By the Time Governments Are Regulating AI, They’re Regulating the Past

By Mark Jamison of AEI.

"AI is changing fast. And spreading fast. Both are problems for people seeking to regulate it.

Regulation works best when regulators understand what they are regulating. That is not the case for AI. Costs are collapsing, having dropped 1000-fold for large language models since 2023. At the same time, capabilities are expanding, business models are changing, and market leaders are turning over rapidly.

These dynamics create both opportunities and problems. Now almost anyone can use AI to manage their household or launch a business. But, as Bill Gates recently noted, they can also create deep fakes, launch phishing attacks, or break into internet sites, as happened to Hugging Face.

This also makes regulation hard: Rules written for the AI that regulators see today will no longer exist by the time the rules take effect.

Nevertheless, many people want regulations that would control AI. The EU has embraced what it calls comprehensive AI regulation, in which regulators judge the relative riskiness of AI applications and systems and then apply controls ranging from outright prohibitions to light-touch oversight. Some people are calling for mandated surveillance of AI users, restrictions on model capabilities, product standards, and computer code review. Gates recommends an international organization layered on top of national all-of-government regulators to oversee all AI risks people imagine. All of these approaches assume overseers who would control innovation.

While it is true that whenever a technology’s costs fall and its abilities grow, people use it more. Sometimes for evil. When Henry Ford put cars within reach of every family, some families created new businesses, but others became bank robbers. When internet service providers spread access across the country, e-commerce exploded, but so did criminal activity on the dark web. In these instances, successful regulatory responses were not to limit cars or the internet, but to use the technologies for regulatory purposes.

The examples of automobiles and the internet illustrate a path forward for AI policy: Let the technology evolve for the good it can do. At the same time, officials and entrepreneurs can protect citizens by developing their own innovations based on a deep understanding of the technologies and their markets.

Recent research published in the Journal of Economic Perspectives provides insights into AI and its markets. The researchers examined the AI most people use, LLMs. LLMs are growing in complexity, now operating in three layers: The Model Layer, where creators such as OpenAI and Meta design and train LLMs; the Inference Layer, where AI providers like OpenAI and Together AI host and run models to respond to user requests; and the Application Layer, where a large ecosystem of startups and established firms embed LLM capabilities into user-facing applications for accounting, legal, retail, and other services.

Activity in each layer has exploded in multiple directions. The Model Layer grew from 1 model in early 2023 to 668 by the end of 2025. There are curious dynamics in this layer. Open-weight models—which allow users to customize systems for specific tasks—charge users 90% less than do closed weight models. Open-weight providers effectively give away their models after spending billions in development and training. Despite what looks like bad economics, there are over twice as many open-weight models as closed-weight models, 449 versus 219.

Customers in the Model Layer also make choices that appear counter intuitive. Even though open-weight providers charge 90% less than do their closed weight counterparts, customers use closed-weight models more than twice as often.

The complexity doesn’t stop there. At the Inference layer, the number of providers grew from 30 to 90 in 2025. These providers largely host open-weight models and their non-price capabilities vary. Closed-weight model creators are more likely to have vertical relationships at this layer.

The growth and interplay of these two layers illustrate why controls can be counterproductive. They limit innovators’ abilities to experiment, meaning that there would be fewer models in both levels. Fewer models at this stage of development means fewer opportunities for customers to express their preferences. And it is unclear whether the two layers will remain separate.

Model diversity is growing in several ways. Measured by the Artificial Analysis Intelligence Index, 80% of the models fell between 0.1 and 0.29 on the scale at the beginning of 2025. By the end of the year, they fell between 0.22 and 0.61, an increase in spread of over 100%.

Market leadership changes often. In the Application Layer, the market leader for science-oriented models changed eight times in 2025, while the market leader for legal services changed five times.

What does this mean for regulators? The innovators, investors, and customers driving AI are creating tremendous value, estimated to be approaching $1 trillion. Regulatory controls handicap legitimate AI providers and create market opportunities for those less inclined to follow the rules.

The lesson isn’t that government has no role in AI’s evolution. It is that seeking to control AI is counterproductive. AI policy should let the government be a leading AI user without limiting legitimate users of AI. This is more like what good governance has always done: punish harmful conduct and protect citizens by adapting its own capabilities as the world changes."