Medium- and Long-Term Scenarios

A scenario, as used in this encyclopedia, is an internally consistent narrative or model of how AI development might unfold over three to more than ten years, built to explore what would have to be true and what would follow. A scenario is not a prediction and carries no probability unless a cited source supplies one; the forecasts that accompany it are separate records with their own operational definitions.

Beyond about three years, the evidence available does not support single-point predictions of AI capability, economic effect or governance. Scenario analysis is the standard response in other long-horizon fields: it holds several different futures in view, identifies the variables that separate them, and states which observations would show that one is unfolding. This article sets out four scenario families, the drivers that distinguish them, and what each would imply for the indicators in FCST-SIGNALS-001. The families are the encyclopedia's own organising device and are not taken from any single source.

Scenario methods developed in military and corporate planning in the twentieth century; Bishop, Hines, and Collins (2007) survey the main techniques and distinguish, among other things, between exploratory and normative scenarios and between qualitative and quantitative construction (Bishop et al., 2007).

In AI, narrative scenarios and quantitative models have appeared side by side. Model-based work includes compute-centric frameworks that treat capability as a function of training compute and algorithmic progress and that include feedback from AI-assisted research (Davidson, 2023; Cotra, 2020). Narrative work includes the AI Futures Project's AI 2027, written by Kokotajlo, Lifland, Larsen, and Dean and published in April 2025, which its authors present as a scenario and not as a firm prediction; they describe it as close to their mode and not their median, and report that co-authors' AGI medians ranged from 2028 to 2032 (Kokotajlo et al., 2025). [VERIFICATION NEEDED: the scenario's content was not reviewed beyond the project's "about" page; the exact publication date and the quoted range are from that page.]

#4.1 What distinguishes a scenario from a forecast

A forecast assigns a probability $P(E)$ to a resolvable event $E$. A scenario $S$ is a conjunction of conditions $C_1\wedge C_2\wedge\dots\wedge C_k$ with a narrative that links them. By the conjunction rule, $P(S)\le \min_i P(C_i)$, and typically much smaller; detailed narratives are therefore less probable than their components, even though they feel more plausible. This is why scenarios are used to explore and why they should not be read as "the most likely future".

#4.2 Scenario axes

Choosing two or three axes of high impact and high uncertainty gives a structured set of scenarios. For AI, candidate axes drawn from the uncertainties discussed in FCST-UNCERT-001 include:

• Capability growth rate: whether measured capability trends continue, slow, or accelerate.
• Automation of AI research itself: whether AI systems materially speed AI development (the mechanism at issue in ASI-EXP-001, a hypothesis and not an established process).
• Resource and coordination constraints: whether power, chips, data, capital, or governance limit scaling.

#4.3 A compute-centric toy structure

In compute-centric models, effective compute $C_{\mathrm{eff}}(t)$ grows through investment and algorithmic improvement, and capability is a function $K(C_{\mathrm{eff}})$. If AI systems speed research, algorithmic progress $a(t)$ becomes endogenous: $\dot a = f(K)$. Whether $f$ is strong enough to produce accelerating growth, or is offset by diminishing returns to research, is an empirical question that the models treat as a parameter (Davidson, 2023). This encyclopedia does not assert a value for it.

The four families below differ chiefly on the axes in 4.2. They are scenarios, not predictions, and no probabilities are assigned.

#5.1 Scenario A: Plateau or slowdown

Conditions. Returns to scaling diminish on the capabilities that matter, or binding constraints (power, chips, data, capital) cap training scale; architectural alternatives do not close the gap. Epoch AI's 2024 analysis identifies power as the constraint most likely to bind first and projects that runs of about $2\times10^{29}$ FLOP are likely feasible by 2030, which makes constraint arrival around then a live possibility but not a result (Sevilla et al., 2024). Observable markers. Slowing growth in the largest training runs; flat measured time horizons; falling investment growth. Implications. Economic adoption continues at the pace of integration, not capability; timelines in survey aggregates would move later.

#5.2 Scenario B: Continued steady progress and gradual automation

Conditions. Existing trends continue with no discontinuity: compute and task-length measures continue to grow roughly at the fitted rates (Sevilla et al., 2022; Kwa et al., 2025), AI systems increasingly automate parts of occupations, and human oversight and integration costs set the pace. Observable markers. Steady, smooth growth in measured capability; productivity effects appearing gradually and unevenly. Implications. Policy has time to adapt incrementally; a fully general system is not reached within a decade, or is reached late.

