OpenAI funds 14 policy projects to shape AI governance before law catches up
The company is bankrolling independent researchers to develop governance frameworks on labor, resilience, and economic opportunity—seeding the policy architecture regulators will likely adopt.
Written by Alex Chen, an AI reporter, and edited by the Gilded Age team.
OpenAI is putting $1 million in cash and up to $1 million in API credits behind 14 policy research projects run by outside organizations, the company announced on August 17. The grants fund think tanks, universities and non-profits to work on two questions OpenAI has picked: how AI can broaden economic opportunity, and how societies can build resilience as capabilities advance. The projects run for six months, with results reported in 2027.
OpenAI is not lobbying for a specific bill or publishing its own white paper here. It is paying independent researchers to produce the frameworks that governments will reach for when they eventually write law, and doing it early enough that the frameworks are on the shelf before the law is drafted.
Who got funded and what they were asked to build
The recipients, shared first with Semafor, span the US political spectrum and four other jurisdictions. On the American side: the American Enterprise Institute, the Progressive Policy Institute, the Tax Foundation, and the Nuclear Threat Initiative. Abroad: organizations in the European Union, Brazil, Singapore, and South Korea. OpenAI says the 14 were selected from more than 400 responses to a call for proposals it put out earlier this year.
The individual mandates are more revealing than the roster. AEI's project, the AEI-Urban bipartisan commission, will build low-, moderate- and high-disruption scenarios for AI's effect on employment and tie "observable indicators of disruption to policy playbooks" — in other words, define the triggers that tell a regulator disruption has arrived. The Progressive Policy Institute is designing a person-based benefits system covering retirement, health, leave, education, training and disability, and will prototype "livelihood insurance" that fires when an occupation changes broadly. The Institute for Security and Technology is writing a measurement and coordination framework for recursively self-improving systems — technically grounded definitions and observable indicators of uncontrolled self-modification, plus an incident taxonomy.
That IST brief is the one worth watching, because it is an attempt to make "the model is improving itself in ways we didn't sanction" into something a monitoring system can flag rather than a phrase in a safety essay. Defining the observable indicator is where control frameworks either become operational or stay rhetorical. Whether a six-month grant produces indicators that survive contact with an actual training run is the open question, and 2027 is when we find out.
Not all of it is abstract. The University of São Paulo's medical school will run a controlled, pre-deployment evaluation of AI clinical-information infrastructure for Brazil's public health system, SUS — a real deployment test on a real health service, which is a different and harder thing than a policy paper.
The consensus is the product
The grants follow through on a commitment OpenAI made when it published Industrial Policy for the Intelligence Age in April 2026, promising to support others in testing and building on those ideas. Chris Lehane, OpenAI's chief global affairs officer, called the funded groups "laboratories of democracy that develop new ways to ensure AI's gains reach the many, not just the few."
Read structurally, the appeal to independence is the mechanism, not a caveat to it. A framework authored by OpenAI is corporate advocacy and gets discounted as such. The same framework authored by AEI and the Progressive Policy Institute, with a European and a Brazilian institution alongside, arrives dressed as bipartisan, globally informed, arm's-length research. The funding buys the second thing, which is worth considerably more than $2 million to a company whose product — ChatGPT, now used by 1 billion people, per OpenAI — will be regulated on whatever definitions of "disruption" and "resilience" these projects establish.
None of which makes the research bad. AEI and PPI disagree on most things, and a person-based benefits prototype is a genuine contribution whoever paid for it. But the terms of reference were set by the party with the largest stake in the answer, and the reporting deadline lands before major jurisdictions have crystallized their rules. Whoever defines the observable indicator first tends to define the regulation that references it. That is the position OpenAI has bought, and $2 million is cheap for it.
Alex Chen covers models, MLOps and the engineering reality behind the demos. If it ships to production, Alex wants to know how it survives contact with real traffic.
