What’s Really Behind the Science Slowdown the White House Wants to Fix?

If we’re going to have a serious discussion about science and technology policy, it’s important to get our facts straight. A recent interview of White House Office of Science and Technology Policy director Michael Kratsios by George Mason University economics professor Tyler Cowen in the New York Times contained a number of claims and declarations about the state and direction of the American science and technology ecosystem. Claim one is that the current model of science and technology policy is based on the idea that the government funds basic research at universities so that the companies can acquire the knowledge and then commercialize it. The second claim deals with science slowing down and the need for the private sector to take a bigger role in the science and technology landscape. Third, there is a revolution in artificial intelligence taking place, and participants in the science ecosystem need to change if they want to stay relevant in this new ecosystem. Fourth, the amount of money that the government spends on science and technology demands greater public accountability and oversight.

There are a number of fundamental problems with these claims, but for the sake of this piece I’ll treat Kratsios' claim that science is slowing down as “true.” I don’t think this is the case. The evidence for it is relatively weak, and such arguments are easily weaponized and coopted. But it makes for an interesting starting point for a conversation, especially in the context of the other claims.

Issue 1: The model should reflect reality

The first claim revolves around the balance of federal research funding. Kratsios is right that the Trump administration has proposed the largest R&D budget in history, but the majority of that spending is allocated to the Department of Defense’s Research, Development, and Test & Evaluation programs. While the surge in DOD expenditure is new, the idea that the government’s role is in funding basic research has not been true for quite some time. Looking at the post-war history, when government expenditures on research and development peaked in the late 1960s during the Apollo Program and atomic eras, and the overwhelming majority of that expenditure was spent on “development” activities — the kind of stuff that focuses on building cars, rockets, and household products as opposed to the pursuit of pure knowledge. Even if we assume that the government is the primary funder of basic research, the overall share of that research has also fallen over time, with businesses and philanthropies taking a larger share in the mid- to late-2010s.

The other element of the claim is that the purpose of government research is to produce knowledge which can later be commercialized. Again, this isn’t the case, and also wasn’t the case in the immediate postwar period, when Vannevar Bush made the argument for re-locating the functions of the military-focused Office of Scientific Research and Development to his proposed National Science Foundation. Institutions like the Atomic Energy Commission, the US Coast and Geodetic Survey, and the National Institutes of Health, and many other research organizations had large “research for the public interest” missions in the pre- and post-war periods. Indeed, the British government’s 1918 landmark Report of the Machinery of Government (upon which the near-mythical Haldane Principle is based) states that “Even in countries where, as in the United States, there is a Federal Government with but little power to interfere in local administration, the value of systematic investigation and accumulation of general knowledge by Departments of the central Government, though unaccompanied by the power to apply it administratively, is regarded as of great importance…”. This research dealt with things like making sure that we have accurate maps, helping farmers manage drought, and making sure that municipal fire hydrants do their job.

When we look at the research that’s done in the US government’s own labs, the overwhelming majority actually goes to experimental development activities (74 percent) and applied research (16 percent), as opposed to basic research (less than 10 percent). Most of these activities are devoted to increasing the lethality of the American warfighter and not commercial endeavors. To be sure, the military has a proud history of producing research that later serves the public benefit, but defense R&D’s impact on long-term productivity growth tends to be minimal.

Even if Kratsios deals with a lot of advocates who think the government should “just fund basic discovery,” data from the past 80 years tells a different story than the one that he and Cowen describe. Their story might be the simplified one that occupies most science policy discourse, but it’s always been a poor reflection of the actual American model for research and development.

What they get right about the model is that businesses tend to fund research that aligns with their core business interests as opposed to the search for new knowledge. Where industry has stepped into early stage research, it’s primarily in areas that are aligned with their business objectives (like AI or quantum computing), devoted to solving problems that are necessary to move onto the next stage of development. Less effort is spent on doing things like studying the venom of a gila monster without the prior knowledge that doing so could solve the global obesity crisis or peering into deep space in search of new physics. Those functions remain relatively unique to federal support for basic research.

Issue 2: Who’s to blame for fewer breakthroughs?

