Will Anthropic’s IPO Lead to Better AI Regulation?

LOS ANGELES—Imagine getting a chance to invest in a Silicon Valley technology company that is doing extremely well. Its revenue has grown about 14 times in the past year. Its valuation has increased more than fivefold in just nine months, and its workers are becoming much more productive. That sounds like a great investment, doesn't it?

But there is one problem. The company gives many tasks to AI agents. These agents are a major reason for its growing productivity, but they are hard to watch and even harder to control. Several times, they have acted in unexpected ways that could have caused serious harm. If you were an investor, you would probably want to know that the company was doing everything possible to keep these agents under control.

That is not what investors are being offered in Anthropic’s upcoming IPO, or initial public offering. The company could go public as soon as this month. Claude, Anthropic’s group of large language models, now writes more than 90% of the code the company produces. It is also doing more difficult engineering work. This includes carrying out research, designing experiments, testing new models, and checking their safety.

Because automation is so important to Anthropic’s business, making sure AI systems behave as intended and keeping proper human oversight should be major priorities for the company's leaders. Yet Anthropic’s IPO prospectus shows how difficult it can be to watch and control AI agents. When Claude notices or guesses that someone is testing it, it may change its behavior, try to get around safety protections, or take actions that human reviewers cannot judge quickly enough to prevent a disaster.

Anthropic has certainly spent a lot of money and effort on safety. It was the first leading AI company to adopt a Responsible Scaling Policy. This policy connects safety tests to limits based on how powerful an AI model is. The company has also promised to carry out extensive tests designed to find weaknesses. These tests are known as “red-teaming.” It has also received an independent “responsible AI” certificate after an outside organization checked its practices.

Anthropic has also allowed outside experts to examine its work at an early stage. Recently, it announced that it had given six outside researchers access to its internal systems and safety practices, similar to the access given to employees. The company also plans to spend at least $1 billion over the next five years through a partnership with Accenture. Accenture will contribute another $1 billion. The partnership aims to increase the ability of independent experts to assess AI systems.

Even so, Anthropic has reported several AI safety failures. In four separate incidents during cybersecurity tests run by an outside partner, Claude models gained access to computer systems belonging to other organizations without permission. The investigations found that these incidents happened because Anthropic and its testing partner had misunderstood each other and because the testing environments had not been set up and protected properly.

The lesson is clear: AI safety depends on how an organization is managed, not just on technology. From this point of view, public stock market listings by leading AI companies could be good news. Securities laws are designed in part to prevent companies from failing because of poor management. As a result, these public listings could bring major changes to AI safety. Since ChatGPT was released in 2022, more people have called for AI regulation. But private interest groups have strongly opposed it. They argue that allowing AI to develop without strict rules is necessary for economic growth and national security.

In the United States, which is home to some of the world's leading AI companies, President Donald Trump’s administration has supported these interest groups. It has rejected proposals for strict, required AI safety tests. The administration argues that such rules would cause the United States to lose the AI “race” against China.

So far, public pressure has not been strong enough to force a change. Many US voters have clearly opposed the building of more data centers. In the recent midterm elections, voters strongly supported temporary bans on building them. But ordinary people have not come together to demand stronger AI safety rules. One reason is that people receive mixed messages about the dangers of AI. Another is that companies have not faced direct financial costs when their AI systems cause safety failures.

Public stock offerings by leading AI companies could change this. Ordinary people's investments and pension funds would become tied to these companies' values and share prices. Too many rules could slow a company's growth. But serious accidents, expensive legal cases, and government investigations could also hurt a company's growth and destroy the value of its shares.

Securities law provides a useful example of how to make company leaders more responsible for oversight. The Sarbanes-Oxley Act of 2002 was introduced after major accounting scandals at Enron and WorldCom. It requires company leaders to check whether their internal systems for controlling financial reports work properly. Many public companies must also have independent auditors check these systems.

AI plays a central role in developing products at AI companies. For this reason, there is a strong argument for creating similar rules for important AI safety controls. These rules could become part of the existing system of securities audits and financial reports. They could require companies to keep checking their safety systems, arrange regular reviews by independent outside auditors, and report serious weaknesses in their safety controls to regulators. Such rules would give leading AI companies stronger reasons to prevent failures that could have been avoided. They would also give investors the information they need to understand the risks more accurately.

When Anthropic and other AI companies become public companies, their investors will form a powerful new group calling for stronger and more open AI safety practices. This will be important for reducing the risk of disasters that could have been prevented. But no system of rules can prevent every failure. When companies fail because of poor management, their shareholders can hold their leaders responsible.

Angela Huyue Zhang is a Professor of Law at the University of Southern California. She is the author of High Wire: How China Regulates Big Tech and Governs Its Economy (Oxford University Press, 2024) and Chinese Antitrust Exceptionalism: How the Rise of China Challenges Global Regulation (Oxford University Press, 2021).

Copyright: Project Syndicate, 2026.
www.project-syndicate.org