Anthropic CEO Calls to ‘Cage’ Claude Fable 5 – Is Immediate AI Regulation Needed?

AI Scaling and Policy Gap

Amodei uses a metaphor of hobbits trying to awaken the slow‑moving Treebeard to illustrate the speed mismatch between AI progress and policy making. In four years, AI models advanced from being unable to write a single line of code to generating the majority of code for leading AI companies, reshaping biology, physics, mathematics, finance, law, and translation. The AI scaling law, supported by more than a decade of empirical evidence, predicts that continued compute growth will produce exponential jumps in cognitive ability, potentially yielding a “Powerful AI” – a genius‑level intelligence housed in a data centre. By contrast, legislative processes take years, creating a widening gap.

Five Mandatory Policy Questions

Regulation & Public Safety

Transparency alone is insufficient; hard regulation is required. AI is compared to aircraft, drugs, and automobiles—technologies whose design or operation errors can cause massive loss of life. Models that exceed a defined compute threshold must undergo mandatory third‑party risk assessments in four domains: cybersecurity, bio‑weapon risk, AI‑control risk, and automated R&D. Government agencies (e.g., FAA‑like bodies) should have authority to block deployment of high‑risk models, enforce security standards, conduct regular red‑team and penetration testing, and require incident reporting. A possible future step is to treat top‑tier AI systems as nuclear‑material‑level assets.

Macroeconomics & Taxation

AI could overturn the long‑standing belief that economic growth is fragile and hard to achieve. If AI surpasses human cognition across most tasks, accelerated scientific, technological, and operational efficiency could generate rapid, strong growth, while massive labor displacement creates extreme inequality. Amodei stresses that durable employment displacement is dangerous. Anthropic operates an Economic Index to measure AI’s labor impact; governments need richer data dimensions. Proposed policy levers include precise measurement of AI‑driven employment effects, wage‑insurance schemes, retention tax incentives, retraining subsidies, and, if displacement becomes permanent and large‑scale, mechanisms such as universal basic income or capital‑gain taxes.

Accelerating Positive AI Impact

There is an asymmetry: too little regulation of AI itself, yet excessive regulation of downstream technologies. In biomedical drug development, FDA and EMA pipelines typically take 7–8 years, based on pessimistic assumptions about candidate viability. AI can dramatically increase the number of drug candidates, improve effect size and safety metrics, enable therapies for previously untreatable diseases, and create entirely new therapeutic classes. Without adapting the approval pipeline, AI would merely create a bottleneck. Amodei recommends that regulators begin incorporating AI‑driven simulation and analysis methods—AI pharmacokinetic modeling, AI toxicology prediction to replace animal tests, precise dose selection, biomarker validation, synthetic control groups, and alternative endpoints—and adopt more aggressive accelerated‑approval mechanisms for interventions that demonstrate sudden, high efficacy.

National & Citizen Aspects

Powerful AI in the wrong hands could become a weapon of mass destruction. Risks include autonomous drone swarms executing illegal commands and AI‑enhanced surveillance that infers deeply private details about individuals, outpacing existing civil‑liberty safeguards. Proposed measures are: accountability rules and outright bans for fully autonomous weapons, closure of data‑broker loopholes that allow mass data harvesting, guaranteeing citizens access to AI tools comparable to state capabilities to protect legal rights, and checks on corporate AI power that could grant quasi‑state authority.

Maintaining Leadership

Amodei elevates AI’s strategic importance to the level of nuclear weapons. A nation possessing advanced AI could dominate militarily and economically; the advantage over a nation without such AI is likened to WWII navies versus medieval swordsmen.

Developer Backlash

Developers report that Claude Fable 5 often silently degrades to older versions such as Opus 4.8 or becomes less capable on AI‑related topics, prompting widespread criticism on social platforms.

Reference: https://x.com/DarioAmodei/status/2064781775247950326