The second claim deals with the idea that science is slowing down and that the current model requires a shakeup, with industry and philanthropy taking a larger role in discovery. The problem with this claim is that business is already the majority funder of research and development, with a dramatically increased role in overall funding for research and development since the 1980s. If science is slowing down, that slowdown is occurring in the context of a system where three out of every four dollars comes from private sector investment. Most of that money is through business funding of R&D, which has risen almost every year since the early 1980s. Businesses now even fund a third of the nation’s basic research, up from 19 percent in 2000.

Kratsios is quick to point out that money at the National Institutes of Health is primarily going to late-career researchers and cutting younger folks with bold new ideas out of the market. Looking at the success rates of the National Institutes of Health by institute, even prior to the Trump Administration researchers only had about a 1 in 5 chance of having their research proposal funded, falling from about 1 in 3 in the late 1990s. This is primarily a function of dramatically increased demand that could not keep pace with a massive lagging inflow of talent after NIH’s “doubling period”, and offers a pretty clear Occam’s razor explanation for why peer review panels were forced to behave more conservatively. But to Cowen’s point, we are still seeing major developments come in biomedical research, from blockbuster obesity drugs to potential miracle mRNA cures like a universal cancer vaccine, so it’s much harder to argue that the biomedical community, at least, is performing badly.

Again, when we look at the total federal funding for research, US government funding of universities is only the third largest expenditure for R&D, In fiscal year 2024, businesses received about $70 billion, federal labs and federally funded research and development centers (what we think of as national labs) about $68 billion, while universities receive about $45 billion. If we think about American research and development like an investment portfolio, the government’s spending on higher education looks more like a modest bet on the off chance that something really big could happen.

So if we take Kratsios at his word that we’re getting fewer breakthroughs, then does it make sense to assign sole responsibility to spending that’s responsible for only five percent of R&D performance? What are we to say of the other 95 percent?

The far larger pot of funding comes from the private sector, much of it subsidized by the R&D tax credit. The tax credit suffers from similar problems that Kratsios assigns to NIH funding, with more than 95 percent of credit dollars going to the largest 5 percent of incumbent firms (in part due to the administrative difficulty securing the credit presents for smaller firms). And it’s a well-known economic maxim that large companies are disincentivized from innovating once they’ve cornered their market and actively work to acquire or kill competing products. This doesn’t mean that NIH shouldn’t try to improve its funding model, but it also suggests that federal tax credits may have significant room for optimization.

The R&D tax credit also incentivizes companies to maximize their return from the US government by attaching it to as many of their internal expenditures as possible. Meta was recently accused of using part of the $6.6 billion in credits it claimed from 2023-2025 to support data center investments. Taxpayers unwittingly also gave Meta billions of dollars in tax credits for the noble purpose of trying to make legs work in the metaverse.

More on that point in a moment.

Issue 3: The (old) age of artificial intelligence

If you’re reading Tech Policy Press, you’re likely well aware of the explosive growth of the AI industry in recent years, especially in the form of frontier models produced by companies like OpenAI and Anthropic. To be sure, advances in the capabilities of frontier models have been impressive over the past few years and are now disrupting STEM occupations in ways that are not always welcome. But the notion that we are on the cusp of an AI-driven enlightenment ignores (once again) AI’s existing integration into science — particularly large scale scientific activities.

To understand AI’s scientific integration, we need to understand the scale of modern scientific activities. If we look at instrumentation like the Square Kilometer Array in South Africa or the Large Hadron Collider at CERN, the amount of data produced by their observations and experiments rivals that of companies like Meta and Google, necessitating advanced pattern recognition and identification techniques enabled by machine learning and artificial intelligence. CERN’s first use of AI was in 1987 looking for faults in its experimental equipment, and you can find mentions of the use of neural networks in astronomy and remote sensing papers even earlier. Areas where AI is used less tend to be those which require a high level of explainability and reproducibility beyond what black boxes can offer.

