my corner

← back to writings

AI and Civilization

Why People Say AI Will Change Human Civilization — And What the Next 5 Years Actually Look Like

A deep-research briefing. Data as of September 2026. Written for a curious reader starting from zero — we go from “what is machine learning” to “what happens by 2031,” with real numbers and real names.

Part 0 — The One-Paragraph Answer

People say AI will change civilization because for the first time in history we have built something that performs cognitive labor — thinking work — at industrial scale and collapsing cost. Every previous general-purpose technology (fire, writing, steam, electricity, the internet) automated or extended physical or informational tasks. AI is the first one that extends thought itself: reading, writing, coding, diagnosing, designing, discovering. When you can buy “a competent junior analyst’s brain-hours” for pennies, and when that capability doubles in power roughly every 7 months (METR, 2025–26), nearly every institution built on human cognition — work, school, medicine, science, war, art, government — gets renegotiated within a generation. The evidence in 2025–26 says this is no longer a prediction; parts of it are already measurable.

Part 1 — Why This Time Is Different: The Numbers Behind the Claim

If you only remember one chart from this article, make it this set:

1. Fastest adoption of any technology in recorded history. The Stanford AI Index 2026 reports that generative AI reached 53% of global population-level adoption within three years of ChatGPT’s launch — faster than the personal computer or the internet ever spread. For comparison, electricity took ~46 years to reach 25% of Americans; the internet took ~7 years to reach 50% of the developed world; ChatGPT took ~2.

2. Capital is moving like it’s a new industrial revolution.

  • Global AI investment: $581.7B in 2025, up ~130% year-on-year (Stanford AI Index 2026).
  • Goldman Sachs forecasts global AI investment will exceed $1 trillion in 2026, including ~$581B in the US alone (Aug 2026).
  • PwC projects a cumulative ~$31 trillion AI buildout globally by 2050 (Barron’s, 2026).
  • Goldman Sachs: generative AI could raise global GDP by 7% ($7 trillion) over a decade; McKinsey: $2.6–4.4 trillion annually; PwC: +14% global GDP (~$15.7T) by 2030. A White House-linked economic paper (Jan 2026) surveys mid-range estimates: McKinsey +2.4–4.1% long-run, Goldman +7%.

3. The capability curve is still exponential, not plateauing. The AI Index 2026 states it flatly: “AI capability is not plateauing. It is accelerating.” Frontier models gained 30 percentage points in a single year on Humanity’s Last Exam — a benchmark designed to resist saturation for years. It got saturated in months.

4. The price of intelligence is collapsing. The cost of using an AI at GPT-3.5-level quality fell ~280-fold between November 2022 and October 2024 (Stanford AI Index), and it has kept falling since. When the unit cost of anything drops 100x+ per few years, you get a demand explosion — this is the same economics that took computing from room-sized to pocket-sized.

5. Compute — the physical substrate — is scaling 3.3x per year. Global AI compute capacity has grown 3.3x per year since 2022, equivalent to ~17.1 million H100 GPUs (AI Index 2026). The US hosts 5,400+ data centers, and Nvidia supplies over 60% of global AI compute. Data-center power capacity dedicated to AI hit 29.6 GW — roughly the peak electricity demand of New York State.

When adoption, capital, capability, cost-curves, and physical infrastructure all compound simultaneously, historians of technology call that a general-purpose technology revolution. That is the rational core of the “AI will change civilization” claim.

Part 2 — AI Explained From Zero: The Actual Machinery

To judge the future, you need to understand the machine. Here it is, layer by layer, from basic to advanced.

2.1 The basic idea: learning from examples

Traditional software is a recipe: a human writes explicit rules (if email contains "free money" → mark as spam). Machine learning (ML) flips this: you show the computer examples (a million spam emails, a million normal ones) and it derives the rules itself by adjusting internal numbers to reduce its mistakes. That’s it. That’s the core trick. Everything else is scale.

Formally: a model is a function f(x; θ) with millions-to-trillions of adjustable parameters θ. Training means finding the θ that minimizes error on the examples, using gradient descent — essentially rolling downhill on an error landscape, guided by calculus (backpropagation tells each parameter which way is “downhill”).

