IP Tech Knowledgy

Whether AI infringes on the copyright of the underlying works it “learns” from remains uncertain, although a few trends are developing. Some courts have determined that AI is allowed to use underlying works, and some cases have determined that the usage is copyright infringement. The differing facts in each case have led to different results.

The information on which AI websites and other uses base their body of knowledge is often information protected by copyright, whether it is contained in books, newspapers, scientific journals, websites, databases, or other media. Under copyright law, the creators of those works  have the right to control who reproduces, distributes copies of, performs, displays or makes derivative works of their material. However, copyright law also allows a third party to make a “fair use” of underlying materials without needing permission.  

Of the few AI cases decided so far, meeting the fair use standards was a key factor in determining whether the AI company was found to be infringing. Not all cases have proceeded to the final ruling stage, although in interim procedural decisions, such as Summary Judgment, courts have decided whether cases can proceed based on an analysis of copyright liability.

In a situation where AI was generating new material based on the underlying material (“generative use”), a court determined that the use was allowed due to the output being “transformative” fair use, because the new output serves a different purpose than the original work was intended to address. Conversely, even in a case where the judge found the use to be transformative, he cautioned that if the AI’s output is similar enough to the copyrighted works so that it might reduce demand for the original, (“market dilution”) the AI company could be found liable. In a case where the AI company trained its system on a collection of pirated books, the case was not dismissed and is proceeding to trial, because the piracy outweighed the transformative use. When the AI creates answers to questions, based on the copyrighted information it trains on — rather than being a generative use  — a court found copyright infringement, because the new work was not transformative, and it created market dilution.  
 
All cases so far from which a court has issued a written opinion have been at the trial court level. Neither appellate courts nor the Supreme Court have had the opportunity to determine the parameters under which it is a copyright infringement to train AI systems on underlying work. Three cases are currently pending at the appellate level. 

It is still too early in the evolution of the jurisprudence to be certain how a court ultimately will decide in any particular case relating to AI and copyright infringement. This is why some high-profile cases have settled. This is also why entities that use AI generated work should look at their contracts with their own customers and address the extent to which they are warranting that all of the information is accurate and not infringing on third parties.   
 

Ned T. Himmelrich
410-576-4171 • nhimmelrich@gfrlaw.com