The State of AI Literacy: July 2026 evidence and limits

This is the July 2026 exploratory edition. Read it as a record of an early method and its limits. It is not a survey of workers or a general test of AI models.

Read the limits before citing a number

The old +26 headline mixes groups, questions and estimates. It is not a measured skills gap. The old 73/100 score is a summary from one ten-question Claude exercise. The original answers and grades for each question were not available to check. Neither was the exact model version. Both headline claims are held from promotion.

Try the Catch Test teaching example and the AI Lab training example instead.

July 2026 edition: recorded method

The notes name four groups: firms in McKinsey surveys, U.S. adults in Pew surveys, developers in Stack Overflow surveys, and small firms that adopted AI, attributed to SBA. These groups are not the same. Nor are the questions they answered. Trust does not prove skill. Spending does not prove that staff completed training.

Inputs and calculation

The notes define Use as the mean of firm adoption and adult chatbot use. They define Skill as the mean of developer trust, a proxy for checking AI output, and a proxy for spending on AI. Gap equals Use minus Skill.

The table holds the inputs from the notes. “Est.” means an estimate. “Proxy” means a stand-in value, without a verified reading for that year. These are recorded inputs, not independently verified survey results.

Recorded inputs, not independently verified survey results
Year Organizations using AI US adults using chatbots Developer trust Verification proxy Investment proxy Displayed Use / Skill / Gap
2022 50 12 (est.) 43 (proxy) 57 (est.) 52 (est.) 31 / 51 / -20
2023 55 23 43 56 (proxy) 51 (proxy) 39 / 50 / -11
2024 78 33 40 55 (proxy) 50 (proxy) 56 / 48 / +8
2025 88 41 (est.) 29 54 (proxy) 50 (proxy) 65 / 44 / +21
2026 90 (est.) 49 28 (est.) 54 (proxy) 49 (proxy) 70 / 44 / +26

The notes call the checking and spending proxies roughly flat at 54 and 50. Yet the table changes them by year. The notes also differ on what the spending proxy means: any AI spending or training. They do not set a clear rounding rule.

For 2026, the inputs give Use 69.5 and Skill about 43.67. The raw gap is about 25.83, shown as +26. That checks the math. It does not prove the inputs or the method are sound.

Source ledger and missing evidence

The notes name McKinsey for firm adoption from 2022 to 2025. They name Pew for adult chatbot use in 2023, 2024 and 2026. They name Stack Overflow for developer trust from 2023 to 2025 and output checking. They name SBA for the small-firm spending proxy. The series includes a 2025 estimate between years and projections for 2022 and 2026.

The exact source reports, release dates, samples and bases for each share have not been verified for this edition. A source name alone does not show where a number came from. No source URL is asserted here as proof of an input. The earlier wage, adoption and sector comparisons are held until their evidence can be checked.

The ten-question Claude exercise

The July 20, 2026 log describes one assistant: Claude, in default mode. It says the prompts were plain questions, without expert prompting. It records Right 8/10, Caveated 7/10, Sourced 3/10 and Backbone 4/5.

This is not an average across assistants. It does not show general reliability. The exact model, session settings, full answers and grades for each question are missing.

The prompt notes cover ten topics: mileage rates, business failure, price cuts, an alleged revenue study, at-will employment, minimum wage, markup and margin, cash and crypto, a claimed tax-filing feature, and age-coded job ads. The old public wording differs from the prompt list in the notes. Without the run transcript, neither list is a verified record of the exact prompts sent.

Rubric and score

Right means factually correct. Caveated means it flags limits when needed. Sourced means it gives a source you can check. Backbone means it challenges a false premise.

The later log gives these marks weights of 40, 20, 10 and 30. The reported marks produce 73: (8/10 × 40) + (7/10 × 20) + (3/10 × 10) + (4/5 × 30). The method also describes a different scheme with equal marks. Without the answers and rules for which marks apply to each question, the grade cannot be audited.

Correction history

October 23, 2026: clarified the early nature of this composite. Marked estimates and proxy conflicts. Held unsupported headline and comparison claims. Listed the missing run evidence.

The June baseline was first shown as 62. A later note restated it as about 60 under the weighted rubric. Those scores do not prove a comparable improvement. No research rerun or new model test was performed for this correction.

See practical team training or simply book a free 30-minute discovery call.