Five-second answer: If AI multiplies productivity but the gains never return as wages, leisure, public support, or cheaper essential services, ordinary people may experience the revolution as “higher expectations with better software.” Medical progress is different: curing disease, preventing disability, or extending healthy years can raise welfare even without winning another economic competition. The real questions are who receives the productivity dividend and whether technology creates benefits people can receive without outranking everyone else.
1. “AI will take every job” is crude. So is “knowledge work is safe”
The ILO’s 2025 global index estimates that about 25% of worldwide employment is in occupations with some GenAI exposure, rising to 34% in high-income countries. Yet its central conclusion is transformation of jobs is more likely than wholesale elimination, at least at present.[1]
That still matters. A department does not need to disappear for labor demand to fall. Ten people becoming six people plus agents can reshape a labor market.
2. “Humans will just approve the output” is already a plausible workflow
Anthropic’s June 2026 Economic Index found higher AI autonomy in Claude Code than in ordinary chat/Cowork use. Agentic environments shift work from repeated human prompting toward larger delegated tasks.[2]
The pattern can become: human does everything → AI assists → AI executes a bundle → human handles exceptions and final judgment.
Professional licenses and expertise may remain important, but “expert” is not the same as “automation-proof.” Physical last-mile tasks are harder to digitize today, yet robotics can move that boundary too.
3. “Picks and shovels will beat AI labs” is a thesis, not a law
If models commoditize and open models become sufficient for many workloads, value can migrate toward scarce inputs: GPUs, memory, power, data centers, and networks.
But software firms can still capture value through distribution, proprietary data, enterprise integration, agents, APIs, and switching costs. Hardware margins can also compress as capacity expands.
The useful question is not “software or hardware?” It is: what remains scarce, and what becomes easy to copy?
4. “AI cannot create revolutionary businesses” goes too far
AI-generated interfaces, copy, and simple apps are easy to imitate. “We use AI” is rarely a moat by itself.
Yet lower production costs can make previously uneconomic niches viable: tiny markets, personalized services, 24/7 operations, and firms run by very small teams.
Commoditization and new business formation can happen at the same time.
5. The real weakness: productivity does not automatically arrive in your bank account
A GPU working 100 times harder does not pay your rent.
IMF work shows that AI could reduce some wage inequalities under certain assumptions while still increasing wealth inequality through capital returns. Outcomes depend on complementarity, displacement, asset ownership, taxation, transfers, and social protection.[3][4]
When productivity doubles, society can convert it into shorter hours, more output, fewer workers, higher profits, more investment, or redistribution. The benchmark sheet does not choose for us.
The bleak comedy version is: “AI can work 24/7, so humans can finally receive tasks 24/7 too.” No thanks.
6. People tired of competition may want progress that does not require winning
Another race for salary, promotion, credentials, market value, and AI fluency is still a race.
A different kind of progress matters more: fewer deaths from infection, earlier cancer detection, treatment of genetic disease, slower cognitive decline, and more years of independent movement.
Medical access is never perfectly equal, but health technology can improve welfare through a route that is not simply “be better than the next person.”
7. In 2026, medicine has become slightly more science-fictional
In April 2026, the FDA approved Otarmeni, the first gene therapy for severe hearing loss associated with biallelic OTOF variants.[5]
In July, Casgevy’s U.S. indication expanded to patients aged two and older with sickle cell disease or transfusion-dependent beta-thalassemia.[6]
In September, the FDA approved Fayuvi, the first treatment designed to alter the course of Sanfilippo syndrome type A rather than merely manage symptoms.[7]
These therapies are narrow, expensive, and medically complex. But they illustrate a major shift: from managing symptoms toward intervening in underlying molecular or genetic causes.
8. AI’s most interesting medical role may be research speed, not chatbot diagnosis
AlphaFold has made more than 200 million predicted protein structures available to researchers and has expanded toward predicting molecular interactions.[8]
Drug development contains a long chain before anything reaches a clinic: target discovery, structural biology, molecule design, patient stratification, and trial design.
If AI shortens that chain, the meaningful metric is not “how many doctors were replaced?” but “how many years were removed from reaching a useful treatment?”
If all the world’s GPUs achieve is faster spreadsheets, humanity may request a refund.
9. “Live to 150” is less grounded than “extend healthspan”
The U.S. National Institute on Aging treats geroscience as the study of how biological aging contributes to chronic disease, with the goal of improving healthspan.[9]
That is not proof that aging has been “solved.” Senolytics remain promising, but a 2025 Nature Aging commentary noted that clear efficacy in humans has not yet been established.[10]
The realistic ladder is earlier detection → better treatment → delayed age-related disease → longer healthy life → only then ask how far lifespan itself can move.
10. Maybe the goal is not “win with AI” but “remove more human misery”
AI debates obsess over which jobs survive, which companies win, and who uses the tools best.
Ordinary people can ask a simpler question: If civilization is spending this much compute, how much disease, pain, disability, care burden, bureaucracy, and compulsory labor can it remove?
We already understand “work faster.”
Now cure something.
Sources
- International Labour Organization — “Generative AI and Jobs: A Refined Global Index of Occupational Exposure” / research brief (2025-05-20) https://www.ilo.org/publications/generative-ai-and-jobs-2025-update Used for the estimate that roughly one quarter of global employment has some GenAI exposure, roughly 34% in high-income countries, and for the distinction between task/job transformation and wholesale replacement ilo.org
- Anthropic — “Anthropic Economic Index report: Cadences” (2026-06-26) Used for observed differences in AI autonomy between Claude Code and ordinary chat/Cowork usage and the shift toward larger delegated agentic tasks anthropic.com
- IMF — “Gen-AI: Artificial Intelligence and the Future of Work” (2024) Used for channels connecting AI adoption, labor income, capital income, inequality, and the role of policy imf.org
- IMF — “AI Adoption and Inequality” (Working Paper 2025/068, 2025-04-04) Used for model-based scenarios in which wage inequality and wealth inequality can move differently, especially through capital returns and asset ownership imf.org
- U.S. FDA — “FDA Approves First-Ever Gene Therapy for Treatment of Genetic Hearing Loss” (2026-04-23) Used for the approval and indication of Otarmeni for severe-to-profound hearing loss associated with biallelic OTOF variants fda.gov
- U.S. FDA — “FDA Approves First Gene Therapy for Young Children with Sickle Cell Disease” / CASGEVY product page (2026-07-01) https://www.fda.gov/vaccines-blood-biologics/casgevy Used for the expanded U.S. indication of Casgevy to patients aged two years and older with qualifying sickle cell disease or transfusion-dependent beta-thalassemia fda.gov
- U.S. FDA — “FDA Approves First Gene Therapy for Pediatric Patients with Sanfilippo Syndrome Type A” (2026-09-17) Used for the first approved treatment intended to alter the disease course of MPS IIIA / Sanfilippo syndrome type A fda.gov
- Google DeepMind — AlphaFold Used for the availability of more than 200 million protein-structure predictions and AlphaFold’s expansion toward biomolecular interaction prediction deepmind.google
- U.S. National Institute on Aging — Geroscience / Fifth Geroscience Summit materials (2026) Used for the definition and current institutional framing of geroscience around biological aging, chronic disease, and healthspan nia.nih.gov
- Khosla, Monroe & Farr — “Towards a personalized approach in senolytic trials,” Nature Aging (2025-09-09) Used for the caution that early human senolytic trials have produced biological signals but clear efficacy in humans remains unestablished nature.com