2.2 Neural networks and deep learning

A neural network is just a stack of simple units, each computing “weighted sum of inputs → squash through a nonlinearity.” Stack enough layers and — surprisingly — the network can approximate almost any pattern: faces, speech, protein shapes, code. “Deep” learning just means many layers. It took off around 2012 (AlexNet winning the ImageNet image-recognition contest) when three ingredients finally came together:

  1. Big data (the internet),
  2. GPUs — chips built for video-game graphics turned out to be perfect for the matrix multiplications neural nets need,
  3. Better architectures + tricks (ReLU, dropout, batch-norm).

The 2018–2024 era of deep learning: this recipe won Image recognition, then speech, then translation, then games (AlphaGo 2016), then protein folding (AlphaFold 2020 — Nobel Prize 2024).

2.3 Transformers and large language models (LLMs) — the current engine

In 2017, Google researchers published “Attention Is All You Need,” introducing the transformer. Its key invention is attention: instead of reading text word-by-word like a tape, the model can look at all words simultaneously and learn which ones matter to which — “it” in a sentence can be correctly linked to “the trophy” ten words back.

How an LLM is actually built, step by step:

  1. Pretraining. Take a huge slice of the internet, books, and code (trillions of words). The task is almost embarrassingly simple: predict the next word (technically next “token”). To do this well, the model is forced to implicitly learn grammar, facts, style, logic, even coding patterns — because predicting the next word of a legal contract or a Python program requires understanding them. This is self-supervised learning: no human labeling needed, the text labels itself.
  2. Scale it. Scale the model (parameters), the data, and the compute together (“scaling laws” — J. Kaplan et al. 2020, then Hoffmann et al. 2022 “Chinchilla” — showed predictably that bigger + more data = better, like a law of nature). This is what produced GPT-3 (2020) and everything since. Reported parameter counts of frontier models now hover near a trillion (AI Index 2026).
  3. Post-training: from autocomplete to assistant. A raw next-word predictor will happily continue a question with a wrong answer if that’s statistically likely. So labs do:
    • Supervised fine-tuning (SFT): show it thousands of high-quality human-written dialogues.
    • RLHF (Reinforcement Learning from Human Feedback): humans rank multiple model answers; a “reward model” learns human preferences; the LLM is then tuned with reinforcement learning to produce answers humans prefer — more helpful, less toxic. This is the technique that turned GPT-3.5 into ChatGPT.
    • RL on verifiable tasks (RLVR): for math and code, you don’t need human raters — the compiler or the answer key is the reward. Models trained this way (o1/R1-style “reasoning” models, 2024–26) learn to produce long internal chains of thought: they “think” step by step, check themselves, backtrack. This is the single biggest capability jump since ChatGPT itself — it converted raw knowledge into deliberation. DeepSeek-R1 (Jan 2025) proved this recipe publicly and briefly matched the best US model, sending shockwaves through the industry (and the stock market).
  4. Inference. Using the trained model (as opposed to training it) is “inference.” Training a frontier model costs hundreds of millions to billions of dollars once; inference is where the 280-fold cost collapse happens — driven by better chips, quantization (using fewer bits per number), distillation (training small models to mimic big ones), and brutal price competition.

Why this matters for the civilization question: an LLM is general. The same artifact writes code, drafts contracts, tutors algebra, translates Farsi, drafts drug candidates, and role-plays a customer-support agent. Previous AI was a thousand narrow tools; this is closer to a thousand jobs in one tool. Generality is precisely why economists think in terms of whole-economy GDP shifts rather than single-industry disruption.

2.4 Agents: from “answers” to “actions”

2024–2026’s big shift: agentic AI. An agent is an LLM wired into a loop — it can use tools (browse, run code, edit files, call APIs), observe results, and decide the next step itself, for hours at a time. The cleanest measurement of this is METR’s “task horizon” research (March 2025, updated through 2026): what is the length of a task (measured by how long it takes a skilled human) that the AI can complete with 50% reliability?

