Darden scholars propose annual AI model releases when training data is unlicensed

AI companies could keep a commercial head start, but not permanent control over models built from unlicensed material. That is the bargain Mike Lenox and Alex Cowan proposed on November 12: require annual public release of those models’ weights. They argue it would improve safety, fairness and competition. It is a proposal, not an adopted rule.

Writing for the University of Virginia’s Darden School of Business, Lenox and Cowan frame the proposal as an alternative to both fragmented oversight and heavier government intervention. Their timing argument centers on AI companies going public: they want the rules settled before a much larger population of shareholders has a financial stake in resisting them.

An annual deadline, not immediate disclosure

The proposed obligation would cover two groups: firms that train AI on material they have not licensed, and firms that build on models trained with unlicensed data. It therefore reaches beyond the company that originally assembled the training material. The requirement would follow models built on that foundation, rather than apply only to the first developer.

Under the authors’ schedule, companies would retain a lead of six to 18 months, averaging about a year, before releasing model weights publicly. The proposal does not call for companies to surrender that lead immediately. Its fairness principle is that businesses using a shared inheritance of books, articles, research and web material should eventually return something to the public.

For competition, they draw an analogy to patents: an inventor receives a period of exclusivity in exchange for publishing how an invention works. Here, the intended result is that more businesses could build on last year’s leading AI models. The authors expect that access to support new firms, lower prices and more employers competing for workers.

The best argument is that too much disclosure removes the incentive to invest at the frontier.

Mike Lenox and Alex Cowan, in their Darden policy essay

The safety case rests on a disclosure clock

Lenox and Cowan acknowledge that investment objection but reject it. They argue that a year’s lead would remain commercially valuable. They also challenge the premise that regulation should preserve the fastest possible development pace, pointing to public polling they cite as showing substantial support for slowing AI development.

Their safety argument borrows from coordinated vulnerability disclosure, in which a software vendor gets a fixed window to address a security flaw before details become public. They contend that AI models can be copied, leaked and rebuilt, making permanent secrecy a weak foundation for safety. Defenders, in their view, should have at least the tools available to attackers.

A known annual release date would give the developer, its customers and the national security community time to identify dangerous capabilities and prepare defenses, the authors argue. They suggest that the delay could shrink as the field matures. That is their proposed safety mechanism, not evidence that annual releases have already made AI safer.

Set the rules before ownership spreads

The essay uses Anthropic’s October 26 IPO filing as its immediate backdrop. The authors are not calling for AI IPOs to stop. Instead, they argue that public listings would spread financial exposure through index funds, pensions and retirement accounts, making later rules that affect AI companies’ value harder to enact.

They also propose a test of their own approach: examine market concentration and prices after two or three annual release cycles, or 24 to 36 months. If those measures have not moved, they say the public should seek stronger intervention. The proposal thus offers a timetable for judging its competition claims, rather than asking for indefinite faith in public access alone.