Opinions about AI are everywhere; good data is rarer. These recurring reports and datasets are the ones journalists, researchers and policymakers rely on.

1. Stanford AI Index

Published every year by Stanford’s Institute for Human-Centered Artificial Intelligence (HAI). It tracks research output, model performance on benchmarks, training costs, investment, policy activity and public opinion. Best for: long-term trends and neutral, well-sourced charts.

2. State of AI Report

An annual review published by investor Nathan Benaich and collaborators, covering research, industry, politics and safety, with predictions for the following year that are scored the next year. Best for: a fast, opinionated overview of the past twelve months.

3. IEA — Energy and AI

The International Energy Agency’s 2025 report analyzes the electricity demand of data centers and AI, and how energy systems could respond. Best for: credible numbers on the energy footprint of AI.

4. Epoch AI datasets

Epoch AI maintains public databases of notable models with their training compute, parameters and data size, plus analyses of trends in hardware and scaling. Best for: quantitative questions such as “how fast is training compute growing?”.

How to use reports well

  • Check the methodology section before quoting a number.
  • Note the date range — AI moves fast, and a report published in spring often covers data up to the previous year.
  • Look for the primary source behind charts you plan to cite.
  • Compare across reports when numbers disagree; differences usually come from definitions.

Other reports worth knowing

  • National and regional AI strategies published by governments explain policy priorities, funding and regulation plans.
  • OECD AI policy resources track AI policies, investment and indicators across many countries.
  • The International AI Safety Report, chaired by Yoshua Bengio, synthesizes the science on risks from advanced general-purpose AI.
  • Industry adoption surveys from consulting firms and analysts estimate how many companies use AI and for what — useful, but check sample sizes and definitions, since results vary widely.
  • Company reports — model cards, system cards and transparency reports — describe specific models’ capabilities, evaluations and safeguards.

Each type answers different questions. Research-oriented reports track capabilities and compute; policy reports track rules and investment; adoption surveys track business use.

How to read a chart from an AI report

  1. Read the axis labels and units. Is it compute in floating-point operations, cost in dollars, or a benchmark score?
  2. Check whether the scale is logarithmic. Many AI charts use log scales, where a straight line means exponential growth.
  3. Look at what is included. “Notable models” lists depend on the authors’ selection criteria.
  4. Find the data source. Charts often combine estimates from several sources with different methods.
  5. Note the period covered. The newest models may not be included yet.

Turning reports into insight

Rather than reading reports cover to cover, read the summary or key findings first, then the chapters relevant to your work. Note three numbers you might use, with their source and date, and one trend that surprised you. Comparing the same chart across several annual editions is often more revealing than any single year’s figures.

Citing reports correctly

When you quote a figure, cite the report name, edition year and page or chart, and note whether the number is an estimate. This makes your work verifiable and avoids spreading outdated statistics.

What to watch next

  • New editions of the AI Index and State of AI reports each year.
  • Better data on AI adoption and productivity in businesses.
  • More transparency from AI developers about training compute and data.
  • Energy and water reporting for AI data centers.

Key takeaways

  • Annual reports like the Stanford AI Index and the State of AI Report give data-driven overviews of AI.
  • The IEA’s Energy and AI report and Epoch AI’s datasets cover energy and compute trends.
  • Check methodology, dates and definitions before quoting numbers.
  • Use reports as starting points and go to primary sources for citations.

Frequently asked questions

Which AI report should I read first?

The Stanford AI Index is a good starting point: it is free, broad and data-driven, with summaries of key findings.

Are these reports free?

Most of the reports mentioned are free to read online, with downloadable data in several cases.

Why do reports give different numbers?

They use different definitions, data sources and time periods. Compare methods before concluding that one is wrong.