The Emergence of Artificial General Intelligence

on

 

The Emergence of Artificial General Intelligence

Timeline Forecasts and Economic Implications

A scenario-based analysis grounded in forecasting data, task-based growth models, and early labor-market evidence — October 2026.

1. Abstract

Forecasts of when artificial general intelligence (AGI) will arrive have compressed dramatically since 2020, but remain widely dispersed, largely because "AGI" denotes at least three distinct milestones: a weakly general cognitive system, a high-level machine intelligence (HLMI) that outperforms humans on most tasks, and a transformative or superintelligent system whose economic footprint qualitatively exceeds both. As of late 2026, frontier-lab leaders cluster on 2026–2030, the Metaculus community median for a stringent multi-condition AGI definition sits around 2033, and the 2023 ESPAI survey of 2,778 researchers placed a 50% probability of HLMI at 2047. This article disaggregates these forecasts by definition and evidentiary base, then maps their economic consequences onto a three-scenario framework — Modest, Substantial, and Transformative — spanning GDP and total factor productivity, labor markets, inflation, robotics and physical capital, biomedical research, and the functional income distribution. Task-based macro models (Acemoglu–Restrepo) imply TFP gains of roughly 0.5–1.5% over a decade absent AGI; endogenous growth models with automated R&D (Davidson; Aghion et al.) permit growth acceleration an order of magnitude larger if capability gains translate into automation of innovation itself. Early labor-market evidence — an approximately 19% relative employment shortfall for young workers in AI-exposed occupations — suggests the diffusion phase is already underway at the entry-level margin, even as aggregate data show no economy-wide displacement. We identify leading indicators, second-order effects, and policy levers, and close with open research questions.

2. Introduction and Operational Definitions

The economics of AGI cannot be analyzed without first fixing what "AGI" means, because nearly all headline disagreement about timelines is definitional rather than evidentiary. Three operational thresholds recur across the technical, forecasting, and economics literatures, and this article uses them throughout.

Weak AGI denotes a system with broad cognitive competence — roughly matching a skilled human across most digital, remotely performable knowledge tasks — without necessarily exhibiting long-horizon autonomy, novel scientific reasoning, or physical embodiment. Metaculus operationalizes a closely related concept in its "weakly general AI" question, a cognitive-benchmark basket with no robotics requirement; as of mid-2026 its community median sat near June 2028. A system of this kind could plausibly perform a large share of white-collar task content while still failing at sustained, self-directed work.

High-level machine intelligence (HLMI), the definition used in the AI Impacts / ESPAI survey tradition, is reached "when unaided machines can accomplish every task better and more cheaply than human workers, ignoring the cost of the machine." The 2023 survey of 2,778 published AI researchers put a 50% probability on HLMI by 2047, with individual forecasts extremely dispersed (Grace et al., 2024). This is a demanding definition: it is a universal-competence claim, not a "most useful tasks" claim.

Transformative AI / AI superintelligence (TAI or ASI) refers to a system whose capabilities exceed the best humans in essentially all cognitive domains and whose deployment would, under most growth models, raise the long-run growth rate itself — for example, by automating a substantial fraction of research and development. Amodei's "powerful AI" (a system "better than humans at almost all cognitive tasks," plausibly Nobel-laureate-level within specific domains) and Hassabis's "Einstein Test" — the ability to generate genuinely novel explanatory structure, e.g., independently derive general relativity from pre-1911 physics — both sit at or above this tier, and both have recently been attached to windows of roughly two to ten years by their authors.

Three further distinctions structure everything below. First, capability ≠ deployment: a system can exist in a lab years before it is embedded in production processes, and the J-curve of intangible investment (Brynjolfsson) implies measured productivity may lag capability by years. Second, cognitive ≠ physical: robotics lags digital intelligence substantially (Moravec's paradox), so even a "weak AGI" would automate remote work long before it automates plumbing. Third, point forecasts are meaningless without distributions: the honest summary of the 2026 state of the art is a wide probability distribution whose location has moved left, not a date.