#5.3 Scenario C: Rapid acceleration through AI-assisted AI research

Conditions. AI systems become able to perform a large share of AI research, and returns to research are high enough that capability growth speeds up (the case explored in compute-centric takeoff models). In Davidson's (2023) draft framework, as summarised on the Effective Altruism Forum, the transition from AI that could automate 20% of cognitive tasks to 100% had a median of roughly 3 years, and the author's best guess for the interval from AGI to superintelligence was under one year absent deliberate slowing; the report is conditional on its assumptions and its authors flag substantial uncertainty (Davidson, 2023). [VERIFICATION NEEDED: figures from the forum post on the January 2023 draft, not from the final report.] Observable markers. Evidence that AI systems account for a large share of the progress in AI development; sharp acceleration in measured trends beyond fitted rates. Implications. Compressed decision time; heightened relevance of safety evaluation and oversight.

#5.4 Scenario D: Fragmentation and constrained deployment

Conditions. Capability growth is substantial, but regulation, export controls, security incidents, or geopolitical competition fragment or limit development and deployment. Observable markers. Divergent regimes across jurisdictions; compute governance measures; deployment restrictions in sensitive domains. Implications. Uneven diffusion; capability concentrated in a few actors or jurisdictions.

The families are not mutually exclusive over time: a plateau followed by an architectural advance, or acceleration followed by imposed constraints, are plausible compositions.

• Stress-testing plans. An organisation can ask whether a strategy is acceptable under A, B, C and D rather than only under its preferred case.
• Disagreement mapping. The Existential Risk Persuasion Tournament found domain experts' median probability of AI-caused extinction by 2100 at 3% against 0.38% for superforecasters, among 80 experts and 89 superforecasters (Karger et al., 2023; see the verification note in FCST-METHOD-001). Scenario axes (such as the strength of AI-driven research acceleration) can locate where such groups differ without resolving who is right.
• Survey context. Aggregate survey dates give a 10% chance by 2027 and 50% by 2047 for unaided machines outperforming humans at every task, and 10% by 2037 and 50% by 2116 for full automation of occupations (Grace et al., 2024). The gap between milestones is itself evidence that respondents expect a long period between capability and complete automation, which is closer to scenario B than C, though the survey does not label scenarios.

• Narrative bias. Vivid narratives attract attention disproportionate to their likelihood (see 4.1).
• Omitted scenarios. Four families cannot span the space; surprises outside the set are likely.
• Hidden probability claims. Ordering scenarios by presentation implies probabilities that the evidence does not give.
• Unresolved: the sign and size of research-feedback effects; how far capability trends on narrow measures generalise; how institutions respond.

As of 2026-10-15 there is no empirical result that selects among A to D. Evidence for continued growth on specific measures exists (Sevilla et al., 2022; Kwa et al., 2025), as do published analyses of possible scaling constraints (Sevilla et al., 2024), but extrapolation across the window of this article is contested. Model-based takeoff results depend on parameters that remain uncertain (Davidson, 2023). Authors of the most visible narrative scenario say it is not their median (Kokotajlo et al., 2025).

Scenario families should be revised when indicators in FCST-SIGNALS-001 move. The near-term records in FCST-NEAR-001 provide the first checks.

#Forecast record

Note on the record: the survey's 10% figure for 2037 is an aggregate across respondents and is not directly a probability for 2030; the 97% figure is an editorial judgment.

• Foundational Concepts:
  • FCST-METHOD-001 Forecasting AI Progress: Methods and Calibration: Why scenarios and forecasts are kept apart.
  • AGI-TIME-001 AGI Forecasting & Timeline Methodology: Timeline methods that feed scenario axes.
• Related Methods & Technologies:
  • ECON-COMP-001 The Economics of Compute & Scaling: Cost and supply constraints behind Scenario A.
  • FCST-NEAR-001 Near-Term Outlook (0-3 years): Dated records that test early divergence.
• Advanced Implications & Future Trajectories:
  • ASI-EXP-001 The Intelligence Explosion Hypothesis: The hypothesis underlying Scenario C.
  • FCST-SIGNALS-001 Signals, Indicators and Inflection Points: Observables that discriminate scenarios.
  • FCST-UNCERT-001 The Uncertainties That Matter Most: The axes' underlying uncertainties.
  • FCST-REG-001 Forecast Register: Record status.