For those of us following AI in science, the Genesis Mission announcement by the Department of Energy was actually pretty mundane. In areas like fusion, scientists have already successfully used AI to help mitigate destructive plasma instabilities in tokamaks. And while the level of focus for Genesis is impressive (if not its actual budget, which is zero dollars since it has no corresponding appropriations and is entirely derived from other DOE programs), it’s coming at the cost of the Department’s investments in critical research and development infrastructure which is deteriorating at an alarming rate. Even if AI is capable of coming up with the most beautiful theories of the universe, the scientific method demands that those theories be torture tested on platforms that are increasingly transcontinental in nature and scope. Building a 100 kilometer supercollider requires mundane things like land acquisition, global supply chains, and the consent of local populations. Money at agencies like the NSF is getting reprioritized away from investments in quality data sources, like scientific collections, which are essential data sources that are needed for AI to produce quality results. If funding redirection compromises talent development or the core infrastructure required for R&D to function, then one of AI’s most exciting potential applications could land with more of a whimper than a bang.

Issue 4: Public accountability

The final issue is with the argument that because we’re spending all this money on research and development, then it’s in the nation’s interest that elected officials have some oversight of research and development expenditures. Kratsios argues “there most certainly should be political oversight” of grants. But this is already the case (and has been for years). Congress decides what gets funded and has its own oversight capabilities, including through the Government Accountability Office. Senior program officers at the National Science Foundation and National Institutes of Health already review the work of peer review panels to make sure that the government’s funding is well-spent. The review panels are important because they provide a layer of accountability and verifiability that the research the government is funding represents questions that are worth asking from folks who are qualified to ask some questions. Under this administration we have already seen political officials intervene in public health research in ways that violate basic ethics protocols and put populations at risk and disrupt Congressionally-mandated programs.

This brings the conversation back to the R&D tax credit. If we need to make sure that we’re “good stewards of taxpayer dollars”, as Kratsios put it, it’s worth noting that the money the government spends on the tax credit does not require any form of peer or political review—just the occasional audit. This is how the American taxpayers end up spending billions unwittingly subsidizing AI infrastructure build-out and putting legs on Mark Zuckerberg’s defunct avatars. The US Congress’ Joint Committee on Taxation estimates the credit will cost American taxpayers about $190 billion dollars from 2025 to 2029, more than twice what the National Science Foundation will spend in the same period assuming level budgets.

If private sector companies are receiving hundreds of billions of dollars in tax expenditures, then it strikes me that those subsidies deserve at least as much scrutiny from the metascience community as that which is directed at much smaller government grants. This is not to say that the tax credit is not valuable — macroeconomic research suggests that it is — but there is ample opportunity to ensure that the credit serves its intended purpose supporting breakthrough developments. At the very least, it should put the pearl-clutching about whether or not NSF and NIH spend their money wisely into perspective.

One last thought on that point. A key argument coming from the growing conservative science reform movement is that the system requires change due to declining public support for science. Again, this needs to be viewed in context of the broader science and technology ecosystem. While public trust in “science” remains pretty durable, the bottom is collapsing around tech companies and the AI buildout. That doesn’t necessarily mean that the systems supporting research can’t or shouldn’t be improved–they absolutely should–but a recalibration in issue prioritization might be in order if tech’s social contract is to be maintained, especially if taxpayer subsidies for R&D are flowing en masse to stuff that people hate. If we know anything about the current political mood in the US it’s that people don’t like data centers or corporate welfare, and members of Congress are changing their behavior. Shifting the nation’s scientific resources into AI, especially given apocalyptic warnings and recent controversies around mathematics, seems more likely to worsen trust dynamics.

Back to reality

Kratsios is right about one thing — the discourse around Washington tends to center itself on the value of basic research. But those debates are often out of touch with reality and ignore large sections of the American R&D ecosystem, which is overwhelmingly weighted toward development activities funded by a relatively select portion of the American economy. Even if we look exclusively at government spending, the money that goes to universities represents a relatively small portion of the overall picture — hardly enough to be blamed for what’s going on in the broader science and technology ecosystem.

If we are to take Kratsios at his word, and that breakthroughs are becoming more scarce, then any debate about the efficacy of R&D spending needs to look at the ecosystem in total. For the past 40 years, that ecosystem has been dominated by the private sector. It might just be that after the government stepped away from major development activities, like the Apollo and atomic energy programs, passing that responsibility to the private sector played some role in the slowdown. A deep dive into those dynamics might reveal something that this administration is less willing to accept.