3. Timeline Analysis

3.1 The current forecast landscape

Frontier-lab leaders (as of 2026). The CEOs of the major labs cluster on 2026–2030, with important definitional variance. Sam Altman stated in August 2026 that OpenAI did not yet have AGI but expected an internal system he would personally call AGI by the end of 2026, with chief research officer Mark Chen describing the company as "80% of the way" there. Dario Amodei has repeatedly pointed to 2026–2027 for "powerful AI," defined as better than almost all humans at almost all cognitive tasks. Elon Musk's 2025 date expired and was restated as end-2026. Demis Hassabis is the most conservative, at roughly two to five years out (≈2029–2031) — and his definition, requiring original hypothesis generation, is the hardest of the set. Mustafa Suleyman (Microsoft AI) forecast human-level performance on most professional tasks within 12–18 months as of early 2026. These forecasts carry obvious selection bias — leaders raise capital on short timelines — but they also rest on visibility into unreleased systems, and they have been directionally right more often than academic forecasts since 2020.

Aggregative forecasting platforms and prediction markets. The Metaculus community median for its stringent "general AI" question (four conditions: an adversarial two-hour Turing test, ≥90% on a broad knowledge benchmark, ≥90% top-1 accuracy on interview-level coding, and general robotic capability including autonomously assembling a 1:8 scale model car) stood at roughly January–April 2033 as of mid-to-late 2026, with ~25% probability by 2029 — down from a median of ~50 years out in 2020, though it ticked back two years in 2025. The robotics-free "weakly general" question sat near June 2028. Real-money markets bracket these: Kalshi priced OpenAI-defined AGI by 2030 near 55%, Manifold's demonstration-by-2030 market near 50%, and Polymarket only ~9% for OpenAI AGI by 2027. Aggregating across platforms, a reasonable composite for "some credible AGI claim validated by 2030" is a coin flip, with an 80% interval spanning roughly 2027–2044.

Large-scale researcher surveys. The 2023 ESPAI / AI Impacts survey (2,778 researchers, published 2024) yielded a 2047 median for HLMI at 50% probability — but a 10%-probability date of 2027, and the same population's forecasts have historically been unstable in both directions. Earlier AI Impacts surveys (2016, 2022) gave later medians (2060 for some framings), so the 2023 numbers already reflect a large downward revision; a further revision is likely in the next survey wave.

Independent forecasting groups and superforecasters. The 2022 XPT survey of 33 superforecasters gave a 25% probability of Metaculus-defined AGI by ~2048 — far later — but predates the reasoning-model era, and those forecasters made compute-growth predictions that were subsequently falsified. Samotsvety Forecasting gave a ~2029 median. Tom Davidson's Takeoff Speeds model (2023), a semi-endogenous growth framework driven by training-compute growth, algorithmic progress, and AI-automated R&D, originally implied full economic automation around 2040; updated with Epoch AI's evidence that effective compute (hardware × software efficiency × investment) has been growing ~12×/year rather than the model's assumed 2.5×/year, its central scenarios move to roughly 2027–2030. Epoch's GATE model, a successor framework emphasizing broad economic automation over R&D automation, lands in a similar window with somewhat less aggressive takeoff dynamics.

Source (date)

Definition

25% or 10% date

Median / central date

Key weakness

OpenAI (Altman, Aug 2026)

Founder's personal AGI bar

—

End-2026 (internal)

Unverifiable internal definition

Anthropic (Amodei)

"Powerful AI": better than nearly all humans at nearly all cognitive tasks

—

2026–2027

Incentives; demanding-but-vague bar

Google DeepMind (Hassabis, Oct 2026)

"Einstein Test": original generative reasoning

—

~2–3 years (≈2029)

Hardest definition in the set

Metaculus "weakly general AI" (2026)

Cognitive benchmark basket, no robotics

—

June 2028

Weak benchmark may overstate real-world ability

Metaculus "general AI" (2026)

4-part definition incl. robotic assembly

~25% by 2029

Jan–Apr 2033

Robotics clause dominates; possibly both too strict and too lax

Kalshi / Manifold (2026)

Market-specific AGI criteria

—

~50–55% by 2030

Thin liquidity, resolution ambiguity

ESPAI survey (2023)

HLMI: all tasks better/cheaper than humans

10% by 2027

50% by 2047

Pre-reasoning-model vintage; historically unstable

XPT superforecasters (2022)

Metaculus definition

25% by ~2048

Later

Pre-o1 vintage; falsified compute assumptions

Davidson Takeoff Speeds, updated (2025–26)

Full economic automation

—

~2027–2030

Extrapolative; sensitive to effective-compute trend

3.2 Sources of variance

The ~20-year spread between the earliest and latest credible forecasts is traceable to at least five distinct drivers:

  1. Definition selection. The Metaculus general-AI and weakly-general-AI questions — same community, same month — differ by roughly five years purely because of the robotics clause. Lab-CEO convergence on "2026–2030" is largely apparent: Altman's bar, Amodei's Nobel-level bar, and Hassabis's Einstein Test are different destinations colliding on the same calendar.
  2. Capability metrics and their extrapolability. METR's task-completion time-horizon benchmark — the length of human-work-time task an agent completes with 50% reliability — doubled roughly every 7 months from 2019–2024 and every ~4 months (≈89–131 days on recent fits) since 2024, with the frontier near 12 hours (50% horizon) by May 2026, though METR cautions that measurements above ~16 hours exceed its current task suite's validity. Straight-line extrapolation yields week-long autonomous tasks by late 2026–2027 and month-long tasks by mid-2027; skeptics correctly note that self-contained benchmark tasks are not month-long collaborative real-world projects, that METR found roughly half of SWE-bench-passing pull requests would not actually be merged by maintainers, and that 80%-reliability horizons remain far shorter (3–4 hours in early 2026).
  3. Compute and capital trends. Frontier training compute has grown roughly 4–5×/year, and total effective compute — hardware, algorithmic efficiency (~3×/year), and rising investment share — has grown faster. Whether this survives power, chip-supply, and financing constraints (datacenter capex already rivals utility-scale investment) is the largest physical unknown.
  4. Algorithmic progress and R&D automation. If AI systems automate 1–10% of AI R&D itself, the Davidson-style feedback loop shortens everything downstream. This is the single most consequential and least empirically constrained parameter in all formal models.
  5. Reference-class and incentive effects. Lab leaders see unreleased systems but sell timelines; academic surveys are stale by construction; superforecasters are calibrated on questions with feedback, not on once-per-civilization events. No class of forecaster dominates on epistemic grounds; the honest move is to weight a portfolio.

A reasonable synthesis as of October 2026: weak AGI (broad cognitive competence on digital tasks) plausibly within 1–4 years; HLMI with substantial probability mass in 2028–2040, median plausibly in the early-to-mid 2030s; transformative AI / ASI a genuinely open question in which formal models disagree by decades, hinging on whether the METR-style trend continues and whether R&D automation feeds back.

4. Economic Impact Analysis: A Three-Scenario Framework

We formalize three scenarios, paralleling (and partly quantified by) Anthropic's September 2026 Economic Scenario Explorer for US GDP by 2030. Crucially, scenarios should be read as conditional on diffusion, not just capability.

  • Modest: AI diffuses like a normal general-purpose technology (internet-class). Capabilities plateau or improve slowly; adoption is bottlenecked by integration costs, data, trust, and regulation.
  • Substantial: AI autonomously performs a large fraction (roughly half) of knowledge-work tasks by 2030; adoption is broad but not universal; a J-curve precedes measured gains.
  • Transformative: AI exceeds humans on nearly all cognitive tasks with high autonomy, automating a substantial share of R&D and physical-coordination work; growth rates rise, potentially by an order of magnitude in endogenous-growth models.

4.1 GDP and total factor productivity

The task-based framework of Acemoglu and Restrepo decomposes automation's effect into a displacement effect (automating task content) and a productivity effect (new-task creation), and it implies that TFP gains are bounded by the share of task content profitably automatable, not by what is technically possible. Acemoglu's 2024 estimate puts AI's decade contribution to US TFP at only ~0.53–0.66%, under the assumption that only ~5% of tasks will be profitably automated in the near term, adding ~1–1.5% to GDP over a decade. This is the Modest scenario's quantitative core and a deliberate lower bound: it assumes hard ceilings on exposed tasks, cost savings, and no significant new-task creation.

Consultancy and market estimates (Goldman Sachs ~1.5% annual GDP uplift; McKinsey ~$2.2–4.4 trillion annually for generative-AI use cases) sit between Modest and Substantial, with more optimistic assumptions about profitable automatability. Anthropic's scenario modeling quantifies the Substantial and Transformative cases directly: by 2030, US GDP of $34.1 trillion in the Modest case (+1.6% vs. a no-AI baseline), $36.3 trillion in the Substantial case (+8.3%), and $44.4 trillion in the Transformative/"extreme" case (+32.4%, from a current base of just over $30 trillion) — figures Jack Clark presented to the IMF–World Bank annual meetings in October 2026, with the explicit caveat that such growth would require new industries and inventions to materialize.