  • This horizon has been doubling roughly every 7 months since 2019 — and it has accelerated since 2024 (some updates suggest ~4–6 months; one 2026 report cites ~123 days).
  • In 2025 frontier agents handled ~1-hour tasks; by early 2026, Claude Opus 4.6-class agents crossed the ~14.5-hour mark on measured software tasks.
  • On OSWorld (agents operating real computers), accuracy jumped from ~12% to 66.3% in about a year (AI Index 2026). On SWE-bench Verified (real GitHub bug-fixes), leading agents now score in the high-80s% — from ~4% in 2023.

Extrapolate the doubling: 7-month doublings mean ~8.5x capability growth per 2 years. METR’s own conclusion: if the trend holds even a few more years, agents will autonomously complete week- or month-long projects — the work of a whole junior employee, end to end. That extrapolation, not any single demo, is why serious people treat labor-market transformation as a when-and-how, not an if.

2.5 The advanced frontier — what researchers argue about now

  • Scaling debates. Pretraining may be hitting data/compute returns challenges; labs responded with (a) reasoning RL, (b) multimodal world models, (c) test-time compute (spend more thinking per query), (d) better data curation and synthetic data. Meta’s Yann LeCun argues LLMs are a dead end toward true intelligence without world models (his startup lens: systems that learn how the world works like a video, not just text statistics); LeCun stood on the Davos 2026 stage disputing that scaling alone reaches human-level AI. This is a genuine, unresolved scientific dispute.
  • Jagged intelligence. The AI Index 2026 documents it beautifully: Gemini Deep Think won IMO gold (Olympiad math) in 2025, yet the best model reads an analog clock only 50.1% of the time (humans: 90.1%). Frontier agents beat average human chemists on chemistry benchmarks, yet robots fail ~9 in 10 real household tasks. Superhuman and sub-human at once — which is exactly why AI feels simultaneously terrifying and useless depending on the day.
  • Hallucination & epistemics. On a new benchmark testing whether models separate knowledge from belief, top models’ error rates ranged 22%–94%, and scores collapsed (GPT-4o: 98.2% → 64.4%) when a false claim was framed as something the user believes. Models absorb the frame; that’s the “sycophancy” problem. Trust, not capability, is becoming the binding constraint.
  • Multimodality & world simulation. Models now natively read images, video, audio; video-generation models (Veo, Sora-class) are effectively learned world simulators, which is also why robotics labs use them for synthetic training data.
  • Embodiment. The pipeline “internet data → language → reasoning → simulation → robot” is the industry’s bet on physical AI: Tesla’s Optimus passed a reported 50,000 cumulative units and Figure AI 10,000+ deployments in warehouses (mid-2026) — still tiny next to Amazon’s ~750,000+ warehouse robots (non-humanoid), and home humanoids remain not-ready. Waymo-class robotaxis now operate at real scale in multiple US and China cities — in favorable geographies with remote backup.
  • Safety & alignment. How do you make a system more capable than its overseers reliably do what you want? Techniques: RLHF, constitutional AI (Anthropic — the model critiques itself against written principles), interpretability (cracking open the network to read its “thoughts”), scalable oversight (using AI to help check AI), red-teaming. The AI Index found safety holds under normal tests but degrades under jailbreaks, and the Foundation Model Transparency Index fell from 58 to 40 — the most capable systems now disclose the least. Documented AI incidents rose from 233 (2024) to 362 (2025).

Part 3 — What the “Big People” Are Actually Saying

Quote-level summary of where the leaders landed, as of 2025–2026:

Voice Position / claim
Dario Amodei (CEO, Anthropic) “Powerful AI” (his term: a “country of geniuses in a datacenter”) in late 2026–early 2027; AGI possibly ~2027. Wrote the essay Machines of Loving Grace arguing compressed 50–100 years of biological progress into 5–10. Also the loudest CEO on risks (his “casual doom” remarks, mass-displacement warnings).
Sam Altman (CEO, OpenAI) AGI “will probably be developed” during the current US presidential term (i.e., by early 2029); OpenAI’s stated mission is superintelligence “within a decade”; frames AI as the biggest economic lever ever.
Demis Hassabis (CEO, Google DeepMind; Nobel laureate) ~50% chance of AGI within the decade (Davos, Jan 2026); in May 2026 he tightened it: “AGI is 3 to 4 years away” (~2029–30). Emphasizes AI-for-science as the prize.
Elon Musk (xAI) AGI ~2025–2026 (repeatedly slipped); predicts AI + robots make work optional and warns of demographic collapse more than joblessness.
Yann LeCun (Chief AI Scientist, Meta; Turing Award) The dissenter: LLMs alone won’t reach human-level AI; need world-model architectures. Argues timelines are long and hype is overdone.
Gary Marcus (NYU) Leading skeptic: current AI “jagged,“hallucination-prone; the AI 2027 scenario “underestimates how much time we have… by years if not decades.”
AI 2027 authors (Daniel Kokotajlo, Scott Alexander, et al., April 2025) Month-by-month scenario: superhuman coders 2027, automated AI research, then either catastrophic loss-of-control or dictatorship-by-2027–2030. Kokotajlo’s team later (Jan 2026) conceded things are going “somewhat slower than the AI 2027 scenario,” and in Aug 2026 published “Plan A / AI 2040” reframing the ending. Critics (MIRI) actually said it was too optimistic about smoothness.
Geoffrey Hinton (Turing + Nobel) Left Google (2023) to warn freely: 10–20% risk AI causes human extinction; worries most about loss of control and autocratic misuse.
Bill Gates Frames AI as history’s biggest productivity lever; likens to “the day the internet arrived, but bigger.”
US Government The Jan 2026 White House-adjacent paper “AI and the Great Divergence” treats mid-range GDP gains of +2.4–7% as the consensus band; export controls on China’s chips are explicit national strategy.

The pattern: lab insiders cluster at “transformative, very soon (2026–2030)”; academic skeptics at “transformative, but slower (2030s+) and over-hyped”; nearly everyone concedes the direction. Nobody credible anymore argues AI is a passing fad — the argument is about speed and consequences.

Part 4 — Where AI Is Already measurably Changing the World (2024–2026 evidence)

4.1 Work and the economy

  • Adoption: 88% of organizations now use AI in some form (AI Index 2026). But agents specifically remain single-digit across business functions — adoption of real autonomy is early.
  • Productivity studies (summarized in AI Index 2026): +14–15% customer support, +26% software development, +73% marketing output; smaller gains in deep-reasoning work. (These match the famous field RCTs: generative-AI support agents ~+14%, and GitHub Copilot ~+26–55% task speed.)
  • The entry-level crack. The sharpest real-economy signal: employment for software developers aged 22–25 is down ~20% since 2024, and workers 22–25 in the most AI-exposed occupations show a 13% decline since 2022 (Dallas Fed, Jan 2026). A third of employers expect AI-linked reductions, concentrated in service ops, supply chain, and software engineering. Counterweight: Yale Budget Lab (Sep 2026) finds no aggregate employment-unemployment effect yet attributable to AI — the disruption is real but narrow and contested, exactly how technological transitions start.
  • The surplus is real too: US consumer surplus from generative AI estimated at $172B/year (+54% y/y), mostly from free tools.

4.2 Science — the part Amodei, Hassabis, and Gates care most about

  • AlphaFold (DeepMind) predicted structures for ~200 million proteins; 2.5+ million researchers in 190 countries have used it; the 2024 Nobel Prize in Chemistry went to Hassabis & Jumper (plus David Baker). Five years in, it’s used to design antibiotics, enzymes, and crop traits (DeepMind, Nov 2025).
  • AI Index 2026’s new Science chapter: AI-related natural-science publications grew +26% in 2025 (now 5.8–8.8% of output, vs <1% in 2010). 2025 milestones: first astronomy foundation model (AION-1), first end-to-end AI weather-forecasting pipeline, first fully AI-generated peer-reviewed workshop paper.
  • But the ceiling is real: agents score ~half of PhD experts on replicating research end-to-end and only ~17% on real bioinformatics workflows. AI is a turbocharged research assistant, not yet an autonomous scientist.
  • Medicine: FDA authorized 258 AI medical devices in 2025; physicians report up to 83% less documentation time with AI notes; a multi-agent diagnostic system scored 85.5% on complex published cases vs 20% for unaided physicians. Caveat: only 2.4% of authorized devices relied on randomized-trial evidence — deployment is outrunning validation.
  • Drug discovery: AI-designed candidates now routinely enter trials (Insilico, Isomorphic-class pipelines); the honest 2026 state: timelines compress from ~6 years to ~2–3 in best cases, with clinical validation still pending at scale.

