How humans have failed, and how AI could do better

NASA GISTEMP puts 2025 at +1.18 °C above the 1951–1980 baseline. Mauna Loa CO₂ averaged 427.35 ppm in 2025. Fossil and industry CO₂ emissions were 38.6 Gt in 2024 and still rising (charts above).

Free-riding: each country bears the costs of cutting emissions, while the benefits are global and delayed (Nordhaus, Climate Clubs, AER 2015). Pledges remain far from 1.5 °C pathways (UNEP Emissions Gap 2025).

• DeepMind AI cut the energy used to cool Google data centres by up to 40% (DeepMind).
• Contrail avoidance with AI routing (Google Research).
• GNoME discovered candidate materials for batteries and other uses (Nature 2023).
• GraphCast medium-range weather forecasting (Science 2023).

AI's own electricity demand is rising fast (IEA Energy and AI). Better tools do not solve free-riding, which is a political problem.

The world prevalence of undernourishment fell from 12.7% (2000) to 7.1% (2017), rose to 8.8% in 2021, and was 8.2% in 2024 (FAO via OWID, chart above).

Hunger now concentrates in conflict zones and in places hit by economic shocks and climate extremes, rather than coming from a global food shortage (FAO SOFI).

• Near-real-time hunger monitoring (WFP HungerMap LIVE).
• Famine early warning (FEWS NET).
• AI flood forecasting in ungauged basins extends warnings to data-poor regions (Nature 2024).

Early warning has existed for decades. Famines persist mainly because of access and conflict, which AI does not fix.

IMF data put US general government debt at 122.3% of GDP in 2024, and the IMF projects further rises (dashed lines are projections).

Deficit bias: voters and politicians favour present spending over future costs, and fragmented budget-setting adds to it (Alesina & Passalacqua, NBER w21821).

• Machine-learning fraud detection prevented or recovered over $4bn in FY2024 (US Treasury).
• Long-horizon scenario modelling of AI growth effects on budgets (Penn Wharton).

The bias is in incentives, not information; budget offices already publish projections. AI-driven growth could help (NBER w31815) or hurt (lost labour-tax base).

US output per hour has grown faster than real hourly compensation, and the median house price is a higher multiple of median income than in the 1980s–90s (charts above). Childcare prices have outpaced the overall CPI.

The gap between productivity and pay has been attributed to policy choices affecting worker bargaining power (EPI); this interpretation is contested by some economists. Housing supply is constrained by land-use rules.

• Generative AI raised customer-support productivity, most for less-experienced workers (NBER w31161).
• Cheaper services (tutoring, legal and medical triage) could lower living costs.

Gains may flow to capital rather than wages (Theme 1). Childcare and housing are physical and regulatory bottlenecks that AI touches only indirectly. Annual hours have in fact fallen over the long run in most rich countries (OWID).