New research ahead of Anthropic’s IPO on how the company will affect society’s capacity to absorb AI safely

The researchers examine Anthropic’s effect on 45 areas, including jobs and the concentration of economic power, autonomous weapons, climate change, mass surveillance, AI agents acting beyond oversight, elections, young people’s mental health, and the stability of the financial system.

The analysis also finds tensions between what Anthropic says and what it does:

• Anthropic asks governments to slow the development of more capable AI across the industry, yet it has contracted computing capacity on the assumption that demand for its models will keep rising as fast as supply allows.
• The rules it asks governments for point towards a world in which AI is held back. Society would be spared most of the harm, but Anthropic could face heavy losses, because most of its commitments for computing capacity are payable however little it uses.
• Its filing concedes that the rules it advocates could slow its growth, but does not show how its commitments would adjust if they were adopted.
• Its filing’s expectations are closest to a world of rapid growth in which the harms of AI are contained. Yet its scale and its AI agents could instead help bring about a world in which AI spreads without limits and does serious harm while the AI sector prospers, the very world the company argues against.

The research sets out systematically how Anthropic could affect both the use of AI and society’s capacity to absorb it, for example:

• It has committed hundreds of billions of dollars to computing capacity ahead of revenue, which gives it an incentive to fill that capacity.
• It releases new models continuously, which rules and institutions may not keep pace with.
• Its AI agents automate work and risk displacing workers faster than they can find new jobs.

• People who get answers from its Claude models have less reason to visit the publishers whose work those answers draw on, undermining independent publishing.
• Its growth concentrates economic power in a few firms, on which customers and governments come to depend.
• Its financing ties with suppliers that also invest in it or lend to it could spread losses through the financial system.

• Its refusal to let its models be used for fully autonomous weapons or mass domestic surveillance protects civil liberties.
• Its safeguards, designed to block a narrow class of help with biological and chemical weapons (at first biological weapons only), limit catastrophic misuse.
• It calls for industry-wide limits on the pace of AI development, which, if adopted, would give rules and institutions time to catch up. This is a position it advocates, not something it necessarily does.

• Independent oversight is stronger because two of the outside evaluators Anthropic names, METR and SecureBio, may publish their findings without its approval, subject to redactions it can require within set categories.
• Its published incident accounts and risk reports give regulators and the public evidence to act on.
• Its Public Benefit Fund could support workers displaced by AI, if it pays out.

The research also assesses whether Anthropic’s positions align with the different beliefs investors hold about AI, and how its conduct could shape the course AI takes over the next five years. Together, these findings give investors a basis for important engagement with the company.

Dr Ben Caldecott, Director of the Oxford Sustainable Finance Group and the Associate Professor of Sustainable Finance at the University of Oxford, said: “Anthropic has a key role in shaping whether society stays within a safe operating space as AI is deployed. It can grow that space and help keep the use of AI within it, or shrink the space and push the use of AI beyond it.” He added: “These dynamics will bear heavily on Anthropic’s future value, and on how investors can help maximise its positive contribution to society. They are therefore both financially material and critical for impact. The Silicon Boundaries analysis equips investors to assess these issues, complementing more traditional investment research. We make six specific recommendations to help investors as they engage in the IPO process and beyond.”