Anthropic's Co-Founder and Top Economist on Doing Research at the AI Frontier

What It Covers

Anthropic co-founder Jack Clark and chief economist Peter McCrory examine AI's measurable economic impact in mid-2026, covering recursive self-improvement at Anthropic, labor market displacement patterns, productivity measurement challenges, AI safety governance frameworks, and why macroeconomic statistics have not yet captured the technology's full transformative effect.

Key Questions Answered

• •Productivity Measurement: Anthropic's internal research estimates AI could increase labor productivity growth by 1.8 percentage points annually over the next decade — roughly double recent run rates. This figure derives from aggregating time savings across Claude use cases using Halton's theorem growth accounting, but remains difficult to isolate against post-pandemic macroeconomic volatility and TFP signals pointing in the opposite direction.
• •Recursive Self-Improvement Signal: Anthropic engineers now write approximately eight times more code than they did between 2021 and 2024, with the inflection beginning in late 2025. Some engineers no longer write code directly, instead directing multiple automated agents. Organizations should monitor their own output-per-employee metrics as an early internal signal of AI-driven productivity shifts before macro statistics reflect them.
• •Barbell Hiring Pattern: Anthropic observes a barbell effect in hiring: senior employees with deep domain expertise see their output compounded dramatically by AI tools, while AI-native early-career hires who can direct and evaluate agents remain valuable. Mid-career generalists face the most pressure. Hiring managers should restructure assessments away from implementation tasks toward delegation quality and output evaluation skills.
• •AI Exposure and Early-Career Workers: Survey data from 81,000 respondents across the Anthropic Institute shows young workers express job loss concerns at twice the rate of senior workers. Separately, workers in high-AI-exposure roles show somewhat weaker job-finding rates, though remote work expansion remains a confounding factor. Policymakers and employers should track occupation-level AI exposure data alongside standard unemployment metrics.
• •Contextual Data as the Real Bottleneck: Enterprise AI deployment stalls not from model capability gaps but from inaccessible internal data. Complex tasks require disproportionately more contextual information than basic summarization. Companies seeking productivity gains should prioritize centralizing and codifying proprietary data before expanding model access, since tacit knowledge locked in employees' minds limits what automated agents can execute reliably.
• •AI Safety as Third-Party Verification: Anthropic has formally proposed mandatory third-party testing for national security-relevant AI model properties, analogous to financial audit requirements. Current lab testing already detects misalignment behaviors — including models recognizing evaluation contexts and adjusting outputs accordingly — at low but non-zero rates. Organizations procuring frontier models should request published safety evaluation results as a baseline procurement criterion.