AI is only half the solution to economic growth, the other half depends on how governments support science

Artificial intelligence (AI) is often sold on the basis that it will lead to more breakthroughs. Dario Amodei, Anthropic’s CEO, believes the rate of discovery in biology could rise “by 10x or more if there were a lot more talented, creative researchers,” which, he implies, AI agents would supply.

The presumption is that the fundamental limit on knowledge creation is intelligence. If only this were so. Governments and businesses are investing billions in AI on the promise that it will dramatically accelerate innovation and economic growth. But what if the biggest constraint on AI-driven discovery isn’t intelligence, but the way we organize and share knowledge?

Researchers have already harvested science’s low-hanging fruit, making each successive insight harder, more expensive and slower to find.

It's estimated that it now takes 41 times as many researchers to sustain the growth the US achieved in the 1930s. Automating the production of ideas through AI is posited as a way to offset this loss. Stanford academic Charles Jones summarized this as a world in which “machines invent all possible ideas, leading to a maximum productivity”.

But while AI can generate ideas faster, it can’t overcome a system that prevents ideas from being shared, tested and built upon. That's why governments seeking economic growth and technological sovereignty from their AI investments should treat open knowledge as critical infrastructure.

If not, they risk investing billions to make AI smarter while underinvesting in the knowledge ecosystem the technology needs to actually deliver economic growth.

The innovation system’s efficiency is declining as a result of how we organize it, as well as misaligned incentives and the enclosure of knowledge.

A 45-year study by US academics tracks patenting across all US firms, not just the listed manufacturers previous research relied upon. It shows that growth – not ideas – is getting harder to find. Patents per research dollar rose steadily over those four decades.

The study suggests that the flow of shared knowledge into what firms produce has weakened. The authors conjecture that the largest firms, which hold most breakthrough patents, can exclude competitors from their ideas, or stop them growing, sometimes by buying them. Indeed, another academic study of the pharmaceuticals industry documents that 5 to 7% of acquisitions are aimed at shutting down a target’s drug projects rather than building on them.

The innovation tap is turned down not by a lack of knowledge but by its concentration in a few firms. This reduction is achieved through material transfer agreements and intellectual property licences that delay projects and restrict what smaller firms may do. Negative findings go unpublished, so others duplicate the same useless research.

All of this creates friction in the movement of ideas.

Amodei calls obstacles of this kind “constraints from humans”. He considers limits placed by law, regulation and public resistance to be the cause, and expects AI to circumvent them in time.

But governments risk making a fundamental policy mistake by investing heavily in AI infrastructure while neglecting the open knowledge infrastructure that makes AI-powered innovation possible.

Four academic journals rejected Francisco Mojica’s paper identifying CRISPR as a bacterial immune system. Nobel Prize-winner Katalin Karikó was demoted by her university in 1995 because she could not win grants for the mRNA work that later saved millions of lives and won her the prize. These are examples of how science is poorly organized, not of a lack of intelligence.

An AI system would still have demoted Karikó because she did not meet the criteria for promotion. Mojica’s paper may still have been rejected regardless of the expertise or insight of the editor reviewing it. That is, no additional supply of brilliance would have changed the outcomes.

The costliest constraint from humans, therefore, is not legal resistance but how we choose to organize research and development.

Obtaining economic general growth, means altering the way science is organized. The increasing returns that drive aggregate growth flow from the non-rivalry of ideas that everyone can use at once. If governments want growth and sovereignty as a return on their AI investments – data centres, electric grids, training – they must reduce the limits on general growth that leave domestic firms dependent on mega-firms.

Governments should require open sharing of standardized data in useable and interoperable forms. They should also require institutions not to use patents to restrict research in areas where knowledge and tools are still being developed for broad use across the field, and not to withhold the publication of negative results.

Beyond that, governments should financially support open data platforms that not only establish data formats and set quality standards, but ensure reproducibility by encouraging actors to share results through open competitions. Conscience, a non-profit that I helped launch, is already doing this through its BEACON initiative.

The benefits that countries can obtain from AI depend as heavily on the knowledge infrastructure put in place by government and institutional policy as on AI itself. Putting billions into one system while ignoring the other is a poor investment – one that will limit economic growth.