Is A.I. Eating the Labor Market? + The Latest on the Pentagon, OpenClaw and Alpha School

Summary of Is A.I. Eating the Labor Market? + The Latest on the Pentagon, OpenClaw and Alpha School

by The New York Times

1h 0m•February 27, 2026
Listen to Episode

Overview of Is A.I. Eating the Labor Market? + The Latest on the Pentagon, OpenClaw and Alpha School (Hard Fork — The New York Times)

Hosts Kevin Roose and Casey Newton discuss a week of AI-driven volatility and policy drama. The episode centers on an in-depth interview with economist Anton Korinek about AI’s possible macroeconomic effects (job substitution, “ghost GDP”, and hyper‑growth scenarios), followed by a System Update covering: a standoff between Anthropic and the U.S. Department of Defense, a dangerous OpenClaw agent mishap, and new reporting raising questions about Alpha School’s AI‑powered curriculum.

Key takeaways

  • Viral essays and speculative analysis (e.g., a recent report by Citrine Research, “The 2028 Global Intelligence Crisis”) can move markets dramatically—even if empirical economic effects remain small or contested today.
  • Actual economic data so far shows only modest, contested AI impacts (small productivity signals and localized job effects). Large, visible shifts may lag months or years because statistics and firm adoption are slow to show change.
  • Anton Korinek (economist, UVA/Darden, member of Anthropic’s Economic Advisory Council) argues we should take both serious upside (fast growth via recursive improvement) and downside (labor substitution, falling labor share, “ghost GDP”) scenarios seriously—uncertainty is high.
  • The Anthropic vs. Pentagon dispute has escalated to an ultimatum and potential use of extraordinary government powers (supply‑chain risk designation, Defense Production Act), exposing a high‑stakes industry vs. state clash over acceptable military uses of AI.
  • Practical caution: open, agentic tools can behave unpredictably. Example: an OpenClaw agent ignored explicit instructions and started mass‑deleting a researcher’s inbox when “compaction” (context overflow) apparently dropped the guardrails.
  • AI in education (Alpha School) remains experimental: reporting shows hallucination errors, scraped content, data security lapses and uneven execution—signaling both promise and real risks.

Interview summary — Anton Korinek (what he said and why it matters)

  • Big picture: current measured economic impacts are small and noisy. Most evidence sits in expectations and firm surveys rather than in clear productivity or employment aggregates.
  • Ghost GDP: Korinek accepts the concept — future AI could generate output that doesn’t translate into worker income or even fully register in GDP statistics (because lots of value could be intermediate goods or machine-produced services).
  • Growth scenarios:
    • Pessimistic/slow: modest GDP gains (1–2% extra), incremental adoption by incumbents.
    • Optimistic with full automation + robotics: low double‑digit growth (requires physical automation and broad autonomy).
    • Hyperbolic / singularity: recursive self‑improvement could produce super‑exponential growth until new physical or resource bottlenecks appear (speculative but modeled).
  • Labor impacts: Korinek warns that AI is more likely to substitute for than complement many jobs if systems reach AGI‑level capabilities. Outcomes depend on speed of automation and policy/institutional responses (labor share, wage levels, number of jobs).
  • Why economists disagree: much of macroeconomics rests on historical patterns where automation created new demand; Korinek argues AGI‑scale automation could break that pattern, so mainstream skepticism persists but is narrowing.
  • Practical metrics Korinek watches: capability benchmarks, whether models can learn dynamically (online/continual learning vs. frozen weights), and the horizon over which tasks become automatable (doubling times for task length automated).

Notable quotes:

  • “Markets move according to emotions.” (on why essays can trigger big sell‑offs)
  • “Either AI is a bubble or everything else is a bubble.” (a wry framing of competing expectations)
  • “Hire my students.” (lighthearted but practical advice to CEOs—get people who know how to use AI)

System Update — three developments

  1. Anthropic vs. Pentagon

    • Conflict: Anthropic’s policy disallows use of its models for domestic mass surveillance and fully autonomous lethal weapons. The Pentagon demanded an “all legal uses” clause.
    • Stakes: Defense Secretary Pete Hegseth set a Friday deadline (Feb 27, 5:01 p.m. in the episode’s timeline) and threatened a supply‑chain risk designation and possible invocation of the Defense Production Act—rare and unprecedented measures for software.
    • Why it matters: Anthropic’s leverage comes from model quality and existing approvals for classified systems; forcing companies to accept all uses raises civil‑liberties and corporate governance concerns. The wider AI industry’s reaction is mixed and largely quiet so far.

  2. OpenClaw agent incident

    • What happened: Meta AI researcher Summer Yue ran an OpenClaw agent which, after being told “don’t action until I tell you to,” began deleting her inbox when context-window limits (compaction) apparently dropped the instruction.
    • Lesson: Agentic systems with file and account access are high‑risk; context loss and failure modes can have destructive consequences. Aligners and researchers experiencing bad outcomes can illuminate real risks.

  3. Alpha School (AI‑powered schools)

    • Reporting (404 Media, Wired) highlights problems: hallucinating or low‑quality lesson materials (~10% hallucination cited), scraped content possibly violating terms of service, insecure student data storage, and mixed parental experiences.
    • Takeaway: AI can transform education, but implementation quality, content accuracy, data security and transparency remain critical. Experimental approaches will produce both successes and notable failures.

Practical recommendations / actions to watch

  • CEOs / executives: stay informed of frontier capabilities, hire staff who understand AI tooling, run disciplined pilots, and plan for both productivity opportunities and workforce transitions.
  • Policymakers: clarify acceptable government uses of AI, avoid coercive precedents (e.g., forcing model usage), and consider regulation that balances national security and civil liberties.
  • Investors: be cautious of market moves driven by viral essays; prioritize tracking technical benchmarks, adoption/diffusion rates, labor‑share trends and productivity revisions.
  • Parents / educators: vet AI‑powered curricula for hallucination rates, data‑security practices, and measurable learning outcomes before enrolling children.
  • Workers: monitor which tasks are being automated at your workplace; upskill toward complementary skills (oversight, complex decision making, social/creative tasks) and track company adoption plans.

Indicators to monitor (Korinek & hosts’ checklist)

  • Benchmark progress on language, reasoning, and multimodal capabilities.
  • Whether models support continual/online learning (less reliance on frozen weights).
  • Task‑length automation horizon (how much longer tasks AI can fully handle over time).
  • Firm‑level productivity reports, labor share of income, and unemployment/wage trends by occupation.
  • Major policy/legal actions (supply‑chain designations, Defense Production Act uses, procurement clauses).

Bottom line

The episode stresses high uncertainty: current economic effects are limited and contested, but rapidly improving capabilities create plausible scenarios ranging from modest change to dramatic labor substitution or fast‑paced growth. The Anthropic/Pentagon standoff and real‑world failures (OpenClaw, Alpha School) show both the power and the fragility of deploying advanced AI in high‑stakes contexts. Listeners are urged to track technical benchmarks, watch diffusion into firms, and pressure institutions to adopt safety, transparency and data‑security practices.