Eyes Above The Waves

Robert O'Callahan. Christian. Kiwi. Software guy. All content human-authored unless otherwise noted.

robert@ocallahan.org ... Follow on X, Follow on Bluesky

Wednesday 18 November 2026

The AI Credit Laundering Problem

AI is interacting with copyright law in interesting ways, so that most new software is probably not copyrightable, in the USA at least. But there is an even bigger and more general problem, which is that AI has broken the informal systems we have always used to assign credit for intellectual work.

At Google we had a situation where person A developed system X the hard way, then person B used AI agents to develop system Y to solve the same task. System Y gave better results than system X, and included a number of techniques similar to those present in system X. The AI agents had access to system X, both directly via tool calls that retrieve code, and because system X was in the training data. The question is, should person A receive some of the credit for the results of system Y? In the pre-AI days, person B would have had a good understanding of how much they had learned from system X and, being a reasonable person, would have assigned credit to A accordingly. Now they probably just don’t know how much credit to assign. Person A will justifiably feel aggrieved.

The problem is even bigger than this. In the real case I’m thinking of, A and B were known to each other, but in general they might not be; person B might never know system X existed. In the limit, it seems fair to say that everyone who contributed to the training data or whose contributions might have been accessed by the AI agent should be eligible for credit.

This is a severe problem. It has come up forcefully in mathematics, where AI-powered proofs seem inspired by specific results from specific human mathematicians but we don’t know for sure.

It may be possible to analyze transcripts, tool call logs and (if available) reasoning tokens to do some credit assignment. That would be a pretty interesting problem to work on! But it certainly wouldn’t capture learning from pretraining, so it won’t solve the problem.

We certainly need to acknowledge the problem with humility: in general we don’t know where credit should go.

I think we probably need to take the expansive view above, which generalises the “AI-generated code is not copyrightable” position: when you use AI to solve a problem, most of the credit for that solution diffuses out to everyone who contributed to the training data and other accessible resources. When AI proves mathematical theorems or generates code, most credit should not go to the people commanding the AI, nor the AI itself, but to the large fraction of humanity (living and dead) on whose collective works the AI was grown.

This also suggests an economic fairness principle: because it’s so much harder to assign credit to specific antecdents of AI-generated work than for human-generated work, we should tax AI work for the benefit of the public more than we tax human work. And of course, in general, we should not allow companies to distill the public intellectual commons into their models and capture all the profits for themselves.