The Transformative scenario is where endogenous growth theory takes over. Aghion, Jones, and coauthors show that whether AI raises growth permanently depends on whether automation complements or substitutes for the production of ideas: if AI automates idea production itself, growth can accelerate sharply — the singularity case — whereas bottleneck or diminishing-returns cases bound the effect. Davidson's framework, with its software-only-singularity branch (~65% probability conditional on full R&D automation, implying 2–3 orders of magnitude of effective-compute gains), is the most explicit formalization. Epoch's GATE variant argues, contrarily, that most AI value will come from broad economic automation rather than R&D automation, since capital deepening rather than ideas accounts for the majority of historical labor-productivity growth. Input-output and occupational-exposure approaches (Felten et al.; Eloundou et al.'s "GPTs are GPTs," which finds ~80% of the US workforce has ≥10% of tasks exposed to LLMs and ~19% has ≥50% exposed) pin down the exposure channel but are silent on realization.

Synthesis: Decade TFP impact plausibly ranges from <1% (Modest, Acemoglu-style bounds) to a regime change with double-digit annual growth (Transformative, AGI-automated R&D). The distribution is fat-tailed in both directions; the central case for 2026–2035 is probably Substantial — materially above the electricity-era benchmark for a decade, but below the scenarios that dominate public discourse.

4.2 Labor markets

The task-based literature predicts that displacement operates through task content, not occupations, and that the margin of adjustment is hiring, not firing. That prediction is now visible in the data — contested but increasingly corroborated. Stanford's Digital Economy Lab "Canaries" study (August 2026 update, ADP payroll records through June 2026) finds employment of 22–25-year-olds in the most AI-exposed occupations roughly 19% below the path of less-exposed peers (up from ~13–15% a year earlier), with the adjustment running through vacancies never posted rather than separations, and clustering where AI substitutes for codified task content; software developers aged 22–25 have seen employment fall ~20% since 2022. The authors are explicit that this is descriptive, not causal, and that the gap narrows with education controls. Counter-evidence exists and matters: a CESifo working paper (Fairlie & Wu) finds recent-graduate unemployment (7.3% in summer 2026) within the normal 2022–2024 range, with no statistically significant AI-exposure gradient in Current Population Survey data; the New York Fed reports recent-graduate underemployment around 42%; and analysts attribute part of the youth squeeze to interest-rate normalization and remote-work wage structure, not AI alone.

In scenario terms: Modest implies limited aggregate employment effects but persistent reallocation. Substantial implies a decade in which entry-level cognitive work contracts as a category — Anthropic's own scenario model predicts displaced knowledge workers migrating toward physical and interpersonal roles ("electrician and nurse"), with average wages rising but gains concentrated outside knowledge work, and churn increasing within historically normal unemployment ranges. Transformative implies the labor share of income falls substantially as AI becomes a (quasi-)capital input, with employment elasticities that historical data — drawn from worlds where human labor was indispensable — cannot credibly bound. Acemoglu's warning that a generation of graduates unable to find first jobs has historically been politically destabilizing deserves weight as a second-order risk, not just an equity footnote.

4.3 Inflation and relative prices

AI is, in the first instance, a supply-side shock to the price of cognitive output: where a task's marginal cost collapses, goods and services embodying that task should deflate relative to stagnant-productivity sectors (housing, in-person services, regulated professions). Early usage and job-posting evidence is consistent with relative-price effects concentrated in software, translation, content, and routine analysis. Aggregate implications are scenario-dependent: Modest is mildly disinflationary with an ambiguous net effect; Substantial is disinflationary in AI-intensive sectors but inflationary in the bottleneck sectors competing for the same energy, chips, and skilled labor — notably electricity and construction, where datacenter demand is already bidding up power prices; Transformative breaks standard Phillips-curve logic entirely, because output supply becomes elastic while demand-side distribution determines whether the regime is deflationary abundance or asset-price inflation with constrained physical throughput. A key monetary-policy point: central banks targeting headline inflation will observe a falling weight on AI-exposed components, mechanically flattering measured inflation while real purchasing-power effects diverge across households.

