Does GDP per Capita Really Predict AI Adoption? Adjusting Anthropic's Estimates

A Critique

Authors: Ilan Strauss AI Disclosures Project (v.2)

---

Summary: Anthropic's Claims vs. Our Critique

Anthropic's Claim	Critique	Our Evidence
"Uneven [AI] adoption remains well-explained by GDP per capita" (global β=0.71)	GDP per capita explains little of middle-income country adoption	Middle-income group: R²=7% (with Seychelles), R²=17% (without Seychelles)
AI usage concentration "essentially unchanged" (Aug–Nov 2025)	Gini captures a static snapshot; differential growth rates compound into widening gaps	Adoption changes across two time periods show potential inequality emerging: High-income +26%, Middle-income +14%, Low-income +22%
"Divergence in living standards" possible (McCrory to FT)	Speculative: no second-order effects estimated (AI → productivity → growth)	Cross-country data used is a snapshot from Claude front-endusage (closer to consumer), not the API (closer to firm-level), and not panel (countries over time)
Implied by FT coverage: data reflects global AI adoption	Data only measures Claude usage, not total AI adoption broadly	Brazil and Thailand saw decreased Claude usage. But could be from competitive use of alternatives (ChatGPT), not reduced AI adoption
Single global relationship applies to all countries	Different income groups have different relationships, making pooling biased (Hsiao 2022)	Separate regressions show relationships (slopes) and R² (fit) vary by income group

---

📖 Read Our Analysis

Format	Link	Description
✍️ Blog Post	BLOG_POST.md	Full blog post
🌐 Blog (Easy Read)	View Online	Simplified analysis

---

Overview

We take issue with two claims or inferences:

1. CORE ANTHROPIC CLAIM 1: Anthropic's Economic Index January 2026 Report claims "Worldwide, uneven [AI] adoption remains well-explained by GDP per capita."

• We show that the data does not support this. When broken down by income-group the fit is substantailly weaker and non-existant for middle-income countries, roughly one-third of the country sample.

2. CORE FINANCIAL TIMES CLAIM 2: The Financial Times covered Anthropic's research with the headline: "Rich countries’ greater use of AI risks deepening inequality, Anthropic warns".

• This is misleading since none of the Anthropic research covered in the article in question provides evidence on this claim. The head of economics at Anthropic warns in the article: "If the productivity gains...materialise in places that have early adoption, you could see a divergence in living standards."

• It is unclear how this relates to the research they have conducted which analyzes adoption of Claude, focusing on consumer adoption not firm-level (API), and not attempting to assess if changing adoption patterns reflects greater usage of ChatGPT or competing products instead

Weak and Highly Uncertain Relationship with Little Explanatory Power

Using two data releases from Anthropic (August 4-11 and November 13-20, 2025), we show that the GDP per capita - AI adoption relationship is weak and highly uncertain, across both time periods:

Table: Regression Results — Anthropic's Global Estimate vs. Our Country-Group Estimates

Income Group	Period	Slope (β)	Std. Error	p-value	Significant?	R²	N
Global	Aug 4-11	0.69	0.04	<0.001	Yes	0.71	114
	Nov 13-20	0.71	0.06	<0.001	Yes	0.56	116
Low income	Aug 4-11	0.76	0.19	<0.001	Yes	0.30	38
	Nov 13-20	0.85	0.18	<0.001	Yes	0.37	39
Middle income	Aug 4-11	🔴 0.44	0.18	0.019	Yes	0.14	38
	Nov 13-20 (with Seychelles)	0.73	0.44	0.105	No (~10%)	0.07	38
	Nov 13-20 (excl. Seychelles)	🔴 0.44	0.16	0.011	Yes	0.17	37
High income	Aug 4-11	0.63	0.20	0.004	Yes	0.21	38
	Nov 13-20	0.67	0.16	<0.001	Yes	0.33	39

Key observations:

• Middle-income relationship for the latest data release (Nov 13-20) is highly uncertain / not statistically significant (p = 0.105).

• Middle-income relationship weakens greatly when the huge outlier of the Seychelles is excluded (B = 0.44). The same estimate holds for Anthropic's previous data release (Aug 4-11).

• R² (goodness of fit) for middle-income fall is just 7%, or 17% when the Seychelles is excluded: in other words GDP per capita explains very little about relative AI adoption among middle-income countries in either case.

Repository Structure

anthropic-econ-critique/
├── README.md                    # This file
├── BLOG_POST.md                 # Full blog post
├── analysis_full.py             # Main Python analysis
├── analysis_results.csv         # Processed results
├── data/                        # Anthropic's original data
│   └── release_2026_01_15/      # November 2025 data (from Jan 2026 report)
├── original_report/             # Their report
└── figures/                     # Generated figures

Reproducing Results

pip install pandas numpy statsmodels scipy
python analysis_full.py

Data Source

Original data from Anthropic's HuggingFace repository:

• https://huggingface.co/datasets/Anthropic/EconomicIndex

Contact

• Ilan Strauss: ilan@aidisclosures.org
• AI Disclosures Project: https://ai-disclosures.org/