

Samples are clear cut—it is excerpting from copyrighted work. A wave form is then copyrightable, in that a work of recorded music is simply an arrangement of waveforms in some combination and sequence.

If the output is procedural, and a process that can be automated, that would enter the domain of patent law, utility patents specifically.

Design output would also be in the realm of design patents, which are specific/generalized arrangements of elements in products that are not processes.

A note on liability:

Law is a process, where precedent plays a major factor. Because law uses natural language, there is ambiguity, and jurisprudence is the process of clarifying the meaning and application via challenges. Especially when in uncharted territory "no one knows" until a case has been ruled on, and rulings can be challenged all the way up to the Supreme Court, which may or may not accept a suit.

Damages for copyright are 3x, but damages can be very hard to prove. Intellectual Property litigation is also enormously expensive, thus unlikely to be pursued unless there is a potential financial benefit. (The first step in any potential copyright violation is typically a "cease and desist" letter, with no legal action if acceded to. In a case such as "Blurred Lines", where there is significant financial return to the alleged infringer, the award or settlement presumably exceeds the cost of litigation.)

However, deep-pocket players can use litigation, or threat of litigation, to disincentivize competitors, and, if the target has limited resources, can produce the desired outcome regardless of the strength of the claim. (It's not uncommon in law in general to file frivolous suits as a strategy, although there is typically a financial penalty such as reimbursement of the defendant's legal fees if the suit is found to be frivolous and dismissed.) "Patent trolling" became so much of a problem before 2013, the entire philosophy of patentability, what is patentable, had to be revisited.

†Who has rights to an algorithmic process is also a legal question.

Software is patentable, being simply a type of computer, regardless of medium, but most software is not patented, and protected instead as "trade secrets" and via non-disclosure agreements. (This may be due to the glacial pace of the patent process compared to the software development process, but also per the from the necessity of making patents public. i.e. if Google patented their search algorithm, it would be an instruction manual on how to exploit it.)

Software is copyrightable as the specific lines of code. Copyright naturally resides with the creator, but the process of copyright is be utilized to document the right against potential infringement.

  • 1
    $\begingroup$ Stanford did patent the google page rank scheme patents.google.com/patent/US6285999B1 and patents.google.com/patent/US6754873B1 includes equations! $\endgroup$ Commented Oct 3, 2020 at 0:39
  • $\begingroup$ There may be a can of worms here: Rights inhere to the creator, but the creator is an algorithm, and legal rights can only be afforded to natural person. $\endgroup$
    – DukeZhou
    Commented Oct 3, 2020 at 1:48
  • 1
    $\begingroup$ The patents George posted are a good example of the type of novel software functions one might want to patent, in that they are likely "basic patents" (patents subsequent patents build on.) In this case, the patent excludes competitors from using the identical process, where alternatives may be less efficient (greater Kolmogorov complexity.) $\endgroup$
    – DukeZhou
    Commented Oct 5, 2020 at 0:21
  • 1
    $\begingroup$ The first one US 6285999 has been cited by over 850 other subsequent patents or applications according to google patents. $\endgroup$ Commented Oct 5, 2020 at 2:20

You must log in to answer this question.

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