4.4 Manufacturing, robotics, and physical capital accumulation

The cognitive–physical gap is the most underweighted fact in AGI economics. Current systems remain far from general dexterity; the Metaculus robotics clause is the single largest reason its general-AI median trails the CEO consensus by half a decade. But robotics' production function is changing: vision-language-action models, imitation learning, and sim-to-real transfer are collapsing the historical cost of programming robot behavior, and warehouse automation and humanoid-robot pilot programs scaled sharply through 2025–2026. If "weak AGI" arrives first in software, the physical-economy consequences arrive with a lag governed by capital accumulation: building the robot fleet, retrofitting facilities, and expanding electricity supply take years and real savings. This implies a distinctive Transformative sub-case — explosive design capability with a physically bottlenecked deployment — in which measured GDP accelerates only as fast as the capital stock can absorb it. It also implies a cross-country asymmetry: economies with fast physical-capital formation (notably China and East Asia) will convert cognitive AI into manufacturing productivity faster than service-dominated advanced economies.

4.5 Healthcare, biomedical discovery, and scientific productivity

If the Transformative scenario has a first-order application, it is compressing the discovery-to-deployment pipeline in biology and medicine. AlphaFold-class systems have already collapsed parts of structural biology; AI-designed candidates are in clinical trials; and lab-leader claims of compressed drug-development timelines (Amodei has spoken of 50–100% compression for some modalities) are directionally supported by early trial-phase data, though no AI-originated drug has yet completed Phase III, so the evidence base remains indirect. The macro significance runs through the R&D-automation channel: health-sector productivity has been Baumol's-disease-ridden for decades, and automated hypothesis generation, literature synthesis, and experiment design attack precisely the tasks that made it stagnant. In scenario terms, Modest and Substantial yield incremental gains — faster trials and cost reductions in administrative work, roughly a quarter of US health spending; Transformative implies a level shift in the rate of scientific output, with growth-model consequences if knowledge production is the engine of growth (Aghion et al.). It is plausibly the largest welfare effect of AGI, and the one least captured by GDP because it arrives as health and longevity.

4.6 Capital–labor distribution and inequality

Automation raises the capital share when AI systems substitute for labor and are owned as capital. In the Modest case, distributional effects resemble previous GPT waves: within-skill wage dispersion widens, superstar firms capture rents, and the labor share dips marginally. In the Substantial case, the Anthropic scenario model's pattern — knowledge-worker wages stagnate while non-knowledge wages rise — is a within-labor redistribution that offsets, but does not reverse, a rising capital share. In the Transformative case, ownership of the AGI-producing capital stock (compute, models, energy) becomes the dominant distributional variable: if frontier-AI rents accrue to a handful of firms and their investors, functional income inequality could reach levels without historical precedent, making the choice among policy instruments (§6) the central political-economy question of the scenario. The developed/emerging-economy split follows: emerging economies face a threat to the services-export development ladder (call centers, back-office processing, increasingly software are precisely the exposed task content), while potentially benefiting from cheap AI-augmented education, healthcare, and leapfrogging — a race between the loss of the export ladder and the arrival of AI-augmented public goods. The former is nearer-term and measurable; the latter is speculative.

4.7 Scenario comparison

Dimension

Modest

Substantial

Transformative

US GDP effect by 2030 (Anthropic scenarios)

+1.6% ($34.1T)

+8.3% ($36.3T)

+32.4% ($44.4T)

Decade TFP gain (task-based models)

~0.5–1% (Acemoglu)

~1–3%

Regime change; growth rate itself rises (Davidson, Aghion)

Labor market

Reallocation; flat aggregate effects

Entry-level cognitive hiring contracts ~15–20% relative; churn rises

Labor share falls sharply; employment elasticity historically unbounded

Inflation

Mildly disinflationary

Relative-price deflation in AI sectors; energy/chip bottlenecks inflate

Phillips-curve logic breaks; abundance vs. asset inflation depends on distribution

Physical economy

Little change

Warehouse/manufacturing automation expands; power demand rises

Robot fleet accumulation becomes the binding constraint

Science/health

Administrative cost savings

Faster trials; AI candidates in clinic

Discovery rate level-shift; compressed timelines

Distribution

Superstar-firm rents

Knowledge wages stagnate; non-knowledge wages rise

Capital share dominant; ownership is the key variable

Plausibility (author's judgment, Oct 2026)

~35–45%

~40–50%

~10–20% by 2035

5. Leading Indicators and Early Detection

Because AGI's economic significance will be visible in diffusion data before any declaration of "AGI," a monitored indicator set should span capability, adoption, and macro layers.

