Alaska Is a Model for Taxing AI as an Extracted Natural Resource

As artificial intelligence continues to revolutionize industries, it’s easy to lose sight of the fact that it’s built on a resource we collectively own—human knowledge. Like oil or mineral deposits, data is mined from vast natural reserves every day.

AI models get their value by ingesting articles, art, books, and other human-created works. Viewed through this lens, we should consider how to ensure AI’s harvest of knowledge is preserved for the public good, just as states have safeguarded the profits from oil extraction for future generations through investment in sovereign wealth funds.

This is where tax policy provides an answer. By taxing businesses that generate revenue from AI models, and then investing that revenue in a US sovereign wealth fund, we can help ensure the windfall is more fairly distributed over time. Sovereign wealth funds are state-owned investment entities that serve exactly this purpose—preserving wealth from today’s profits for future generations by generating long-term returns or investments.

Companies building AI models are investing substantial sums into their design, but the models depend on other people’s data to be of any use. It’s only fair that these companies share the benefits more widely. Just as industries that rely on natural resources face additional taxation to benefit the public, so too should AI endeavors contribute to a broader societal gain beyond the routine corporate taxes they pay.

Microsoft Corp., for example, was taxed on its profits from the sale of Windows, but Windows wasn’t written using the collected works of Stephen King and code from millions of projects on GitHub. The idea that society is due no more from Google LLC for its development of Gemini as from Microsoft for Windows seems to devalue the work of society writ large.

Taxing AI models today, when much of their value derives from publicly available data, and funneling that revenue into a sovereign wealth fund, would allow countries to start protecting the long-term value of this collective resource.

Managing a resource for the public good through tax policy isn’t new. The obvious policy analog in the US is the Alaska Permanent Fund, which uses taxes on oil to benefit future citizens, even those born after all crude oil has been extracted from the state. Every barrel of oil extracted and refined is one less left for future generations, but some of its value is preserved through taxation and investment.

Alaska’s example is funded through a severance tax on oil revenue. It was established in 1976 in the wake of an oil boom, with an understanding that the profits from oil and gas extraction should benefit all Alaskans. Each year, the fund generates a dividend for every Alaskan resident, with the 2024 dividend amounting to $1,702.

On the other side of the world, Norway’s Government Pension Fund acts as a long-term savings program and ensures economic stability during bust cycles. Oil and gas taxes finance the fund, with a tax rate for oil and gas exploration as high as 78%. This has helped make Norway’s sovereign wealth fund the most valuable in the world, with nearly $1.8 trillion in total assets.

The success of these funds shows how tax policy can step in to preserve long-term value from shared public resources. Although such resources can’t be perfectly preserved for future generations, some of the value derived from them can be. The US could take a similar approach to manage the wealth generated by AI’s extraction from the well of human knowledge.

While AI will, and has, benefited society by improving products and services, the economic value generated by AI models likely will be concentrated among a small number of private corporations. This concentration of wealth conflicts with the collective knowledge and resources that fuel these models being publicly derived.

Public resources should generate public value. Their method of creating value places AI companies as distinct from other tech companies and more like those in the resource extraction industry.

The most effective structure to implement a tax on AI may be based on their parameters, the fundamental building blocks of language learning models. Parameters represent the individual bits of information the model has gleaned from data it was trained on. Large models, such as OpenAI’s GPT-4, have trillions of these parameters.

The key to this structure would be to view a tax on model parameters as akin to Alaska’s severance tax—each barrel of oil represents an amount of the resource that has been extracted, and so too for model parameters. In the latter case, the resource being extracted is human knowledge, and each parameter represents a fraction of information that has been extracted and, for lack of a better term, consumed.

Such a tax would start at a miniscule rate, in the millionths or billionths of a penny per parameter, to ensure it doesn’t stifle AI innovation. However, given the size of the leading models, any parameter tax would generate substantial revenue from each large model developed—in the tens or hundreds of millions of dollars.

Channeling resources into a US sovereign wealth fund would ensure profits from AI’s development aren’t lost to private corporations at the expense of taxpayers. A sovereign wealth fund based on AI revenue likely wouldn’t rival the massive investment vehicle in Norway or even Alaska.

But even a modest-sized fund could provide targeted safety nets for future generations, such as funding subsidies for broadband or investments in next-generation communications infrastructure.

AI development is moving rapidly—the best time to tax AI models was yesterday, but the second-best time is today.

Andrew Leahey is a tax and technology attorney, principal at Hunter Creek Consulting, and adjunct professor at Drexel Kline School of Law. Follow him on Mastodon at @andrew@esq.social

Read More Technically Speaking