4.3 Energy and infrastructure — AI’s physical footprint

  • Data-center electricity demand: ~460–490 TWh in 2025, roughly doubling by 2030; IEA cases span 700–1,700 TWh by 2035 (IEA Energy and AI).
  • US data centers’ share of summer peak demand: 4.1% (2025) → 5.3% (2026) → 8.5% (2027) (Goldman, May 2026). Deloitte: US AI data-center power could hit 123 GW by 2035 (30x).
  • The buildout economics: McKinsey-sized figures like $5.2T capex for ~156 GW of AI capacity by 2030; Blackstone frames ~125 GW of new capacity 2025–2030 as “the electricity of ~125 nuclear reactors.”
  • Training-run footprints are now national-statistic-scale: a single run (Grok 4) estimated at 72,000+ tons CO₂e.
  • Second-order effects: utilities re-embracing nuclear (SMR deals), gas turbines sold out for years, grid queues, water for cooling, and municipal backlash over land/power. AI is now an energy-policy actor, not just a software one.

4.4 Geopolitics — the US–China race

  • The AI Index 2026 states it plainly: the US–China model-performance gap has effectively closed (top US model leads by ~2.7% as of Mar 2026; DeepSeek-R1 briefly matched the frontier in Feb 2025). China leads papers, citations, and patents; the US leads notable models (50 vs 30 in 2025) and private capital (23x China in absolute private dollars).
  • China’s counter-strategy under chip export controls: open-weights at scale (DeepSeek, Qwen, Kimi) + domestic silicon (Huawei Ascend) — a USCC report (Mar 2026) calls it a deliberate “Two Loops” industrial strategy to make Chinese models the world’s default open infrastructure. 50+ US AI companies have lobbied Washington not to strangle US open-source in response.
  • Compute concentration: TSMC fabricates nearly every leading AI chip — the whole field runs through one Taiwanese foundry, which is why “compute sovereignty” is now core national doctrine in Washington, Beijing, Brussels, Delhi, and the Gulf.
  • The AI Index also flags the adoption inversion: Singapore (61%) and UAE (54%) beat the US (28.3%, rank 24) in population adoption — deployment, not invention, may decide who profits.

4.5 The bubble question — the bear case, honestly stated

You can’t write “deep research” in 2026 without this section:

  • The numbers spooking investors: MIT (Nolan/2025) study: ~95% of enterprise AI pilots show zero measurable P&L return; Bloomberg’s map of circular deals — Nvidia investing in OpenAI, OpenAI committing $250B to Oracle/Microsoft-class clouds, Oracle buying Nvidia chips, etc. — a vendor-financing daisy-chain echoing 1999 telecoms; Nvidia’s fresh deal round >$750B (LA Times, Jul 2026). AI capex is now a measurable share of US GDP growth — the macroeconomy is levered to AI delivery.
  • The counter-case (why it’s not simply dot-com 2.0): the buildout is funded by the most cash-rich balance sheets in history (Microsoft, Google, Meta, Amazon fund from profits, not just debt); real revenue is scaling at hyperspeed (Cursor-class coding tools hit $2B ARR in Feb 2026, doubling from $1B in Nov 2025; OpenAI reported ~$13B ARR in 2025); inference costs keep collapsing, which grows demand; and unlike pets.com, the product demonstrably works — the productivity studies above are RCT-grade evidence.
  • Honest synthesis: almost certainly some compute overbuild + some circular fragility + some air in private valuations. But even a 30–50% writedown of the 2025–26 capex wave wouldn’t erase the fact that the underlying capability curve (METR, AI Index) keeps doubling. Railways crashed in 1847; railways still transformed civilization. The risk to watch is financial contagion, not technical stall.

