On its current trajectory, the American AI industry is set to commit the largest, most disruptive theft of value from creators and the commons in modern history. Large language models (LLMs), the very basis of the technology, are built and trained on the pirated gatherings of copyrighted content, only to make authors and editors redundant. Software engineers train AI systems to autonomously write code, only to see their creations replace them. Mayors allow Waymo to ingest municipal data and scan city streets, only to put legions of drivers out of work. And all the while, CEOs of AI firms are out in the media spooking the world with dire predictions, only to see their valuations rocket into orbit.
As AI related job losses start to mount, and the content of our culture gets masticated into LLM modules, federal and state lawmakers remain paralyzed—overwhelmed by the scale of the challenge, flummoxed by the erratic Trump administration, and cowed by the threat of a SuperPAC backlash. Yet something must be done. Unregulated and unchecked, AI presents an almost existential threat to what we call civilization.
For candidates in this mid-term election cycle who dare to step up and advocate just, systemic approaches to reining in the worst of AI’s deleterious effects, we offer three proposals:
The first is a statutory collective licensing law for all copyrighted content used by the industry. LLM’s are built upon the assembled knowledge of humanity—news articles, essays, research papers, books, blogs, Reddit posts, photography, art, movies, music, user queries and countless other categories of content that are simply scraped off the internet. Much of it is protected under copyright law. But with few exceptions such as Anthropic’s recent settlement with authors for copying books from pirate sites or OpenAI’s deal with Reddit for user content, AI firms pay nothing for it. Their defense of this practice rests on a legal definition of “fair use”—that in training AI models, the content is so “transformed” as to be not subject to copyright law. In the few cases that have been contested, the courts have ruled in their favor.
Copyright law should be updated to require that all copyrighted content used to train AI models be licensed and compensated—through royalties, equity or both. Because the transaction costs for entering into agreements with every content creator would be too high, a “collective” licensing arrangement makes most sense. There is already precedence for this in the music industry. In 2018, Congress passed the Music Modernization Act (MMA), which effectively created a mechanism for streaming services to obtain one single license covering all songs. The MMA established a nonprofit that negotiates with the streamers every five years to set a significant revenue-based royalty, which is then distributed to musicians and producers based on pro rata listenership. A similar framework can be established for content used to train AI. Under this scenario, AI firms would pay a share of their revenues annually to an entity that would, in turn, distribute those revenues to content creators. This would, of course, be far more complex than music, given the range of content AI models are trained on. The policy would also need to establish clear and acceptable metrics for calculating each content creator’s share of royalties—such as token-based measures or assessments of relative importance. No doubt, this will require nuance and is sure to be contested, but without such a mechanism, we will see a massive shift in wealth from human content creators to AI firms. It has already begun. Plus, we will destroy any incentive for humans to produce original work; that is, we will snuff out the wellsprings of our culture.
The second is an equity tax. Here, AI firms would be required to give a substantial percentage of ownership, through the issuance of new shares, to a government sponsored entity that would then directly distribute the dividends of those shares to taxpayers. This is akin to how the Alaska Permanent Fund works, which distributes a percentage of the state’s oil and gas revenues directly to taxpayers. Those dividends could be distributed equally across taxpayers or via some income-based formula. Taxpayers would then participate proportionately in the gains produced by AI. This is justified: in addition to copyrighted material, AI is trained on unprotected content such as user queries, short expressions from social media posts and ideas, processes and systems. This content belongs to the people. Moreover, much of the foundational research for AI led by scientists including Marvin Minsky and John McCarthy was supported by taxpayer funded institutions such as the Department of Defense and the National Science Foundation. Perhaps most importantly, AI will rapidly lead to rising unemployment across many sectors of the economy. Without societal participation in the value creation of AI, we risk ever greater economic inequality and disruption with consequent political instability.
Third, AI firms should be subject to stakeholder governance and be legally chartered to place human interests above maximizing profits. When founders and investors have complete power over big tech firms, history shows that the costs to society can be devastating. At Meta, a firm that impacts billions of people around the world, one person has de facto executive control and the right to appoint the majority of its board. The resulting brutal externalities of this arrangement are a matter of public record. Current predictions by AI executives themselves project that the social and economic carnage of AI will dwarf that of social media. Society must play a role in governing this force. Corporate law should mandate that AI firms be governed by multi-stakeholder boards that, in addition to investor interests, include content creators, nonprofit leaders, academics and members of civil society. There is precedence for this too. Anthropic is governed by a Long-Term Benefit Trust, whose trustees include such stakeholders. It may be why Anthropic tends to take the higher ground on many business decisions. However, Anthropic’s charter can be altered by a supermajority of its shareholders, which include the who’s who of Silicon Valley and several of the Magnificent Seven. Stakeholder governance shouldn’t be optional—it needs to be mandatory. Given what is at stake, this is the least that society should demand of the industry.
AI has been built on the cumulative output of human expression and culture, and AI has the potential to spread disruption and social costs on a catastrophic scale. If changes are not made, the gains of AI will accrue to the very few founders and the wealthy investors behind them, while the economic carnage and social pathologies will be visited upon the rest of us.
Candidates, it’s up to you: the hour is late and voters are terrified. They need policies, not platitudes. These proposals can make a difference.
by:
Hans Taparia, Clinical Professor, NYU Stern School of Business
Bruce Buchanan, C. W. Nichols Professor of Business Ethics, NYU Stern School of Business