Even though you’ve seen it a million times already, we can’t illustrate this story with Shutterstock photo number 297886754. The image (best known as the distracted boyfriend meme) is owned by photographer Antonio Guillem and is not included in our Shutterstock licence.
What we can do is feed Guillem’s photo into the Midjourney AI image generator, which creates a limitless number of distracted boyfriends:
Is the above a breach of copyright? Well, who knows. Creating derivative works from a protected image seems intuitively wrong on two levels, but technology has outpaced the scope of laws as well as the capacity for artists to pursue legal action.
For Nasdaq-listed Shutterstock, this legal grey area has become central to its business. The company licenced its stock image library to OpenAI in 2021 (though the deal was only announced last October, in combination with a revenue warning). Earlier this year, Shutterstock added text-to-image generation to its website using a version of OpenAI’s Dall-E 2 tool.
Jarrod Yahes, CFO of Shutterstock, stopped by the offices of Redburn this week to explain. He argues that Shutterstock offers its customers commercial indemnification, effectively giving a guarantee that their AI generations will contain nothing dodgy.
Redburn analyst Nick Delfas says:
The longer-term threat is that AI technology will completely eliminate the need for stock photos and steal Shutterstock’s customer base. So far Shutterstock has seen the reverse – the demand for licensing of its content is accelerating with technology companies seeking more data access and increasing their content budgets as the commercial uses for AI rise. Training of AI models is not a finite process – to stay relevant the machine will require ongoing access to large volumes of fresh data.
Furthermore, we consider it unlikely that Shutterstock’s role as an aggregator of a broad range of content types with legal copyright protections could be easily replaced. Customers value the legal safety of using Shutterstock rather than a platform that might be ingesting images it has not correctly licensed.
More than 8mn AI generated images were added to Shutterstock’s library in the first 45 days after Dalle-2 was added to the website, Redburn says. Humans upload about 50mn photos a year to Shutterstock so, at the current run rate, approximately three AI images are added for every two conventional ones.
And increasingly, the image library is used for AI training rather than visual decoration. Large-language-model and neural network developers put a high value on third-party databases of standardised, censored and sanitised content that have detailed text descriptions attached. That means a few customers have been paying a multiple of the median contract value: Shutterstock’s average deal size went from $22k in 2020 to $310k in 2021, and to $1.3mn in 2022.
Approximately four-fifths of the value of Shutterstock’s AI contracts is booked upfront, with the remainder recognised over the (typically five year) contract life as new photos are uploaded. AI trainers pay almost nothing per image but Shutterstock has a contributors’ fund that bumps up royalties to an average rate, so for the moment the arrangement is gross margin neutral.
Photographers are still boned, of course. Not only does robot competition cannibalise their marketplace royalties, the one-off compensation payments Shutterstock offers are distributed pro rata to library size rather than sales. In effect, the value of Distracted Boyfriend to a bot trainer becomes the same as Circumcised Banana or Futuristic Woman Injects Corn.
It’s a model Shutterstock likes. The company this week announced a partnership with Nvidia to develop an AI-powered 3D model generator, and has been working with Meta and LG on projects involving AI-generated stock video and music.
In each case, the product is the safety of familiarity. Redburn’s Delfas writes:
Shutterstock acts as a connecting point between a large volume of content contributors and a large volume of advertisers. The value it creates sits in the volumes of sales it facilitates — reflected in its high 70 per cent gross margin.
Marketers coming to the platform are likely to be agnostic to the source of content that meets their requirements — it could be taken by a human or generated by a machine.
Having a broad library of content with an integrated AI-generation tool provides Shutterstock’s customers with a higher level of assurance that they will be able to find the content they need. The risk of churn from customers migrating to a competing AI-only tool seems unlikely. It is therefore unlikely that any AI platform will be able to replicate the customer traffic and volume of sales offered by Shutterstock (or Getty Images).
…against which we offer a few counterpoints.
What happens when those landmark deals with OpenAI and Meta expire? Would they be renewed on similar terms, or on the much lower annual rate they currently pay for their drip-feed updates? Since the main value of the library has been sold upfront, how much negotiating power does Shutterstock have left?
Is the value of the library being protected? Paying creators by the yard has potential implications for quality, as does the generation of AI content from AI content. Potentially, when viewed in terms of clean data, the setup is somewhere between the horsemeat scandal and mad cow disease.
What is the guarantee of commercial indemnification worth in practice? Shutterstock classes itself as an internet service provider under the US Digital Millennium Copyright Act. Contributors to its marketplace are bound only by a terms-and-conditions form and takedown requests involve a DCMA notice. Given Shutterstock effectively disclaims responsibility for the nature of the content it sells, is its promise of a ethical AI any more than a placebo?
Historically, Shutterstock has made more revenue from individual users (20 per cent) and SMEs (30 per cent) than from large enterprises (19 per cent). Should these customers be worrying en masse about their as-yet-undefined exposure to litigation from using generative AI? Until intellectual property law catches up with events, we simply don’t know.
And in the meantime . . .
— Art and artificial intelligence collide in landmark legal dispute (FT)
— Thirty AI-generated imaginings of what banks might look like (FTAV)