No One Has Been Paid for the Text That Trained the Models

The oddest thing about the biggest fight in AI might be that nobody has been paid anything yet. The models were built on an enormous corpus of copyrighted work, and the price of that corpus is, right now, zero — because no one has paid for it. The Authors Guild says that is not an accident but the point: it frames the generative-AI boom as built on copyrighted works "without licenses — that is, without giving authors any compensation or control" over their use. That is not a damages complaint so much as a description of the cash-flow structure, and the structure is that authors are on the wrong side of a transfer they never agreed to.

The basic point is that this looks much more like a tax on the people who wrote the training data than a licensing market. The companies took in the copyrighted corpus at a license cost of zero; the money that would have gone to authors instead stayed inside the model builders' own capital-expenditure stack. The Authors Guild and the named writers who joined it hold no contract with OpenAI and no ongoing compensation stream. Their one enforceable lever is a copyright-infringement suit asserting an uncompensated taking — a claim that a copy was made and nothing was paid for it.

Follow who actually holds the bargaining power on that boundary and the whole dispute snaps into focus. The firm controls access to the training data it already copied at no cost; the author's specific work may not be individually necessary to any general-purpose model — no one can point and say "this book is what did it." So the author's individual claim to value is weak, and the only force that can concentrate it is collective: a class action, waged through the courts rather than through any market that currently exists.

That is why the procedural details are worth watching as closely as the legal theories. The Authors Guild first filed its class action against OpenAI in September 2023; on December 4, 2023 it filed an amended complaint adding Microsoft as a defendant. The move up the chain is the tell. Microsoft is not a training-data consumer in the same way OpenAI is — it is the capital and distribution behind the product — and adding it extends the uncompensated-use claim from the company that built the model to the company that pays for and sells it. That is a measure of how far the suit thinks the liability runs.

An investor valuing an AI lab should read this as an open, unpriced line item, not a settled one. There is no licensing payment, no damages award, and no settlement establishing anything yet. That means no licensing benchmark exists that anyone could use to model the expected cost — and if the unlicensed-copying claim prevails or resolves into a settlement, a retroactive license or damages cost lands on top of training capex that currently carries no such line. You cannot price a liability with no disclosed magnitude and no comparable transaction. The suit is the pricing event in waiting; so far it has priced nothing.

So the honest position is not that litigation will set a price — no settlement or damages figure exists to support that — but that no price exists yet at all, and the authors' ability to collect anything runs through the courts rather than through a market. Whoever holds the bargaining power on that unsettled boundary decides whether a licensing market ever forms or whether the transfer stays unilateral. Do not forecast the outcome; just watch for a ruling or a settlement, because that event, not the training data itself, is what would finally put a number on it.