Methodology

How the scores are computed. The methodology replicates Simpson (2026). All code is open source.

Data source

The Anthropic Economic Index is published roughly quarterly on HuggingFace. Each release contains country-level data on Claude usage patterns collected over a one-week sampling window: conversation counts, collaboration patterns, task success rates, use case classifications, and education-level estimates.

Countries are included when they have at least 200 conversations in the sampling window. Population data for per-capita calculations comes from the World Bank Open Data API.

Agency composite

The agency score is the mean of up to four min-max normalised components:

• Co-creation rate: task iteration + learning + validation as a share of all conversations
• Directive rate (inverted): share of purely directive conversations, flipped so less directive = higher agency
• Task success rate: share of conversations classified as successful
• Prompt education level: mean education-years estimate of prompt content

Each component is min-max normalised across all countries in the release to a [0, 1] range. The composite is the arithmetic mean. When a component is not available in a release (see variable availability), it is excluded from the average.

Access score

Access is calculated as:

access = log₁₀(usage_count / population)

The log transform compresses the right-skewed distribution of per-capita usage rates, making the scatter plot readable across several orders of magnitude.

Stage assignment

Countries are assigned to stages using sequential threshold rules applied to use-case share data. The rules are applied in order; the first match wins:

1. Stage 1 (Full Dependency): coursework share > 30%
2. Stage 2 (Elite Empowerment): work share > 48% AND coursework share < 25%
3. Stage 3 (Passive Dependency): personal share > 38%
4. Residual: agency composite score ≥ 0.55 → Stage 2; ≥ 0.42 → Stage 3; otherwise Stage 1

When use-case data is not available (as in the September 2025 release), all countries are classified using the residual agency-score rules.

Variable availability by release

Limitations

• Single platform. The AEI captures Claude only. Countries where ChatGPT or Gemini dominate will look lower-access than they are.
• Selection bias. Claude skews English-speaking and higher-income. Anglophone access scores are likely inflated.
• One-week window. Each release samples a single week, which may not reflect longer-term patterns.
• Threshold sensitivity. Stage rules use fixed cutpoints from the paper. Small changes can move individual countries between stages.
• Variable availability. Earlier releases have fewer agency components, so cross-wave comparisons are less precise.

Papers

The AI Matrix framework is developed across two foundational papers. Two further papers provide empirical support using the same data sources this site draws from.

Foundational

Supporting evidence

Citation

If you use this data or methodology, please cite:

@article{simpson2026aimatrix,
  title={The AI Matrix as Diagnostic: Access, Agency,
         and Adoption},
  author={Simpson, Ewan},
  year={2026},
  publisher={Zenodo},
  doi={10.5281/zenodo.18181372}
}

@article{simpson2026decomposing,
  title={Decomposing the Capability Overhang: Access, Agency,
         and the Geography of AI Adoption},
  author={Simpson, Ewan},
  year={2026},
  publisher={Zenodo},
  doi={10.5281/zenodo.19250306}
}