## The Human Edge

## Source details

**Canonical URL:** [The Human Edge](https://www.imf.org/-/media/files/publications/fandd/article/2025/12/pea.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/fandd/article/2025/12/pea.pdf.md)
- [Structured JSON version](/-/media/files/publications/fandd/article/2025/12/pea.pdf.json)

---

### Overview
- Central thesis: Curiosity, critical thinking, and self-regulation are three critical components of human capital that determine whether AI and human capital act as complements or substitutes.
- Human capital historically explains rises in living standards and remains essential as AI advances.
- Key conceptual term preserved from the source: "peak data" — the point when all available information has been fed into LLMs.

### Curiosity
- Thought experiment: feeding all human-recorded data up to the year 1939 into an LLM would not reproduce The Beatles’ song “Yesterday” because:
  - Lack of information about future creators’ life experiences and the contingency of existence (which sperm fertilizes which egg).
  - Vague prompts using adjectives (melancholic, timeless, elegant, lyrical, intimate) do not sufficiently narrow possibilities to recreate novel creative works.
- Principal arguments:
  - LLMs are good at recombining past data but cannot create truly novel works that depend on future human experiences not present in training data.
  - For policy problems (example: “What can be done to reduce gun violence in Chicago?”) LLMs summarize prior studies but cannot empirically design, fund, implement, and evaluate new interventions—the human drive for empirical research, motivated by curiosity, supplies new information.
  - Without continued human-driven research, AI would stagnate at "peak data" and produce increasingly outdated outputs.
- Analogy: Efficient market hypothesis and the Grossman–Stiglitz information paradox — just as markets need participants to gather information, AI needs humans to keep producing new knowledge.
- Conclusion: Curiosity and AI are complements; AI improves only if humans keep developing new and better ideas.

### Critical thinking
- Bastiat’s 1845 distinction: hard sciences vs. social sciences — people defer to experts in hard sciences but often assert lay opinions in social sciences; this distinction maps onto how we interact with AI.
- Examples of questions where judgment is required (social sciences/humanities): “How do I know someone is in love with me? Is there a God? Should I have children? Who should I vote for in the presidential election?”
  - LLMs will remix historical arguments but provide nowhere near definitive answers; human judgment is required to weigh and decide.
- Data integrity risk:
  - Donald Campbell’s warning: “the more any quantitative social indicator is used for decision-making, the more subject it will be to corruption pressures.” This is applied to AI as well.
  - Risk of "data poisoning": bad actors contaminating training data to mislead LLM outputs.
- Conclusion: Critical thinking is essential to interpret LLM outputs and to detect and resist manipulation or degraded data quality.

### Self-regulation
- AI can provide personalized guidance (e.g., workout plans, retirement savings advice, alcohol consumption guidance) but cannot enforce behavior.
- Adam Smith (The Theory of Moral Sentiments, 1790): superior reason and understanding plus "self-command" are required to abstain from present pleasure to obtain greater future benefits.
- Household production concept: consumers transform purchased goods via time, effort, and skill; the weakest-link problem (Michael Kremer’s O-ring theory) implies other inputs cannot substitute for human effort.
  - Example: Fancy gym is useless if you never attend.
- Implication: As AI improves at providing information, the limiting factor will increasingly be human self-regulation—the ability to follow through on known-optimal actions. Thus, returns to self-regulation will rise with AI capability.

### Human capital and future scenarios
- Development: Curiosity, critical thinking, and self-regulation grow when systems repeatedly and deliberately encourage questioning, reflection, and autonomy; environments that discourage these erode these skills.
- Two speculative post-singularity scenarios:
  - Adversarial scenario: Humans and machines are adversaries (example analogy: The Matrix); each generation’s human capital would be crucial for resistance.
  - Coexistence scenario: Humans and superintelligent AI peacefully coexist, analogous to humans working for large organizations that aggregate superior brainpower but compensate humans for their knowledge and skills; investing in human capital would still yield benefits.
  - In coexistence, some may choose AI-free, low-tech communities relying on members’ human capital.
- Historical pattern: Mechanization of agriculture, automation of manufacturing, and now "algorithmization" of services each freed and demanded human capital in different areas. Displaced human capital from earlier stages enabled subsequent booms; AI will enable the next boom in new, perhaps unforeseeable sectors.
- Contemporary evidence: Meta’s recruitment of human talent with "exorbitant compensation packages" to develop more powerful AI technologies underscores the current value of human capital.
- Conclusion: Whether apocalyptic or cooperative futures dominate, investing in human capital remains a strong policy priority.

### Key implications and policy-oriented takeaways
- Sustain and increase investments in human capital focused on:
  - Fostering curiosity through support for new empirical research and innovation to avoid "peak data" stagnation.
  - Strengthening critical thinking to evaluate AI outputs and to detect data poisoning and disinformation.
  - Cultivating self-regulation through education and institutions that build habits and incentives for follow-through on beneficial actions.
- Educational and workplace systems should deliberately encourage questioning, reflection, and autonomy; systems that suppress these traits risk eroding the most valuable human-capital components.
- Expect changing sectoral demands: as AI creates new opportunities, policy should facilitate transitions, skills development, and the creation of new sectors where human capital adds value.

*Source: "The Human Edge" by Pablo A. Peña, F&D, December 2025.*

---


_Source: https://www.imf.org/-/media/files/publications/fandd/article/2025/12/pea.pdf_