While the public may think of taxation as primarily a way to raise money for governments, Professor Reuven Avi-Yonah notes that the bulk of the federal tax code regulates activities. He proposes using that second function as an effective way to control emerging artificial intelligence (AI) technology.
In recent articles and lectures, Avi-Yonah makes his case, specifically focusing on autonomous AI—“basically, AI that is not entirely under the control of the people who program it,” he said.
Avi-Yonah—the Irwin I. Cohn Professor of Law—recently answered five questions about his plan:
1. Why is there a need to regulate autonomous AI?
There are lots of AI programs, including the large language models (LLMs), that are not entirely under the control of their programmers. If they were, they probably wouldn’t be doing the bad things that we attribute to them, like hallucinations and defamation and copyright infringement.
AI has also been guilty of racial bias and other forms of discrimination. In some cases, people commit suicide or harm themselves because of interactions with AI agents.
Then, there are broader issues about AI displacing human beings in the workforce and about the energy consumption of the data centers that LLMs require. The list goes on and on of the bad things that unregulated autonomous AI can do.
2. Your proposal makes a distinction between taxing the AI itself as opposed to its corporate owner. Why is that important?
The US corporate tax is now well over 100 years old. It has accumulated all kinds of problems over time, and those problems are not easy to fix.
More importantly, there is a well-documented history of corporations like the big tech companies being able to avoid the corporate tax very effectively—for example, by shifting their profits overseas and by lobbying. They have a lot of power in Washington.
There’s another problem, which is that all of these companies, other than OpenAI, have other revenue sources that are not AI-based. If you’re trying to specifically address AI, then adjusting the general corporate tax is not the way to do it—because you’ll be affecting other things that have nothing to do with AI.
3. How would your idea work in practice?
The model is the way we treat corporations. Over the years, we have given them various legal attributes—a sort of personhood—in order to achieve particular legal ends. My thought is to take an AI agent and imbue it with a very limited legal personality.
Let’s say the AI is treated as a separate legal person for purposes of opening a bank account. Then you give the AI the instruction to maximize the money in its bank account by all lawful means.
There would be a tax on that, and you could adjust the tax rate up or down depending on the behavior of the AI agent. It would be scored—by the government or an independent nonprofit—on the harm that it causes, and that would affect its tax rate.
If the AI is directed to maximize the amount of money in the bank account, the AI will be motivated to do no harm, in order to minimize the amount of tax.
4. What advantages are there to this model?
This is different from “command and control” regulations—legislation that attempts to regulate AI by saying what it can and cannot do. The problem is that AI is advancing so fast, it’s almost impossible for the government to keep up.
There are other ideas, such as the government taking stock in the AI companies. But I think that is also problematic, in that the government would have too much power over the companies and not enough information to use it wisely.
Fundamentally, my proposal is more flexible. It depends on incentivizing the AI itself to behave in particular ways.
5. Are there any disadvantages to this model?
What worries me the most is who will set the scores. If you let the government—that is, the IRS—set the scores, that gives the IRS a lot of power. We generally don’t like the government setting tax rates without the involvement of Congress.
I would prefer some kind of outside agency to have that power. I think regulatory taxes can be effectively delegated in that way.
If you swallow the basic idea of taxing AI, then I’m sure Congress will get involved and there’ll be all kinds of negotiations about this. But the first step is really to accept the idea that an AI agent can be a separate legal person from the corporation that owns it.