NBER paper: AI answer systems could push the open web subcritical

A new NBER working paper puts a formal frame on something publishers have been complaining about all year, that AI answer engines keep the reader and quietly starve the source. In w35344, published this June, Alex Chan argues that an AI platform which underinternalizes future content reproduction "retains too little referral traffic and can make costly open-web information subcritical, even with truthful content." The word to notice there is subcritical, meaning below the threshold where the open web keeps regenerating itself.

The mechanism Chan lays out is self-reinforcing. AI answer systems "can improve user experience while diverting visits that finance publisher content and generate source-level quality signals." As those visits and signals dry up, conventional search gets worse at telling good sources from bad, users lean harder on the AI, and the loop tightens. It is a theory paper, not an empirical estimate, so read it as a mechanism you should now expect to argue about rather than a measured collapse.

Chan's proposed remedies are more concrete than most of what has been floated in the press. He points to "visitor-replacement royalties, audited provenance, human-information audits, and keystone-topic compensation." Translated: pay publishers for the click you swallowed, prove where a fact came from, audit whether humans still generate the underlying knowledge, and reserve special treatment for topics where the loss would be societally expensive.

The honest caveat is what a paper like this does not give you. There are no numbers on how far along the decline actually is, no ranking of which categories tip over first, and no sizing of what a visitor-replacement royalty would have to be to keep a marginal publisher solvent. Those are the questions the empirical follow-ups will have to answer.

What it does do, if the argument holds, is give regulators, licensing negotiators, and platform product teams a shared vocabulary. "Underinternalized content reproduction" is now a phrase to argue about, and "visitor-replacement" is a cleaner ask from publishers than a vague piece of the LLM. That is the useful thing about theory papers arriving at the right moment.