The Rentier’s Trap, Or: Why Foundational Models are the Economic Fulcrum of AI

Basic considerations about a fair legal regime for all of GenAI.

The Rentier’s Trap, Or: Why Foundational Models are the Economic Fulcrum of AI
Canonical URL
Do not index
Do not index
AI summary

Introduction

The legal argument that will be contended over GenAI is that, in consuming most of the internet one document at a time, the foundational model builders have not created a derivative work of any single copyright holder but have in effect created a derivative work of all of them, thus setting up a class action; or, and this will surely be the view of big tech, the resulting models are not a derivative work at all, that is, are the result of fair use.

Fair Use vs Derivative Work

I agree with the fair use argument. Not because of any arcane reading of copyright law but for policy reasons: it’s better for society that foundational models are fair use. That is the only path forward to collective flourishing, however angry it may cause—wait for it—Sarah Silverman and Stephen King to become.
But what’s sauce for the goose should be sauce for the gander, too.
That is, the rest of us should have the same liberal fair use rights with respect to the foundational models themselves as they will surely claim to countless blog posts, technical reports, textbooks, etc.
Using, say, a distillation technique, I should be able to train n student models from, say, Falcon or Dbx etc. teacher models and that use should be considered fair under ordinary copyright law.

Justice as Fairness: Treating Like Cases Alike

And in a fair, progress-oriented legal regime, one not captured by big tech rentiers, by that effort I will no more have violated fair use than they violated individual copyright holders by creating the models in the first place. It’s not after all prima facie clear that—by reading many documents, each act of which in seriatim, is wholly consistent with ordinary fair use—they must have, in aggregate, created a derivative work.
What is intolerable is a legal result that offers big tech a liberal fair use ruling with respect to individual copyright holders while also offering big tech a strict, derivative-work interpretation of its IP.
Except that too much of the legal regime is the result of regulatory capture of exactly that type.
 

Voicebox talks for you and your data talks back! Grow your business with faster time to insight.

Stardog Voicebox is a fast, accurate AI Data Assistant that's 100% hallucination-free guaranteed.

Signup