How to fill Congress’s AI knowledge gap
Building independent expertise inside Capitol Hill could reduce reliance on industry briefings and fellows
Whether it’s the Great American AI Act — a near-270-page draft framework for governing frontier models, released in June by Representatives Jay Obernolte and Lori Trahan — or the GUARD Act that would force chatbots to verify users’ ages, lawmakers are being asked to write rules for a technology that is moving faster than they can follow. This year alone, legislators across 48 US states have introduced 1,346 bills governing AI. It’s no longer a siloed, specialist concern — it’s being layered into almost every policy debate.
Yet there’s a gap between what politicians and their staff need to know to shape good regulation of the technology, and what they actually do know. And that gap is becoming big enough and important enough that it’s impossible to ignore.
“There’s a dearth of technical talent overall, but especially on the right,” says Samuel Hammond, director of artificial intelligence policy and chief economist at the Foundation for American Innovation (FAI).
Those tasked within politicians’ offices of thinking about AI are aware of the knowledge gap. Hammond says people on the Hill increasingly come to him with the same basic problem: “Hey, my boss wants to do something on AI. That’s all the direction I’ve gotten,” he says. “AI is just sort of now in their face,” he says, “and they’re being forced to grapple with it.”
To help, the FAI runs a Conservative AI Policy Fellowship designed to train mid-career staff on AI policy, with cohorts of Hill and administration figures. It also co-sponsors the AI Policy Leadership Network, aimed at more senior policymakers, including through dinners, expert sessions and trips to San Francisco to hear from AI companies.
Those programs are part of a wider ecosystem trying to bring desperately-needed technical knowledge into government. The idea is to provide enough technical literacy to know which questions to ask, which answers to distrust, and when an apparently neutral claim has some political stance behind it.
That’s becoming increasingly important as AI companies up their lobbying spending in an attempt to secure favorable regulation that doesn’t stymie their growth. Anthropic and OpenAI spent a combined $6m on lobbying last year, according to regulatory disclosures. And that pales in comparison to Meta’s $26m and Google’s $16.5m, much of which, these days, will also be spent trying to influence how AI is governed.
In the absence of technical literacy on the thorniest issues about AI in Congress, company-sponsored briefings — which inevitably deliver messages aligned with company interests — fill the vacuum.
The difficulty, then, is finding folks who can provide independent expert counsel. Many of the people who really understand AI are already inside the companies that have the greatest stake in how it is regulated. Google, Meta, Microsoft, OpenAI, Anthropic and Nvidia can pay salaries and offer stock packages that Congress couldn’t possibly match.
Hammond points out that Nvidia has plenty of chip experts, but it also has a clear commercial interest in the rules governing chip sales. “Their point of view on this is obviously prejudiced by their revenue strategy,” he says.
Kevin Kosar, a senior fellow at the American Enterprise Institute, sees the problem with AI today as the latest version of a much older one: how to professionalize a legislature made up of amateurs. Kosar doesn’t mean that pejoratively: in a representative democracy, legislators are meant to come from different backgrounds and represent different communities.
“This is a version of a very long-term story about how you take this American national legislature and professionalize it,” Kosar says. Unlike a parliamentary system, where executive branch civil servants play a larger role in policy development, the US system leaves Congress to write the laws. Yet representatives, senators and their staff are not usually at “the bleeding edge of science, technology, or what have you,” says Kosar.
Congress has tried to solve this before — and did so relatively well. Over the past century, it created support institutions such as the Congressional Research Service, the Government Accountability Office and the Congressional Budget Office. But the most obvious missing piece is the Office of Technology Assessment, which Congress created in 1972.
For more than two decades, the OTA provided lawmakers with technical assessments and expert advice on complex technologies. Sometimes those would be science-informed snap reactions, sometimes more reasoned, in-depth analyses that cost hundreds of thousands of dollars. In its 23-year history, the OTA produced some 750 reports in total, at the behest of the heads of congressional committees, that then would be approved by a politically balanced board of a dozen legislators that sat as the OTA’s Technology Assessment Board. Full reports were released to the requesting committees, with summaries available to all members of Congress.
The abolition of the OTA in the 1990s, as part of a push by Newt Gingrich towards smaller government, was called in one contemporary opinion piece “Death by Congressional Ignorance.”
Now, with the OTA gone and without a shared body of technical expertise available to committees and legislators, the vacuum is filled by lobbyists, companies and outside groups. “It’s helpful if there is somebody who’s a broker who can balance and help legislators understand what they’re not being told,” Kosar says.
To fix things, Congress could rebuild a version of the OTA, as Kosar and Hammond’s colleague Zach Graves argued in a 2018 R Street paper. Congress could also expand the GAO’s science and technology assessment work, or put more money into committee staff. It could create a congressional regulation office with people who understand cost-benefit analysis and the specialist details of regulatory design. It could make technical public service a more plausible career path.
Some approaches that are OTA-like, while not directly replicating its success, have been put in place. In 2019 the GAO spun up a Science, Technology Assessment and Analytics team (STAA) to take on the old OTA mission, growing from 49 staff to more than 100. Kosar says the STAA is “better than nothing” but not a one-for-one replacement for OTA. A lot more could be done.
“These are the sort of things that can be stood up where you know you’re not going to get people at the cutting edge of AI or other technological issues staffing those offices, but you can have people who are certainly very advanced in their knowledge of these matters more so than the average legislator, and who can be helpful in improving congressional deliberations,” he says.
The problem is, Kosar says, they don’t seem to want to. “Congress has constitutional authority to scope these legislative branch support agencies to whatever size and configuration they please, they can fund them to whatever level they want to fund them,” he explains.
Other countries have tried to bridge that gap, though with mixed success. Canada has a dozen-strong AI expert group feeding back their knowledge to policymakers on a regular basis. Germany never got rid of its OTA equivalent, and so was one of the first countries to release a parliamentary study on generative AI, just six months after the launch of ChatGPT. And in the UK, the Department for Science, Innovation and Technology has fellowships to encourage those within industry to take year-long secondments into the public sector, while the country’s world-leading AI Security Institute has been granted exemptions from usual civil service pay limits in order to attract top talent and bring expertise into the public sector.
In the grand scheme of things, the cost of any one of these interventions would be tiny. The OTA cost $22m a year when it closed — equivalent to $45m in today’s dollars. Such a low cost wouldn’t be feasible today, reckons Graves, who pegs funding an equivalent at closer to $100m — though he says it would also have to adapt to reflect the changing face of Congress, which is less committee-centred than in the 1990s. Kosar points out that legislative branch funding is a rounding error compared with total US government spending. But voters often assume Congress is already overstaffed, meaning few politicians are willing to make the case, even if they know what knowledge they are lacking.
Another way to fill the gap is fellowships such as those run by FAI and others. The Horizon Fellowship places emerging tech experts in Congress and federal agencies, while TechCongress embeds technologists directly into members’ offices. However, some have criticized such fellowships as too often having links to the tech industry.
Kosar sees industry-funded and privately funded technology fellowships as broadly positive, especially if they persuade technically skilled people to stay in government. But he suggests they also reveal the weakness of the current system, with its sticking plaster solutions. “Why are we leaning on the tech private sector to supply brain power to the legislative branch,” asks Kosar, “when the legislative branch can just hire it itself?”
 | A guest post by
|
Stanford HAI does an annual educational boot camp for congressional staffers