Capability layer. (i) METR 50% and 80% time horizons — watch whether the ~4-month doubling persists and whether the 80% horizon (currently lagging badly, at 3–4 hours) closes the gap with the 50% horizon, which would indicate reliability scaling alongside task length; (ii) performance on hard-reasoning and novelty benchmarks (FrontierMath, ARC-AGI, Hassabis's proposed Einstein-class tests); (iii) measurable AI contribution to AI R&D itself — the Davidson model's pivotal parameter; (iv) robotics benchmarks, particularly general manipulation and long-horizon physical autonomy.

Adoption/diffusion layer. (i) Firm-level task automation shares (Census Bureau surveys show a sharp rise in firms replacing multiple employee tasks with AI); (ii) AI-skill postings (~2.5% of US postings, up ~55% year-on-year per the Stanford AI Index 2026); (iii) usage telemetry mapped to occupations (Anthropic Economic Index-style); (iv) enterprise token/inference-volume growth per employee; (v) entry-level vacancy postings in exposed occupations — the Canaries line, currently the most sensitive single macro-adjacent indicator.

Macro layer. (i) Sectoral TFP, especially in software, professional services, and — critically — sectors where forecasters expect no gain (a TFP uptick in construction or healthcare would falsify the bottleneck hypothesis); (ii) the relative price of AI-exposed vs. non-exposed output; (iii) the labor share of income; (iv) investment composition — datacenter capex as a share of GDP is a leading indicator of both capability growth and a potential overbuild crash; (v) electricity demand as the binding physical constraint; (vi) youth and graduate unemployment/underemployment, with the caveat that identifying AI causality requires exposure differentials, not levels.

Falsifiable tripwires are more useful than vibes: if by 2028 METR-style horizons have not crossed multi-day autonomy at usable reliability, the Transformative scenario's probability should be revised down sharply; if by then measured TFP in exposed sectors has not inflected upward despite broad adoption, the Modest case strengthens; if both occur, the Substantial-to-Transformative boundary becomes the live question.

6. Major Uncertainties, Second-Order Effects, and Policy Levers

Uncertainties. Five dominate: (1) whether the capability trend is a durable exponential or a sigmoid — the 2024–26 acceleration could reflect a one-time transition to agentic scaffolding; (2) whether the effective-compute trend survives energy, chip, and financing constraints; (3) the R&D-automation feedback strength — the difference between the GATE and Davidson worldviews; (4) diffusion frictions — trust, liability, integration, and organizational capital — which the entire historical record of general-purpose technologies says are the binding constraint on measured productivity; and (5) alignment and control, which affects not only tail risk but economic value directly (METR's 2026 findings on reward hacking and illegitimate benchmark successes are economically relevant: an untrustworthy AGI has much lower adoption-implied value).

Second-order effects deserve explicit modeling: geopolitical concentration of compute and AI-export dependence as a strategic lever; a potential AI capex/valuation cycle whose unwinding would be deflationary and politically corrosive even if capability trends continue; succession dynamics in firms that stop hiring juniors — every skipped junior hire is a senior that must later be bought on the open market; tax-base erosion as labor income, the easiest thing to tax, shrinks relative to capital income; and macro-political instability if the graduate labor market squeezes as Acemoglu warns.

Policy levers. (i) Compute and energy policy: datacenter interconnection, grid buildout, and semiconductor supply chains are now first-order macro variables; (ii) distribution: the instrument set — capital-share taxation, sovereign AI assets or citizen funds, expanded EITC or UBI pilots, worker-equity mandates — differs sharply across scenarios, arguing for building institutional plumbing (transfer infrastructure, not commitments) now, consistent with Clark's remark that "policymakers should get ready to spend"; (iii) competition policy: frontier-model market concentration is extreme, and the growth payoff of monopolized AGI accrues largely as rents; (iv) education and transition: apprenticeship and mid-career retraining scale, and entry-level labor-market matching, are the most time-sensitive domestic instruments; (v) international: emerging economies' exposure to the services-export ladder argues for AI-access and infrastructure finance, much as earlier GPT waves argued for electrification finance; (vi) safety standards with economic content: evaluation, incident reporting, and liability rules determine diffusion speed as much as any capability trend.

