The AI 2030 Project

1The Question

The aggregate question is easy to ask and almost impossible to answer honestly: what is artificial intelligence doing to human civilization, taken as a whole? The dishonest answers are familiar. One is the single headline verdict — AI is good, AI is dangerous — which averages away every population for whom the opposite is true. The other is the survey that lists every position and commits to none.

This project takes a third route. Instead of writing conclusions and organizing them afterward, it fixes a single evaluation procedure first and then runs that procedure across the map of civilization, one domain at a time.The research beneath each chapter is a live web sweep conducted at compile time and graded by evidence quality — not a static literature review frozen at publication. The document is what the procedure produces. That inversion — framework first, document second — is the entire method, and everything below is a description of the framework rather than a summary of its conclusions.

2The Framework

The framework's shape is chosen to match the shape of its subject. Reality here is neither a clean hierarchy nor a flat list of independent factors, so the framework is a hierarchical-relational hybrid.

2.1Architecture

It is hierarchical because domains genuinely nest: the economy contains labor markets, which contain gig work. Flatten that and you compare “Healthcare” to “Radiology Second Opinions” as though they were peers. The spine is therefore a domain tree.Roughly twelve to sixteen top-level domains, each recursively decomposable. Depth below the top layer is set by how much evidence exists, not by a fixed schema. But domains are not sealed compartments, so a relational graph is layered on top: every chapter carries an explicit list of cross-domain dependencies, tagged by direction and mechanism. It is deliberately not a formal causal model with estimated weights — forcing quantitative precision onto contested social phenomena would manufacture authority the evidence cannot support.

2.2Five mandatory outputs

Feed the framework a domain and it returns the same five outputs, in the same order, every time. This is what makes sixty-six chapters comparable rather than sixty-six loose opinions, and what lets the document be updated as a machine-readable object.

3Calibrating Uncertainty

Most technology forecasting fails by asserting everything in the same confident register. The framework's authority comes from the opposite discipline: it separates what is known from how well it is known, and marks that separation on the page.

3.1Every claim is graded

Each empirical claim carries an inline evidence tag from a six-tier scale, adapted from the GRADE system in clinical medicine and the IPCC's convention of separating evidence from confidence.High agreement among experts on thin evidence is reported as exactly that — not laundered into high confidence. See refs [1], [2]. Verdicts built on thin evidence are barred from confident directional language.

3.2No verdict is a population-wide average

Every verdict is disaggregated — by geography and development tier, by socioeconomic strata, by power position, by place on the adoption curve, by generation, and by the institutional capacity of the surrounding hospital, court, or state.The divergence between frontier and low-connectivity economies is treated as a first-order finding, not a footnote. A verdict that reports only a population average is treated as a framework violation, because that is precisely how false consensus is manufactured.

3.3Time as a compounding uncertainty

Judgments sit on a six-rung timescale ladder, with confidence scaled down explicitly as the horizon lengthens. Beyond fifteen years verdicts become conditional statements; beyond forty, only structural possibility space remains.

3.4Value disagreement is never averaged

On contested normative questions, each verdict is worked out in parallel from at least three value frames — welfare-consequentialist, rights-based, and distributive-justice — plus frames appropriate to the domain.Where the frames converge, that agreement is itself reported — genuine cross-frame convergence is rarer and more informative than any single frame's verdict. The framework does not adjudicate between them to produce one final answer; that is a political act outside its mandate. Where they diverge sharply, the divergence itself is labeled — a signal that the disagreement is about values, not facts, and no new evidence will resolve it.

4The Domain Taxonomy

The taxonomy is deliberately wide. A domain earns its own chapter only when AI's mechanism of effect there is distinguishable from its parent, when the evidence clears at least the theoretical tier, and when it is not already captured by a cross-domain edge. Candidates that fail the last test become edges, not chapters — which keeps the tree from proliferating without limit.

5Selected Figures

Each chapter's visualization specs render to figures like these — every one tied to a specific, dated, evidence-graded claim rather than a general impression.

6Reading & Reproduction

@techreport{shulman2026ai2030,
  title       = {The AI 2030 Project: A Framework-Generated
                 Assessment of Artificial Intelligence Across
                 Human Civilization},
  author      = {Shulman, Nathaniel},
  year        = {2026},
  month       = {7},
  version     = {1.0},
  note        = {Built on the Civilizational Impact
                 Assessment Framework (CIAF).},
  institution = {Independent}
}

7About the Author

§References

Methodological foundations the framework adapts or builds on. This is not the document's full bibliography, which is compiled per chapter.

1. GRADE Working Group. Grading quality of evidence and strength of recommendations. BMJ 328, 1490 (2004).
2. Intergovernmental Panel on Climate Change. Guidance Note for Lead Authors on Consistent Treatment of Uncertainties. (2010).
3. METR. Measuring AI Ability to Complete Long Tasks. (2025).
4. Stanford Institute for Human-Centered AI. Artificial Intelligence Index Report. (2025).
5. Frey, C. B. & Osborne, M. A. The Future of Employment: How Susceptible Are Jobs to Computerisation? (2013).
6. Tetlock, P. E. & Gardner, D. Superforecasting: The Art and Science of Prediction. Crown (2015).