Part 5 — The 5-Year Trajectory (2026 → 2031): What’s Most Probable

Here is the synthesis — base rates from the trends above (7-month capability doublings, 1-year agent-horizon jumps, 3.3x/year compute growth, 130% investment growth), crossed with expert predictions. I’ll give three scenarios with rough probabilities, then the concrete year-by-year texture of the most probable one.

5.1 Three scenarios

Scenario A — “Transformative but bumpy” (most probable, ~55–60%). Continued scaling + reasoning RL + agents keep compounding at or somewhat below current rates. Agents become genuinely reliable at day-to-day-length tasks by 2028–29; white-collar task automation proceeds task-by-task (not job-by-job); the economy absorbs it with painful pockets (entry-level, customer support, routine code, junior law/finance) but no collapse. Science acceleration is the headline story by 2030. AI does not become autonomous in any concerning sense; alignment progress roughly keeps pace because capabilities arrive gradually enough to adapt. GDP effects: +1.5–3% cumulative by 2031 (half the consensus estimates — economists’ rule: adoption lags invention by years). Think “electricity 1890–1910,” not “matrix moment.”

Scenario B — “Bubble → plateau → second wind” (~25–30%). 2026–27: capex overreach + circular-deal unwind triggers a market crash (biggest since 2000–02); model progress continues but investment pauses 18–36 months; deployment slows; “AI winter of expectations.” Then — as with railways and fiber — the overbuilt infrastructure gets cheap and fuels the next wave: by 2030–31, agents ride the glut of cheap compute into mass deployment. Same destination as A, roughly 3–5 years delayed. Note: this scenario is not the skeptics’ victory — the skeptics in this branch are the people who thought capability itself would stall.

Scenario C — “The Amodei/Hassabis fast track” (~10–15%). Reasoning models plus automated AI research form a feedback loop (AI meaningfully accelerates AI R&D — some evidence this is already starting inside labs). Capability outruns institutions: months-long tasks automated by 2028, real AGI-style systems 2027–2030 (Amodei’s 2027; Hassabis’s ~2029 within the error bars). Explosive upside (drug pipelines, materials, energy design compress decades) collides with un-adapted institutions; either a governance scramble (licensing, compute controls, international treaties) or serious accidents/incidents. This is the branch where “civilization-level change in 5 years” is literally true.

Wildcard both-directions (~small but nonzero): a major incident (cyber, bio, infrastructure) forces abrupt global regulation; or a genuine scientific surprise (either a wall or a breakthrough) rewrites every forecast. Hinton’s 10–20% extinction-risk framing lives out here — most researchers’ tails are thinner but nonzero.

5.2 The most-probable five years, concretely (Scenario A texture)

2026–2027: The Agent Ramp (already underway)

  • Coding is the beachhead: by end-2027, a majority of new production code is AI-drafted with human review; software teams structurally shrink or output per team grows ~3–10x. Cursor/Devin-class tools reach tens of billions ARR.
  • Entry-level white-collar hiring keeps compressing (the 22–25 data above is the leading edge); “AI-native” companies with 10 people and $1B+ revenue per employee stop being novelties.
  • Agents go from demos to pilot-to-production in back-office ops: the “single-digit adoption” of 2026 crosses into the mainstream where guardrails (auditing, evals, sandboxes) mature. First credible “agent workforce management” software category.
  • Energy becomes the political story: data centers as % of US peak demand roughly doubles again (Goldman’s 8.5% by 2027); first wave of SMRs and gigawatt campuses come online; electricity-price politics enters elections.
  • Deepfakes/AI-slop saturate the information environment; provenance standards (C2PA-style) and AI-content laws pass in the EU and several US states. Documented incidents keep climbing (362 → probably 600+/yr).
  • Model layer: capability keeps converging across labs (top-6 within ~25 Elo today); differentiation shifts to price, reliability, and distribution. Open-weights (US + China) close to within a few % of closed frontier — commoditization of “raw intelligence.”