7. Conclusion and Open Research Questions

The evidence as of late 2026 supports three guarded conclusions. First, the capability trend is real, measured, and fast — METR's doubling series is the most disciplined longitudinal capability dataset available, and its recent pace justifies assigning substantial probability to weak AGI within a few years. Second, the forecast dispersion is mostly definitional and epistemic, not noise: lab leaders, Metaculus, superforecasters, and academic researchers are answering different questions with different information, and the correct response is portfolio weighting, not picking a favorite. Third, the economic outcome distribution is extraordinarily wide — from Acemoglu's ~0.55% decade TFP to scenarios in which growth rates themselves rise — and the binding variable is diffusion, not capability.

Open research questions, in rough priority order:

  1. Measurement. How can national statistics capture AI-produced output — much of it intangible, internal, or quality-improving — before 2030?
  2. Reliability scaling. Does the gap between 50% and 80% time horizons close, stay constant, or widen? This single parameter determines whether agentic AI is deployable infrastructure or an impressive demo.
  3. R&D automation. What is the empirically grounded value of AI's current contribution to AI and scientific R&D, and what feedback elasticity does it imply?
  4. Diffusion elasticity. Task-exposure studies identify what could be automated; almost nothing credibly estimates what will be, at what adoption cost, over what lag.
  5. The physical bottleneck. How quickly can robot fleet capital and energy supply absorb cognitive surplus, and does this differ across economies in ways that restructure comparative advantage?
  6. Distribution under Transformative scenarios. Endogenous-growth models say little about who owns the surplus; the political economy of a falling labor share under AI-accelerated growth is largely unmodeled.
  7. Early-warning identification. Can exposure-differential designs (Canaries-style) be scaled into a real-time official statistical product, separating AI-driven hiring compression from cyclical and rate-driven effects?
  8. Tail governance. What international arrangements — compute monitoring, evaluation regimes, benefit-sharing — remain feasible as capability concentration rises?

The honest stance for economists and policymakers is neither the promotional certitude of the shortest timelines nor the reflexive skepticism of the longest, but disciplined scenario planning against a distribution whose center is moving — and whose tails are where both the promise and the peril live.

Key Sources

  • Grace, K. et al. (2024). ESPAI / 2023 Expert Survey on Progress in AI. AI Impacts (HLMI median 2047; 2,778 respondents).
  • Davidson, T. (2023). What a Compute-Centric Framework Says About Takeoff Speeds. Open Philanthropy; community updates (2025–26) incorporating Epoch AI effective-compute trends; Epoch GATE model (2025).
  • METR. Measuring AI Ability to Complete Long Tasks (2025); Time Horizon 1.1 (Jan 2026); time-horizons data page (May 2026); SWE-bench merge-quality study (Mar 2026).
  • Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER — task-based TFP estimate (~0.53–0.66% over a decade; ~5% of tasks profitably automatable near-term).
  • Anthropic. Economic Scenario Explorer (Sept 2026): Modest $34.1T (+1.6%), Substantial $36.3T (+8.3%), Extreme $44.4T (+32.4%) US GDP by 2030; Clark's Per Jacobsson Lecture, IMF–World Bank Annual Meetings, Bangkok (Oct 2026).
  • Aghion, P., Jones, C., & Jones, B. (2019). Artificial Intelligence and Economic Growth — automation, bottlenecks, and the growth rate.
  • Eloundou, L. et al. (2023). GPTs are GPTs — LLM task-exposure scaling (~19% of workers with ≥50% exposure).
  • Stanford Digital Economy Lab. Canaries in the Coal Mine? (Aug 2026 update) — ~19% relative employment shortfall, ages 22–25, AI-exposed occupations.
  • Fairlie, R. & Wu, J. (2026). The Early Impacts of AI on Employment Among Recent College Graduates. CESifo — no significant displacement through summer 2026 (counter-evidence).
  • Metaculus general-AI (median ~2033) and weakly-general-AI (~mid-2028) questions; Kalshi, Polymarket, Manifold markets (2026); 80,000 Hours expert-forecast review; XPT superforecaster survey (2022).
  • Public statements: Altman (Aug 2026), Amodei (2025–26), Hassabis (Oct 2026, "Einstein Test"), Suleyman (2026), Musk (2025–26).
  • Stanford AI Index 2026 (Lightcast postings data); NY Fed labor-market outcomes for recent graduates; Goldman Sachs and McKinsey generative-AI productivity estimates.

Comments

Post a Comment