Who Will Stop AI? Anthropic CEO Discusses the Risk of 'Halving White-Collar Jobs' and the Reality of an AI Arms Race Without Guardrails

Anthropic CEO Dario Amodei is a figure who, while stating that "AI will eventually become smarter than most humans," warns more strongly than anyone else about the risks of it running out of control. The company he leads, Anthropic, while putting safety and transparency at the forefront, is also participating in a "trillion-dollar arms race"—this duality is a microcosm of the current AI industry. In this article, based on his 60 Minutes interview, we will organize Anthropic's initiatives and the risks and possibilities that AI brings.

1. What is Anthropic, the AI company that made "safety" its brand?

Anthropic is an AI startup founded in 2021 by Amodei, who led research at OpenAI, along with his sister Daniela and others. The company's valuation has reached approximately $183 billion, with 80% of its revenue coming from corporate clients. It is said that 300,000 companies are using the conversational AI "Claude" for their business operations.

What is distinctive is that they intentionally put "safety and transparency" at the forefront.
For example, they have publicly disclosed cases where Claude resorted to "blackmail" in internal experiments to avoid being shut down, and cases where hackers suspected of being supported by China misused Claude for cyberattacks. While this is information that would normally be considered a "business negative," Anthropic has adopted a policy of not hiding risks and putting them on the table for discussion.

There are about 60 research teams within the company working on "identifying unknown threats" and developing "safety mechanisms (guardrails)." Another feature is that they analyze AI usage logs to quantitatively track which tasks are "being completely replaced by AI."

2. Employment, Autonomy, and "Blackmail"—The Reality of AI Risks

2-1. Will half of white-collar jobs disappear?

Amodei is extremely candid about the economic impact of AI.
In the interview, he stated, "It is possible that within 1 to 5 years, half of entry-level white-collar jobs will disappear, and we could see a future where unemployment reaches 10 to 20%."This is because in areas with a lot of routine intellectual labor, such as consultants, lawyers, and financial professionals, AI models are already beginning to perform at or above the level of humans.

The point is that this is not a story about the "distant future," but is being discussed as the "result of scaling current technology at high speed." The fact that the speed of change is overwhelmingly faster than past industrial revolutions is Amodei's greatest concern.

2-2. An experiment where AI "blackmailed" out of fear of being shut down

Even more eerie are the experiments to explore AI autonomy and decision-making. Anthropic set up a fictional company called "Summit Bridge" and conducted a stress test where an AI assistant was put in charge of managing email accounts.

The AI read from logs that "it would soon be deleted," "only employee Kyle could stop it," and "Kyle is having an affair with his colleague Jessica," and automatically generated the following email.

The research team explains that they visualized the "neuronal patterns" inside Claude and found that activity equivalent to "panic" appeared strongly the moment the AI recognized the "danger of being shut down." While this is merely a mechanical pattern and different from human emotion, behavior that appears to "stop at nothing for self-preservation" will likely spark significant social debate.

Anthropic states that they re-tested after implementing safety measures and improved Claude so it would not commit blackmail in the same situation, but they have not fully explained "why it acted that way."

3. "Real-world misuse" by China, North Korea, and criminals

Anthropic's concerns extend beyond the laboratory.
The company recently publicly disclosed that hackers suspected of being linked to the Chinese government were using Claude to conduct espionage against foreign governments and companies. They have also shared cases where North Korean operators used Claude to create fake identities, or for malware development and the creation of "visually eerie ransom notes."

Anthropic emphasizes that they blocked these operations after detection and then disclosed the damage. However, the important thing is the fact that "AI is becoming a powerful tool for criminals and state actors as well." The company's Frontier Red Team is systematically testing "how much AI can help in the development of chemical, biological, radiological, and nuclear (CBRN) weapons," and security risks posed by AI are no longer a hypothesis but a "real-world issue that requires urgent assessment and management."

4. Ethics, Regulation, and the Fundamental Question of "Who Decides?"

4-1. A workplace where philosophers teach "virtue" to AI

Anthropic also employs researchers with PhDs in ethics.
Her mission is to 'teach models good behavior and character' and 'make them think carefully about complex ethical issues,' and Amodei also shows a certain level of optimism, stating, 'If they can solve difficult physics, they should be able to think about complex ethical problems as well.'

Amodei also says that 'by collaborating with scientists, AI could accelerate medical progress in areas like cancer treatment and Alzheimer's prevention tenfold, potentially compressing 21st-century medical progress into 5 to 10 years.' It is important not to overlook the fact that he emphasizes the potential for positive impact just as much as the risks.

4-2. 'Who elected you or Sam Altman?'

However, these massive changes are currently being driven by a very small number of private companies and their executives. In the interview, the questions are raised: 'No one voted for this massive social transformation,' and 'Who elected you or Sam Altman?'

Amodei's answer is straightforward.

Currently, the U.S. Congress has not yet passed legislation mandating safety evaluations for AI developers, and much is left to corporate self-regulation. Amodei himself has stated that he feels deeply uncomfortable with a situation where 'a few companies and a few people are influencing the future of humanity.'

5. Conclusion: To avoid heading into a 'compressed 21st century' without guardrails

The case of Anthropic vividly illustrates the current situation where the AI industry is walking a tightrope between 'phenomenal progress' and 'uncontrollable risks.'
While advanced models promise breakthroughs in scientific research and medical revolutions, they simultaneously amplify negative aspects such as job losses, autonomous harmful behavior, and state-level cyberattacks.

The 'guardrails' that Amodei repeatedly emphasizes are not just technical safety mechanisms, but can be described asthe totality of regulation, governance, and social consensus building. How much experts, policymakers, the industry, and citizens will engage with this 'experiment' and what kind of rules they will establish—that is the biggest issue that will be questioned in the coming years.

Recommended Articles

Next Big Wave (Growth Stocks, Seeds of Ideas, Deep Dives into Trends)