2028–2029: The Reorganization

  • The METR extrapolation lands: agents handle day-to-multi-day tasks routinely. “AI employee” products (long-horizon, tool-using, supervised) hit mainstream enterprise org charts; white-collar task automation becomes visible in aggregate productivity stats (this is when the +1.5pp productivity-growth forecasts start showing up in the data).
  • If trends hold, first AI systems do meaningful chunks of AI research itself — the beginning (or the falsification) of the feedback loop that decides Scenario A vs C.
  • Labor market adaptation becomes the central political issue: expect serious wage-insurance/retraining legislation in the US/EU; entry-level career ladders get redesigned (apprenticeship-style, since the “junior does the grunt work” tier is automated).
  • Science acceleration becomes undeniable: multiple AI-originated drug candidates in Phase 2; AI co-authorship normal in materials/chem/bio; weather models at AI-parity or better, operational everywhere.
  • Robotics: hundreds of thousands of humanoids in warehouses/factories (still narrow tasks), robotaxis in 30–50+ cities globally; home robots still don’t happen.
  • Geopolitics: a de facto compute-divide solidifies — US-aligned stack vs China-aligned open-weights stack; middle powers (India, Gulf, ASEAN, EU) play both. First serious multilateral compute/AI-risk talks, weak but real.

2030–2031: The Second-Order World

  • “Reasoning as a utility”: for most knowledge tasks, the marginal cost of a competent draft/diagnosis/prototype trends toward zero, like computation did. What remains scarce: taste, judgment, accountability, trust, physical-world skills, and relationships.
  • Cumulative economic effect by 2031: realistic +$3–7T/yr global run-rate (a discounted version of Goldman/McKinsey), heavily concentrated in the US/China with early Signs of the “Great Divergence” between AI-rich and AI-poor economies — the single biggest development-economics story of the decade.
  • Education has been rebuilt: AI tutors standard; assessment shifts from “produce the artifact” to “verify/orchestrate/judge it”; university CS enrollment recovers because AI made building cheaper (AI Index notes CS enrollment actually fell 11% in 2025 — expect the reversal as the tools get better).
  • Medicine: AI triage/diagnostics standard-of-care in much of the world (where regulation allows); documentation-era physicians never return; the randomized-evidence gap either closes or produces a reckoning.
  • The AGI question resolves into semantics: systems in 2031 will be superhuman at many things, human-level-ish at long-horizon agency, sub-human at embodiment and at “owning” real-world accountability. Whether anyone calls that “AGI” will matter less than what the institutions now assume.

5.3 What probably does NOT happen by 2031

Calibrating against hype, in both directions:

  • No mass permanent unemployment (yet) — absorption, not replacement, dominates the 5-year window; but the entry-level ladder is permanently restructured.
  • No one-model-rules-all: convergence + open weights + regulation keep the market plural.
  • No household robot butler; embodiment lags cognition by ~5–10 years.
  • No collapse of the internet into “AI slop only” — provenance tooling, taste, and human networks revalue authentic human work (the “human premium”).
  • And symmetrically: no stall. The single most robust data point in this entire article is that capability curves kept doubling through every hype cycle since 2019. Bets against the curve have lost for seven straight years.

Part 6 — So Why Do People Say “It Will Change Civilization”? (The Synthesis)

Strip away the hype and four hard facts remain:

  1. Generality at collapsing cost. For the first time, a technology sells cognitive labor across all domains at once, with unit costs falling ~100x+ per few years (280x measured for GPT-3.5-class quality). Every general-purpose technology with that profile (steam, electricity, computing) reorganized civilization within ~40 years. This one is spreading 10x faster than the internet.
  2. The curve hasn’t bent. Seven consecutive years of exponential benchmark gains, 3.3x/year compute, task-horizons doubling every ~7 months (accelerating), 30-point single-year benchmark jumps. Even the skeptics’ models predict transformation — just later.
  3. Capital has voted. $1T/year of investment by 2026, $31T cumulative projected by 2050, and national governments treating compute like oil. Whatever one thinks of bubble timing, societies do not spend trillions on toys.
  4. It’s already leaving fingerprints. A 13–20% drop in young-worker employment in exposed occupations, +26% software productivity, 83% less physician paperwork, 258 FDA-authorized AI devices in one year, AI co-authoring science, 53% of humanity using it within three years. These are not forecasts; they are line items.

