AI’s growth forecasts still need guardrails

Anthropic's latest economic promise comes with a warning about control. Speaking in Bangkok on October 17, co-founder Jack Clark said an internal model envisaged US output approaching $44 trillion by 2030, compared with just over $30 trillion today, Reuters reports.

That scenario depends on AI continuing to improve and new industries emerging. It is a company scenario described by Clark, not a measured productivity gain or an independent consensus forecast. Clark also called for international cooperation on regulation and warned that the window for safeguards was narrowing.

The useful question is not simply how large the upside could be. It is who can identify unacceptable risks, require changes and stop consequential activity before harm occurs.

From a metaphor to responsibilities

At the IMF meetings a day earlier, Kristalina Georgieva compared AI's acceleration to racing that needs protective barriers. Clark argued that companies cannot make these decisions alone, according to AP's October 16 report. Agreement that safeguards matter does not establish agreement on what should be required.

A less spectacular development offers something practical to examine. On October 15, the Financial Stability Board finalized twelve practices for financial institutions adopting AI. They cover accountable leadership, documented uses and risks, data quality, performance checks, human oversight, cybersecurity and dependence on outside providers.

The report's introduction makes its limits explicit: these practices are not a binding international standard, and following them does not replace local legal obligations. Its treatment of agent and multi-agent risks is still developing. This is guidance to assess, not evidence that finance has made AI safe.

A useful test for your own work

Our reading is that economic forecasts and operational permission belong in separate decisions. A convincing growth story cannot tell you whether an agent should have access to customer records, approve a payment or make an irreversible change.

For a team using ChatGPT or another agent, start with one consequential workflow. Name its owner, document the information and actions it can access, decide which outputs require review, and establish how a person can interrupt it. Keep a record of failures and the changes made in response. NIST's risk framework and companion Playbook provide a broader starting point.

Those are practical review questions, not a safety guarantee or a substitute for public oversight. We can value useful AI while asking its developers and the institutions deploying it to demonstrate control. Promised prosperity does not answer that demand.