Back to Dashboard
Ecosystem Monitoring / 8.2

8.2 Prediction of Future Developments and Impacts

2026 Governance Status: Data-rich forecasting; policy-grade prediction still open

Original Problem in the Paper

Paper motivation/open problem: “Anticipating the trajectory and potential impact of AI systems may allow policymakers to proactively set governance priorities, determine the urgency of addressing specific issues, and allocate resources accordingly”; foresight supports “adaptive and anticipatory” governance. The paper identifies future work on extending empirical measurement of trends beyond training compute and algorithmic progress, assessing prediction accuracy, estimating impacts before deployment, estimating economic impacts, developing technical tools to safely and ethically experiment with or simulate outcomes, and understanding AI-specific hardware developments and compute governability.

July 2026 Update & Trajectory

Forecasting is more data-rich by July 2026: Stanford AI Index 2026, OECD.AI live data, METR time-horizon scaling, IEA Energy & AI, and International AI Safety Report syntheses provide richer policy-relevant trend data and analysis. But policy-grade prediction remains open: the checked sources themselves flag uncertainty from evolving methods, external validity, data transparency, and the gap between measured capabilities and real-world impact. Hardware and compute-governability forecasting also remains open because chip counts alone do not capture changing hardware access, model-use patterns, distributed computation, or algorithmic/software efficiency; the sources reviewed here do not establish a validated mechanism-level forecast for those substitutions.

Deployed / Operationalized

  • Stanford AI Index 2026 tracks technical progress, economic influence, societal impact, investment, adoption, and governance/evaluation/preparedness gaps for policymakers and other audiences.
  • OECD.AI live data tracks AI research, demographics, investments in AI and data, jobs/skills, software development, search trends, news, compute, knowledge flows, models/datasets, and patents; the page flags that the live-data section is being updated and that AI-measurement methods remain evolving.
  • METR operationalizes one forecast-relevant capability dimension: the human-time length of tasks AI agents can complete. Its Time Horizon 1.1 page is the current 2026 measurement, and its March 2025 write-up reports an approximately seven-month doubling time while also flagging external-validity and model-error concerns.
  • Frontier-lab safety frameworks operationalize anticipatory risk monitoring in different ways: OpenAI uses tracked capability categories, threat models, thresholds, evaluations, safeguards, and safeguard-sufficiency decisions; Google DeepMind uses early-warning evaluations, Tracked Capability Levels, Critical Capability Levels, inherent/residual risk assessment, and mitigations; Anthropic uses Responsible Scaling Policy capability thresholds, risk reports, and required safeguards.
  • IEA Energy and AI operationalizes electricity-demand and AI-energy analysis using global and regional modelling and datasets, including projections for AI electricity consumption and analysis of energy supply, emissions, security, innovation, and affordability.

New Tractable Vectors

  • Forecast autonomous-agent capability using task time horizons rather than only benchmark percent-correct scores.
  • Backtest capability forecasts against model release data, benchmark saturation, time-to-complete tasks, and AI Index/OECD trend datasets.
  • Estimate energy and grid impacts of AI data centers using IEA regional modelling and data-center datasets.
  • Compare lab safety-framework thresholds and evaluation triggers against observed capability changes, while preserving each framework’s lab-specific terminology.
  • Scenario-test policy thresholds against uncertainty in algorithmic progress, model-use patterns, and hardware/software efficiency rather than relying only on chip-count projections.

Key Open Questions

  • Predicting economic and labor impacts before deployment when adoption, workflow redesign, and organizational learning may lag model capability.
  • Forecasting tail risks and emergent capabilities from frontier models when public information and external access are limited.
  • Estimating hardware and compute governability as actors adapt through alternative access channels, changing hardware mixes, distributed computation, and algorithmic/software efficiency improvements.
  • Validating simulations of social and political impacts without causing the harms being studied.
  • Forecasting long-horizon agent and tool-use failures when benchmarks underrepresent real-world task duration, scaffolding, and deployment complexity.

Evidence & Primary Sources

  • Reuel, Bucknall et al. frame anticipation of AI trajectory and impact as useful for governance prioritization, urgency setting, and resource allocation; they identify empirical trend measurement, prediction accuracy, pre-deployment and economic impact estimation, safe/ethical simulation, and hardware/compute governability as future work. (arXiv PDF): https://arxiv.org/pdf/2407.14981
  • Stanford AI Index says its mission is to provide unbiased, rigorously vetted, globally sourced data for policymakers and others; its 2026 report page says AI technical capabilities, investment, adoption, and economic integration are accelerating while governance, evaluation, and understanding frameworks lag and data transparency is declining. (2026 AI Index Report page): https://hai.stanford.edu/ai-index
  • OECD.AI live data provides timely indicators across AI research, demographics, investments in AI and data, AI jobs and skills, AI software development, search trends, AI news, AI compute, knowledge flows, AI models and datasets, and AI patents; it says the live-data section is being updated and that methods for measuring AI are still evolving. (Copyright 2026): https://oecd.ai/en/data
  • METR proposes measuring AI performance by the human-time length of tasks agents can complete, reports that this length has doubled approximately every seven months over the past six years, identifies Time Horizon 1.1 as the current/up-to-date 2026 measurement, and discusses methodological and external-validity uncertainty. (19 March 2025; page includes 2026 Time Horizon 1.1 update): https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/
  • International AI Safety Report 2026 describes itself as a comprehensive review of current general-purpose AI capabilities, risks/impacts, and mitigation/risk-management approaches for decision-makers and policymakers. (3 February 2026): https://internationalaisafetyreport.org/
  • IEA Energy and AI says it uses global and regional modelling and datasets to fill data gaps and analyze AI electricity consumption, energy supply, emissions, energy security, innovation, and affordability, including projections for AI electricity consumption over the next decade. (Published 10 April 2025): https://www.iea.org/reports/energy-and-ai
  • OpenAI Preparedness Framework v2 describes OpenAI’s approach to tracking and preparing for frontier capabilities, with tracked categories, threat models, measurable thresholds, evaluations before deployment and during development, safeguards, and safeguard-sufficiency decisions. (Last updated 15 April 2025): https://cdn.openai.com/pdf/18a02b5d-6b67-4cec-ab64-68cdfbddebcd/preparedness-framework-v2.pdf
  • Google DeepMind Frontier Safety Framework v3.1 defines Critical Capability Levels and Tracked Capability Levels, uses early-warning evaluations, assesses inherent and residual risk, and addresses CBRN, cyber, harmful manipulation, ML R&D, and misalignment risks with mitigations. (Published 17 April 2026): https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/strengthening-our-frontier-safety-framework/frontier-safety-framework_3-1.pdf
  • Anthropic’s Responsible Scaling Policy page lists RSP version 3.3 as effective 26 May 2026 and describes capability thresholds, risk reports that quantify risk across deployed models, external review authority for Risk Reports, off-cycle updates, and required safeguards. (Last updated 26 May 2026): https://www.anthropic.com/responsible-scaling-policy