🔍 Read the full analysis: What We Count When We Evaluate AI on ThorstenMeyerAI.com
TL;DR
A November 11 essay from ThorstenMeyerAI.com uses the history of national income accounting to examine how AI is evaluated. Its central argument is that market prices and benchmark scores capture only part of the value people receive; the source does not provide a detailed AI evaluation framework or independently verified findings.
ThorstenMeyerAI.com published an essay on November 11 arguing that the limits of gross domestic product offer a useful way to think about what AI evaluations count—and what they leave out. The piece traces national income accounting from the Great Depression and applies its lessons to AI, but it does not set out a new benchmark or report original evaluation results.
The essay begins with the creation of a national income estimate during the Great Depression. In 1932, the U.S. Senate asked the Commerce Department to estimate national income. Economist Simon Kuznets led the work, and a report covering 1929–32 was submitted to the Senate on January 4, 1934. It estimated that income produced fell from $83.0 billion in 1929 to $39.4 billion in 1932, a decline of 53%, according to the source.
Kuznets also warned that such estimates were not a complete account of social welfare. The report primarily counted activities whose results appeared in markets, leaving out unpaid household work and offering no measure of how income was distributed or how difficult it was to earn. The essay uses that distinction to frame its discussion of AI: a price, revenue figure or test score may be measurable without capturing all the value or effects of a tool.
To illustrate how market measures can miss value, the source cites a 2017 experiment in which the median participant would have required $17,530 to give up search engines for a year, despite paying no direct price for them. It also cites a Stanford working paper estimating $172 billion a year in value to U.S. consumers from AI chatbots. These are figures reported by the essay; they do not, on their own, establish a common method for comparing AI systems or a definitive measure of AI’s social value.
Infographic 01 · The sector clock, by output
China: how GDP changed shape
Share of GDP (value added at current prices) by sector, 1978 to 2025. In 2023 farming still held 22.8% of all jobs but produced 7.1% of GDP.
- Farming27.7%
- Industry47.7%
- Services24.6%
Reform starts
- Farming14.7%
- Industry45.5%
- Services39.8%
- Farming9.1%
- Industry45.4%
- Services45.5%
Services overtake industry
- Farming6.7%
- Industry35.6%
- Services57.7%
Communiqué 2025
Measured by output, not by jobs: the sector clocks of earlier episodes count workers. Jobs vs. output for farming, 2023: 22.8% of employment (year-end) vs. 7.1% of GDP.
Sources: NBS China, China Statistical Yearbook 2024, tables 3-2 and 4-2; NBS, Statistical Communiqué 2025 (28 Feb 2026).
Powered by Thorsten Meyer AI
The Limits of AI Value Measures
The essay’s main point is that what is easy to count is not necessarily the whole outcome. AI assessments often rely on observable measures, such as performance on defined tasks, user adoption, or commercial revenue. Those can answer specific questions, but a high score or large user benefit estimate does not by itself show how a system affects overall welfare, who benefits, or what costs accompany its use.
The comparison matters because AI tools can provide services at little or no direct charge, while their value to users may not appear as revenue. Conversely, a benefit estimate does not settle questions about reliability, harms, distribution or the quality of the underlying research. The source’s GDP analogy is a warning about interpretation, not proof that AI evaluation and economic accounting are identical.
For readers, the practical takeaway is to ask what an evaluation measures, who or what it leaves out, and what its baseline is. This is relevant to companies choosing systems, policymakers weighing their effects and users trying to understand claims about AI’s performance or economic contribution.
Popular right now
Ninja Luxe Café Mini Plus Espresso and Coffee Maker with AutoFroth, Bone & Champagne
Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Bluetooth 5.3, USB-C Charging, and IPX4 Water Resistance
Dyson Ball Animal 3 Upright Corded Vacuum
As an affiliate, we earn on qualifying purchases.
From Kuznets to AI Accounting
The source places the 1934 report in a longer history. National income estimates were developed to help officials understand the economic collapse, when available indicators included stock prices, freight-car loadings and incomplete industrial production measures. In the 1940s, U.S. accounts expanded to help plan wartime production. The first international System of National Accounts followed in 1953, and the United States adopted GDP as its headline measure in 1991.
GDP records the market value of final goods and services produced within a country, organized in the U.S. account as consumption, investment, government spending and net exports. It is a measure of production, not a direct score of wellbeing. The source emphasizes that this distinction was stated by Kuznets alongside the early estimates, rather than discovered after GDP became widely used.
The essay then turns to unpaid work and free digital services as examples of activity that can be useful without producing a corresponding market price. It presents AI chatbots as a newer case in which consumer value may exceed provider revenue. That comparison supplies the article’s framing; the cited estimates are not a substitute for system-level testing or a full accounting of costs and benefits.
What the AI Figures Cannot Establish
The essay’s cited figures do not establish how the estimates were calculated, what assumptions they use, or how uncertainty was handled. The supplied material identifies the chatbot valuation as coming from a Stanford working paper and the search-engine figure as coming from a 2017 experiment, but it does not provide the studies’ full methods, sample details or ranges of results.
It is also unclear what the essay means by a complete evaluation of AI in operational terms. The source excerpt does not specify a standard set of benchmarks, a method for weighing benefits against risks, or a way to compare systems across tasks. It does not report new tests of AI products, nor does it show that consumer-value estimates can measure performance, safety or social impact.
Accordingly, the figures should be read as estimates attributed to outside research, not settled measurements of AI’s total value. The source also does not establish that the valuations apply equally across users, products or countries.
Questions for Future AI Evaluation
The essay’s argument points toward a continuing debate over AI measurement: evaluations need to state clearly what is being measured and what is outside their scope. For any specific system or claim, readers should look for the task tested, the comparison baseline, the population represented and whether results cover only performance or also costs and effects.
The source does not announce a follow-up study, policy decision or new evaluation standard. It presents the GDP history as the opening installment in a series on basic concepts, with later discussion expected to address productivity. For now, the next step for readers is to distinguish measured task performance, estimated consumer value and broader social welfare rather than treating them as interchangeable.
Key Questions
What is the actual development in the article?
ThorstenMeyerAI.com published an analysis on November 11, 2026, using the history and limitations of GDP to discuss what AI evaluation measures and may omit. It is an argument and historical account, not an announcement of a new AI benchmark.
What did Simon Kuznets’ 1934 report estimate?
The report estimated that U.S. income produced fell from $83.0 billion in 1929 to $39.4 billion in 1932, a 53% decline, according to the source. Kuznets cautioned that national income was not a complete measure of welfare.
Why does the essay compare GDP with AI evaluation?
Both raise questions about the scope of a measure. GDP counts market production but does not capture every form of work or wellbeing; the essay argues that AI scores and revenue can likewise leave out parts of user value or wider effects.
Does the article prove that AI chatbots create $172 billion in value each year?
No. It cites a Stanford working paper’s estimate of $172 billion a year in value for U.S. consumers. The supplied material does not give the study’s methods or uncertainty range, so the figure should be treated as an attributed estimate, not a definitive total.
Source: ThorstenMeyerAI.com
