In a previous article, I posted on the copyright hazards of AI. It is still very much up in the air. While it is sexy to talk about what is broken, the problem still exists that AI can and will move forward. Regardless of the state of the law, this is one technological tool that will never go away. I generally try to stay on target for covering what to do to stay in the clear, and I realize I have missed the mark on that one.
Schrodinger’s Cat of Copyright
We are in a difficult position right now where we do not know where AI will fall on the copyright legality spectrum. We do not know whether existing licenses will cover the previously non-existent use of material for training AI. It is somewhat careless to rely on courts to determine who can use your work, as a rightsholder, though. Licenses for copyright should anticipate and allow or disallow the use of the material. If you are a rightsholder and are explicit that you do not want your material to be used for a specific purpose, then you can explicitly disallow it and not rely on the courts to decide whether or not training an AI is within the scope of the license.
For material that is already out in the wild, whether you can use that or not is still a gray area. If you have a body of material in mind to use, the safe bet would be to contact the rightsholders and make them aware of your use. That contact would give you a good temperature measure as to whether they would want to come after you for infringement if the courts come down against AI use of the material you do not hold a commercial use license for. It may not be a gamble if the rightsholder doesn’t mind, but it would be incumbent upon you to convince them that it is not a harm to their rights. Avoid the argument of law and work on other, more central issues of the material in question.
Moving Beyond Copyright
Copyright is only one of the four intellectual property protections we enjoy in the US. As someone looking to protect rights, it is good to not become too fixed on copyright for AI generated media. The other three forms of IP are patents, trademarks, and trade secrets. Trademark has completely different implications in AI (except for input scrubbing – a la Getty Images marks showing up on output), but patents and trade secrets afford protection where copyright falls short.
Patent
AI generated work on patents are in a funny place, but the Patent Office seems poised to allow AI to be listed as an inventor so long as there is at least one human inventor. The office has never been too hung up on the how of the inventors arriving at their invention, which seems reasonable. But patents are both expensive and have a limited scope in time for their use. The PTO publishes every patent and after the expiration, your work goes into the public domain. I have discussed patents with clients, but the cost, the need to police your own invention, and the limited duration make this a less than desirable form of protection, particularly for the individual or small innovator.
Trade Secret
Unless you intend to produce AI output for entertainment or marketing purposes, trade secret law may be the best law for what you seek to protect. Anyone who has ever signed an NDA (or signed a contract that includes an NDA) should have at least a vague idea of trade secrets, although they may not have thought of that as Intellectual Property protection. This is the most vigorously enforced area of intellectual property law, where if the owner of the secret can show they were careful to protect their secrets, law enforcement can get involved to pursue state and federal criminal penalties for the theft.
The key here is protecting the secret. Fortunately, for data all you need is reasonable computer security and NDAs, including an NDA in any contract for sales of any data you produce. You would need to show that you not only protected the data, but you had procedures for keeping the data safe and strictly followed those procedures. Then again, you should already be familiar with this if you handle or process any personal information.