Copyright shield – are generative AI tools now risk free?

The recent board-level wranglings at OpenAI have been well publicised. Those following developments will have seen Sam Altman ousted as CEO, only to be reinstated at the time of writing.

Altman announced, in what appeared for a short while to be one of his final public appearances, a new ‘copyright shield’ that would be extended to OpenAI’s paying users. Altman’s statement was subsequently confirmed with the following blogpost:

‘[OpenAI] will step in and defend our customers, and pay the costs incurred, if you face legal claims around copyright infringement. This applies to generally available features of ChatGPT Enterprise and our developer platform.’

Legally speaking, the ‘copyright shield’ should be understood as an indemnity, and OpenAI are not the first in their field to offer such protection. Microsoft has implemented a ‘Copyright Commitment’, and Google Cloud it’s generative AI indemnification.

Also of note are the indemnities offered under Adobe’s image generation tool Firefly and Getty Images’ ‘Generative AI by Getty Images’, both of which promise ‘commercially safe’ usage (though, having only been trained on each company’s respective image libraries, these tools should be distinguished as the risk of third-party infringement claims would appear lower).

OpenAI’s Terms of Use

Under the updated Terms of Use, if the indemnity applies to a particular claim, OpenAI will cover damages awarded against their customers or the settlement amount agreed. Such payments would also not be subject to OpenAI’s usual liability cap.

On top of the expected exclusions (ie invalidation due to modifications to OpenAI’s products, etc), it is noteworthy that indemnity will not apply where a claim arises from the following circumstances:

1. the nature of the customer’s input data or any training data provided to OpenAI;
2. if the customer has combined OpenAI’s products with non-OpenAI products; or
3. if the claim would not have arisen but for the customer’s application.

Point 1 makes it clear the copyright shield will not apply if OpenAI can show the input materials themselves infringed a third party’s copyright or those materials were not properly licenced.

On point 2, what is meant by the ‘combination’ of OpenAI products with other products is unclear, though on first impressions it would appear to provide OpenAI with a potentially very wide exclusion (not least because OpenAI’s paying products are focused on integration with customers’ own platforms). The ‘but for’ exclusion under point 3 is also potentially very wide and could be applied to any number of claims based on the customer’s application and seemingly without the need to show the customer knowingly used its application in a manner that would infringe third-party intellectual property rights.

Commentary

It is questionable whether the ‘copyright shield’ goes much further than the protections currently provided by Microsoft and Google Cloud. What is now clear is that all three companies will (subject to exclusions) cover any claims based on their own training data.

By way of example, were a user to prompt ChatGPT to produce a country song in the style of Johnny Cash, it would be OpenAI and not the user who would be responsible for paying out on claims simply by virtue of the user a) using the tool, b) making the prompt, or c) publishing the output materials.

That OpenAI, Google Cloud, and Microsoft are willing to indemnify their users can be read in several ways. It could speak to a collective confidence that their training data sets are fully licenced and will not be exposed to third-party claims. It could also be said OpenAI, Google Cloud and Microsoft have taken a view that using their tools will not in itself constitute an act of infringement. The collective move to indemnify customers could also be seen as a commercial calculation by OpenAI, Google Cloud and Microsoft to overcome any residual caution surrounding intellectual property infringement thereby removing a final barrier to the wholesale uptake of their tools by commercial users.

Until there is an authority on whether generative AI tools infringe intellectual property rights when they are fed training data, the extent of the risk OpenAI, Google Cloud and Microsoft have promised to indemnify against remains unclear. We hope Getty Images’ claim against Stable AI for infringement of its image library and trademark and Universal Music’s claim against Anthropic will offer clarity in this regard.

Nonetheless, in the context of OpenAI, Google Cloud and Microsoft as providers of generative AI tools, a firm benchmark has been set for the protection of customers using tools trained on providers’ own data sets.

This article was first published in Tech+, a newsletter from our tech and innovation team designed to help readers unpack complex topics in the tech space and keep up-to-date with the changes across this rapidly evolving sector. Be the first to receive the next edition and subscribe here.