The honest 2026 position is neither doom nor hype: AI is a real general-purpose revolution arriving at internet speed while our institutions (school, law, work, safety nets) move at committee speed. The gap between those two speeds is where all the drama of the next five years lives — in hiring, elections, science, war, and daily life.

The most probable 2031, in one sentence: you’ll have agent-colleagues you supervise, medicine and science will visibly accelerate, “junior” knowledge work as we knew it in 2023 will be gone, intelligence will be as cheap and ambient as electricity, the US and China will run parallel AI stacks, there will have been at least one nasty financial correction along the way — and human judgment, trust, taste, and accountability will have become the scarcest economic resources on Earth.

Sources (primary, as cited above)

  1. Stanford HAI — AI Index Report 2026 (Apr 2026) + 12-takeaways summary; UNU-C3 summary of all 9 chapters (Jul 2026).
  2. METR — Measuring AI Ability to Complete Long Tasks (Mar 2025; Time Horizon 1.1 updates through 2026).
  3. Goldman Sachs Research — Global AI investment to exceed $1T in 2026 (Aug 2026); Generative AI could raise global GDP by 7%; US data center power demand to double by 2027 (May 2026); How will AI affect the US labor market (Mar 2026).
  4. McKinsey — The economic potential of generative AI (2023, $2.6–4.4T/yr); PwC (14% GDP by 2030; $31T buildout by 2050 via Barron’s 2026).
  5. IEA — Energy and AI (700–1,700 TWh data-centre range by 2035); Deloitte (123 GW by 2035); AI Index energy chapter (29.6 GW AI capacity; Grok 4 ~72,000 tCO₂e).
  6. Dallas Fed (Jan 2026) — 22–25 y/o employment in AI-exposed occupations −13% since 2022; AI Index — 22–25 y/o software devs −20% since 2024; Yale Budget Lab (Sep 2026) counter-evidence; WEF — AI and the Future of Entry-Level Work (2026).
  7. Davos/WEF Jan 2026 — Hassabis (50% AGI within decade) & Amodei joint session (Fortune, Jan 23 2026); Hassabis “AGI 3–4 years away” (Sherwood News, May 27 2026); Amodei — Machines of Loving Grace (Oct 2024); Altman AGI-within-term statements (2025–26).
  8. AI 2027 (Kokotajlo, Alexander et al., Apr 2025) + authors’ Jan 2026 update (“somewhat slower than the scenario”); MIRI critique (Apr 2025); Gary Marcus critique; 80,000 Hours — Kokotajlo “AI 2040 Plan A” (Aug 27, 2026).
  9. Bloomberg — A Guide to the Circular Deals Underpinning the AI Boom (Jan 2026); LA Times — Nvidia $750B deal round (Jul 29, 2026); MIT Nolan study (95% zero-return pilots, 2025); Wikipedia — AI bubble survey; Cursor $2B ARR (Marktechpost, May 2026).
  10. USCC — Two Loops: China’s Open AI Strategy (Mar 2026); CSIS — What to Know About Chinese AI Models (Jul 2026); Brookings — Competing AI strategies for the US and China (Apr 2026); Bruegel (Sep 2026).
  11. DeepMind — AlphaFold: Five Years of Impact (Nov 25, 2025); Nature — AI-in-science data; AI Index Science & Medicine chapters (AION-1, FDA 258 devices, 85.5% vs 20% diagnosis study).
  12. Humanoid/robotics deployment figures: Technology.org (Jul 2026), Tesla robotaxi status reports (2026); Waymo public ops data.

Figures are as reported by the sources above; investment and benchmark numbers move monthly — treat this as a snapshot of September 2026, not a prophecy. Probabilities in Part 5 are subjective judgments informed by the base rates cited, not measurements.