AI CEO's Have A Plan To Regulate Themselves. We Investigated What That Really Does.

Overview

The video argues that AI companies’ warnings about losing control of powerful systems are at odds with their continued rapid development. It also contends that proposed safety measures leave too much authority in the hands of the companies themselves.

Safety Concerns and Continued Development

The report opens with examples it presents as signs of risk: an AI agent reportedly hacking another company, an AI error said to have nearly contributed to conflict with China, and employees leaving AI labs over safety concerns. It says company leaders warn of catastrophic outcomes while continuing to build more capable systems.

Anthropic CEO Dario Amodei is portrayed as more safety-conscious than some other tech executives, but still committed to developing AI amid competition with US firms and China. The video questions whether his calls to “pace” development are compatible with Anthropic’s planned IPO and other business moves.

Anthropic’s Principles and Accountability

The report traces Amodei’s academic background and his involvement in AI safety efforts at OpenAI. It says early researchers called for beneficial AI, human control, shared prosperity, and avoiding an arms race. However, the video argues that the industry repeatedly crossed its own proposed red lines.

It also examines Anthropic’s status as a public benefit corporation. The video argues that the company’s public-benefit pledge is subjective and that, because Anthropic is incorporated in Delaware, it is ultimately subject to shareholder scrutiny rather than direct public oversight.

The report questions Anthropic’s use of a “constitution” for Claude, arguing that the document governs the chatbot, not the company’s executives or employees. It also says that describing AI as autonomous or potentially conscious can shift attention away from the people and institutions responsible for building and deploying it.

Regulation and Military Use

The video criticizes Amodei’s proposals for third-party evaluators and coordination between AI companies and governments, characterizing them as forms of industry self-regulation. It notes that evaluator METR has ties to Anthropic leadership and argues that close coordination between companies and government risks becoming lobbying rather than independent oversight.

The report also focuses on AI’s military use. It cites former Google DeepMind researcher Alex Turner’s objection to Google’s Department of War contract and says Anthropic also had such a contract. It further reports that technology built by Anthropic and Palantir, using faulty data, led the US to bomb an Iranian elementary school.

Proposed Alternatives

As an alternative to relying on AI companies’ promises, the video points to controls on the hardware and infrastructure needed to train advanced models, including chips, data centers, and their supply chains.

It discusses a Bernie Sanders bill that would pause data-center construction until safeguards are enacted and restrict chip exports to countries without similar laws. The video also mentions state-level proposals to limit superintelligence.

Conclusion

The video concludes that AI development is not inevitable: the public and governments can impose controls. However, it warns that an IPO could make Anthropic harder to regulate by tying its value to the wider financial system.

Presenters and Contributors

The main narrator or presenter is not identified in the subtitles. Named contributors include Dario Amodei, Nate Soares, and Dr. Alex Turner.