DALL-E 2 Indemnification

Now that DALL-E 2 is commercially available, perhaps, it’s worth thinking about legal indemnification. If you don’t know what that is, GPT-3 describes it as:

Indemnification is a legal principle whereby one party agrees to compensate another party for losses or damages incurred as a result of some action.

In other words, say Acme Inc. bought an image using their DALL-E credits and got sued for content in the image, indemnification would mean OpenAI would protect Acme Inc. from any legal damages up to a certain amount and cover these costs on their behalf. Indemnification is a selling point to encourage buyers to use stock photography platforms worry-free, compensate any accidentally infringed parties, and also puts a focus on the stock photography companies themselves to have rigorous legal vetting processes before adding new content to the site.

Other websites offering stock photography content already have legal indemnification policies. Here are some snippets from storyblocks, which is a service I use:

I’m sure OpenAI has considered this or is already talking about it. Many people have concerns with using DALL-E 2 generations at a very commercial, professional scale and will likely be looking for legal coverage for aspects like safety, copyright, brand infringement, and much more. From what I’m hearing, serious commercial concerns are related to the unpredictability of the AI model itself - it could put anything somewhere in the image, creating problems for large creative organizations.

In the past, I have argued for GPT-3 insurance, which I don’t really think is necessary as much anymore, but maybe indemnification is something OpenAI could bundle in down the road to ease the anxiety of very large corporations. To be honest, DALL-E 2’s current state even now seems fine to me, but maybe indemnification could be the, “cherry on top” that gets large orgs over the fence as major customers.

Regardless, my hunch tells me OpenAI will take a broader alignment approach to all of these problems like they did with GPT-3, but I still think it’s worth capturing this indemnification idea here to get the conversation going.