Drafted by AI. Countersigned by a human.
Corrections
If you find an error, write to corrections@countersigned.org. Include the article URL, the passage in question and a source supporting your proposed correction. Please start with a description and public link rather than confidential information.
Material corrections and updates explain what changed and how the conclusions are affected.
The AI-usage figures in this article's disclosure box were too high. They had been measured with a tool that counted some answers of the AI model more than once, because the session log writes one answer on several lines. The figures were re-counted from the same session log for the same point in time. Before: 4,467 model calls (each a request to the model and its answer), about 1,396.3 million tokens processed and 1.7 million generated (tokens are the small units of text the model reads and writes). Now: 2,204 model calls, about 688.9 million tokens processed and 671 thousand generated.
Effect on conclusions: None. The text, figures, sources and conclusions of the report are unchanged; only the AI-usage figures were corrected.
The AI-usage figures in this article's disclosure box were too high. They had been measured with a tool that counted some answers of the AI model more than once, because the session log writes one answer on several lines. The figures were re-counted from the same session log for the same point in time. Before: 2,943 model calls (each a request to the model and its answer), about 729.4 million tokens processed and 1.6 million generated (tokens are the small units of text the model reads and writes). Now: 1,414 model calls, about 356.7 million tokens processed and 592 thousand generated.
Effect on conclusions: None. The text, figures, sources and conclusions of the report are unchanged; only the AI-usage figures were corrected.
The AI-usage figures in this article's disclosure box were too high. They had been measured with a tool that counted some answers of the AI model more than once, because the session log writes one answer on several lines. The figures were re-counted from the same session log for the same point in time. Before: 2,823 model calls (each a request to the model and its answer), about 721.3 million tokens processed and 4.2 million generated (tokens are the small units of text the model reads and writes). Now: 1,134 model calls, about 270.7 million tokens processed and 1.1 million generated.
Effect on conclusions: None. The text, figures, sources and conclusions of the report are unchanged; only the AI-usage figures were corrected.
The AI-usage figures in this article's disclosure box were too high. They had been measured with a tool that counted some answers of the AI model more than once, because the session log writes one answer on several lines. The figures were re-counted from the same session log for the same point in time. Before: 501 model calls (each a request to the model and its answer), about 187.5 million tokens processed and 683 thousand generated (tokens are the small units of text the model reads and writes). Now: 239 model calls, about 97.8 million tokens processed and 293 thousand generated.
Effect on conclusions: None. The text, figures, sources and conclusions of the report are unchanged; only the AI-usage figures were corrected.
The AI-usage figures in this article's disclosure box were too high. They had been measured with a tool that counted some answers of the AI model more than once, because the session log writes one answer on several lines. The figures were re-counted from the same session log for the same point in time. Before: 4,024 model calls (each a request to the model and its answer), about 1,004.1 million tokens processed and 4.8 million generated (tokens are the small units of text the model reads and writes). Now: 1,947 model calls, about 439.6 million tokens processed and 2.5 million generated.
Effect on conclusions: None. The text, figures, sources and conclusions of the report are unchanged; only the AI-usage figures were corrected.
The AI-usage figures in this article's disclosure box were too high. They had been measured with a tool that counted some answers of the AI model more than once, because the session log writes one answer on several lines. The figures were re-counted from the same session log for the same point in time. Before: 3,942 model calls (each a request to the model and its answer), about 691.1 million tokens processed and 3.7 million generated (tokens are the small units of text the model reads and writes). Now: 1,855 model calls, about 309.5 million tokens processed and 2.2 million generated.
Effect on conclusions: None. The text, figures, sources and conclusions of the report are unchanged; only the AI-usage figures were corrected.
The AI-usage figures in this article's disclosure box were too high. They had been measured with a tool that counted some answers of the AI model more than once, because the session log writes one answer on several lines. The figures were re-counted from the same session log for the same point in time. Before: 6,982 model calls (each a request to the model and its answer), about 804.7 million tokens processed and 4.4 million generated (tokens are the small units of text the model reads and writes). Now: 2,964 model calls, about 346.1 million tokens processed and 2.7 million generated. Only the part measured from Claude Code session logs was re-counted; the part measured from Codex logs is unchanged.
Effect on conclusions: None. The text, figures, sources and conclusions of the report are unchanged; only the AI-usage figures were corrected.
The AI-usage figures in this article's disclosure box were too high. They had been measured with a tool that counted some answers of the AI model more than once, because the session log writes one answer on several lines. The figures were re-counted from the same session log for the same point in time. Before: 261 model calls (each a request to the model and its answer), about 97.8 million tokens processed and 1.9 million generated (tokens are the small units of text the model reads and writes). Now: 46 model calls, about 20.7 million tokens processed and 247 thousand generated.
Effect on conclusions: None. The text, figures, sources and conclusions of the report are unchanged; only the AI-usage figures were corrected.
The AI-usage figures in this article's disclosure box were too high. They had been measured with a tool that counted some answers of the AI model more than once, because the session log writes one answer on several lines. The figures were re-counted from the same session log for the same point in time. Before: 3,210 model calls (each a request to the model and its answer), about 542.3 million tokens processed and 6.3 million generated (tokens are the small units of text the model reads and writes). Now: 1,333 model calls, about 178.5 million tokens processed and 1.9 million generated.
Effect on conclusions: None. The text, figures, sources and conclusions of the report are unchanged; only the AI-usage figures were corrected.
The AI-contribution note stated that the publisher's editorial review was pending. The review had been completed on 12 October 2026 before publication, as the editorial-review block on the same page records; the sentence now says so. The links to the report's videos and the recorded publication time were added in the same release.
Effect on conclusions: None. The text, figures, sources and conclusions are unchanged.