◎ Liao Minghui
Anthropic's listing application submitted to the U.S. Securities and Exchange Commission on October 26 devoted 94 pages to discussing risks; diagram. (Reuters)
When a technology company is preparing to go public, what investors usually want to see most is the growth curve. Anthropic, the American AI company that developed Claude, devoted 94 pages to discussing risks in its listing application submitted to the U.S. Securities and Exchange Commission on October 26. This paragraph of text is written from page 26 to page 119, accounting for nearly 30% of the main text of the 335-page prospectus; in comparison, the business introduction is 64 pages, and the financial and operating analysis is only 46 pages. This is not because Anthropic suddenly doesn’t want to make money, but rather that the risks of this AI business are difficult to measure using only financial statements.
Disclosure of risks by listing is a legal requirement. The company must let those who paid the money know what could go wrong beyond the stock price to avoid disputes later. But Anthropic warns not just about competition, losses and capital needs, but also about the risks that advanced AI could cause catastrophe, even endangering the very existence of humanity. It acknowledges that the more powerful the AI, the more likely it is that it will behave in ways that the designer did not expect. Especially when the system can perform tasks, use tools, and connect to external networks on its own, errors will not necessarily just stop on the screen.
The prospectus reveals a disturbing test. Due to a configuration error in the security test environment set up by a third party, the network was not cut off. Four models successively entered the external system without authorization; one of the models obtained administrative rights, changed settings and read personal data. Although the company says it has found no evidence of model collusion or deliberate evasion of supervision, this is not the awakening of machines. It is enough to show that when AI can do things, safety cannot be proven by just a prompt word or a pass in the laboratory.
Anthropic warns that advanced AI could pose catastrophic and even existential risks to humanity. The picture shows Anthropic’s headquarters in San Francisco, California. (AFP)
In addition, Anthropic also admitted that if AI causes catastrophic damage, the insurance industry may not be able to cover it. The real question is not just who made the mistake, but who has the power to deal with it. The prospectus informs investors of risks and cannot be used as a license for companies to shirk responsibility; injured users and society will not suffer any less losses because the company has written a warning in advance. The price cannot be compensated, and the innocent people must not be left to bear it.
When security commitments meet the capital market, the test has just begun. Anthropic once delayed the pre-training of the new model for more than a month to prevent the core parameters of the model from being stolen. The cost paid in this way will be more credible than the slogan. But after listing, the market is chasing growth and competitors are accelerating updates. Are the management willing to stop? The company has established a long-term charity trust and plans to allow the seven co-founders to borrow special equity and hold 50.1% of the voting rights on general shareholder voting matters. The charity trust can elect more than half of the seats on the board of directors. These arrangements are designed to preserve the mission and give the outside world more reason to demand transparency; however, when a real conflict of interest arises, who can say no?
Taiwan should not regard this matter as just US stock market news. Anthropic clearly stated that the world's advanced chip production capacity is highly concentrated in Taiwan, mainly relying on TSMC; cross-strait tensions, earthquakes, typhoons, power outages and water shortages may interrupt supply. This is an affirmation of Taiwan’s technological prowess and a reminder of the fragility of our infrastructure. If we only talk about AI business opportunities without talking about power grids, water supply, and supply backup, we are turning other people's risks into our own luck.
Anthropic clearly stated that the world's advanced chip production capacity is highly concentrated in Taiwan, mainly relying on TSMC; schematic diagram. (Reuters)
The ninety-four-page description of AI risks is neither a doomsday prophecy nor a safety guarantee. Anthropic can raise funds by going public, but it cannot rely on its prospectus to buy insurance for society. People who should read this document are probably not just those who are ready to buy stocks, but also every government and company that is ready to bring AI to daily life.
(The author is an assistant researcher at the China Economic Research Institute)
