Generative AI reached 72% of organizations before copyright law settled who owns the output

Copyright starts with a human creator. A prompt supplies thin material for one. The user provides the idea, and ideas are generally not protected by copyright, Morey Haber writes ^[15]. Enter a basic prompt, accept what comes back, and there may not be enough human creative contribution to establish copyright in that output ^[4]. Substantially select, arrange, rewrite, edit or transform the material, and those human contributions may qualify ^[5]. Haber does not say how much editing is enough.

A licence settles what two parties may do with each other's property, not whether copyright exists in the first place. "A platform's terms of service cannot create copyright protection where the law does not recognize it," Haber wrote ^[6].

So a company shipping AI-assisted work has two separate problems. One is that it may not own the asset it paid for. The other is that permission from the provider does not necessarily mean the output is clear of third-party intellectual property or copyright risk ^[8]. Haber's example is a corporate mascot generated to look suspiciously similar to an existing copyrighted character, where the provider's terms may well permit use of the image ^[9]. The character's owner stands outside that agreement.

The population facing the question grew quickly. Half of 72 is 36, so "nearly double the share two years earlier" places the earlier figure a little above 36% of organizations, and the two-year increase a little under 36 percentage points ^[14]. The same survey shows 28% reporting no generative AI in any business function ^[13]. "We invent new capabilities before we completely figure out the rules," Haber wrote ^[12].

Registration is what a company needs when it wants to stop a competitor reusing an asset it commissioned. In the mascot example, the third-party claim arises whether or not anything was ever registered ^[8]. The record for all of this is one contributed column on Forbes' Technology Council, written by BeyondTrust's chief security advisor, an identity and technical evangelist with more than 25 years in the IT industry ^[11].

This quarter will end with the training-data question still open. Courts, regulators, creators and technology companies continue to grapple with when training constitutes permissible use, when licensing is required, and whether outputs can infringe existing works ^[10].