AI and Hollywood Accounting
This week Harvard Business Review published an article by E. Glen Weyl and Raul Castro Fernandez, How AI Companies Can Pay Fair Rates for the Content They Need. As a publisher and author, it’s a subject that’s close to my heart, and I was glad to see HBR addressing it. I agree completely with the authors that both AI companies and authors will benefit from the creation of a sustainable market for content for training purposes—in fact, I’d go further and suggest that society as a whole benefits from that. I also concur with their view that collective management organisations (CMOs) have an important role to play in distributing money to publishers and individual authors. But to me there is a fundamental, fatal flaw in their argument that needs to be called out.
The article suggests that publishers and authors should not be compensated from AI companies’ revenues, because this ignores the costs of running models. Instead, they suggest that compensation should be a share of operating profit on a per-model basis, which distributes both upside and risk.
That doesn’t stack up to me if one thinks of content as one of several critical inputs to the operation of an AI model. It cannot operate without training data or resource inputs like electricity and water. No-one believes that the power company should share in the risk of the model, and it is not clear to me why content should be treated so differently. A company that can’t pay its electricity bill isn’t viable, and if paying authors on that basis compromises the business model, maybe it wasn’t such a great one in the first place.
More broadly, the authors compare their proposal with the model of Hollywood studios granting contributors in a film a share of profit. Of course, this can be fantastically lucrative: Sir Alec Guinness famously insisted on 2.25% of the gross from the first Star Wars film, an arrangement that reportedly made him nearly $100 million in his lifetime. But he had the balance of power in negotiations, and was shrewd enough to insist on points on the gross, not the net. In the HBR example, the proposal is operating profit rather than gross profit. The lower end of the income statement is subject to all kinds of accounting variances to reduce the amount of profit a given film (or in this case, AI model) makes. It’s a sufficiently well-known practice in the movie business that there is a very entertaining Wikipedia page for Hollywood accounting: mundane, entirely legal ways to turn a monster hit into a paper loss. Forrest Gump famously grossed nearly $700 million and pencilled out at a loss of $62 million, leaving author Winston Groom with nothing for his 3% net profit share (reader, he sued).
Now transpose this to AI and think about the drivers of operating profit. The biggest cost is probably compute, often bought from a cloud provider that is one of the model-developer’s investors (think Microsoft and OpenAI, or Amazon and Anthropic). The price of that compute is set between the parties, not in an open market. That exact structure—an affiliated supplier charging the producer an internally-determined fee—is one of the key levers for Hollywood studios to shift profit margins. And it is only one example: what proportion of SG&A, R&D, safety or infrastructure costs should a model carry? Each of those is an accounting judgement, profit is a construction and an AI lab’s cost base is arguably far more opaque than a studio’s.
A sustainable market for content as training data is desirable and, I hope, inevitable. CMOs are the right structure for distributing the proceeds. But the whole thing has to be anchored to revenue so input costs are recognised for what they are and so authors and publishers can have confidence in the system. The authors would object that a revenue basis falls hardest on open-weight and thin-margin developers, who earn little revenue relative to the value they create. That is true, but it is simply what it means to treat content as an input rather than a bet: a model that can only exist if its core inputs go unpaid is not, in the end, a viable one. Authors and publishers have spent enough time on the wrong end of the AI value equation, and while it is good that HBR is highlighting the issue, it would be a shame to design further inequity into the system intended to put things right.