A dated, public log of every figure corrected, superseded, or clarified on this site. Nothing is quietly edited away.
22 August 2026 — agent performance figures superseded
◈ WHAT CHANGED
Two figures on this site described agent capability using benchmarks that have since been replaced.
Was: "62% on a standard web-navigation benchmark, up from 14%." Now: the OSWorld computer-use benchmark rose from 12% to 66.3% by March 2026, then saturated. A harder replacement — OSWorld 2.0, released 26 June 2026 — puts the best available model at 20.6%.Stanford HAI 2026 AI Index · OSWorld 2.0 leaderboard, Aug 2026
Was: "24% of real professional tasks completed first attempt." That measurement stands for its own benchmark and date, but pass rates are being superseded by time-horizon measurement: the length of expert task an agent completes at 50% reliability, reported at roughly 2 hours 17 minutes in May 2026.METR time-horizon update, 8 May 2026
Why it was wrong: not an error in the original figures — an error in presenting them as if they described a fixed thing. They described a model against a ruler, and the ruler was replaced.
What changed on the site: the agents page now opens with the instrument change and states it explicitly. Homepage figures updated. Every agent figure now carries its benchmark name alongside its date.
◈ THE RULE
Three things get an entry: a factual error corrected; a figure superseded by newer data; a claim clarified because it was misleading as written. Typo fixes and design changes do not. If a number you read here changed after you read it, you will find out why on this page.
◈ WHERE THIS LOG BEGINS
This log records changes made to the site after publication — every figure revised, superseded or corrected once there were readers to mislead. It starts empty on purpose. An empty log on day one is honest; a pre-filled one would be theatre.
Before publication, every sourced claim on the site was traced to its issuing organisation over six verification passes, and a number of errors were found and fixed — an unverifiable statistic, a percentage that should have been a count, false precision on a figure the underlying research did not support, and a misattributed quote. None of those reached a reader, so none of them are listed below as corrections. They are described on the method page as what the verification process is for. What appears here from now on is the public record: what changed, when, and why.
2026
2026-08-02 · CLARIFICATION · SITE-WIDE
The site described itself as being about "dead internet theory." That framing overstated a claim the site's own content disputes — the record page documents that the widely repeated prediction of a 90% automated web by 2026 did not occur, and that the AI share of new articles plateaued rather than continuing to climb. The site was reframed as a sourced record of measurable change, with dead internet theory covered as one explained subject rather than the site's thesis. No figures changed.
2026-08-02 · DISCLOSURE ADDED
Added an explicit method page stating which sources qualify, what is excluded, and what this site is not — including that it performs no original measurement and aggregates the work of others.
2026-07-31 · RECORD PUBLISHED
First edition of the quarterly record. Baseline figures logged for future comparison: automated share of web traffic 53% (Imperva/Thales, Apr 2026, full-year 2025 data); new articles primarily AI-generated ~50% and plateaued since late 2024 (Graphite, Common Crawl analysis; quarterly figures vary 48–50% across analyses); AI-related fraud losses $893,346,472 across 22,364 complaints (FBI IC3, Apr 2026); 198 frontier models shipped since Nov 2022. Next revision due October 2026.
THE RECORD LINEA line you can check against later. That is the whole point of writing it down.
How to report an error
If a figure here is wrong, outdated, or attributed to the wrong source, it should be fixed. Corrections are welcome from anyone, and a correction that arrives with a primary source attached will be acted on faster than one without. Send corrections to aigptnet@gmail.com. A correction that arrives with a primary source attached will be acted on fastest.
WHY THIS PAGE EXISTS AT ALL
Almost nobody publishing statistics online keeps a corrections log. That is precisely why one is worth keeping: an outlet willing to record being wrong is making a claim that can be checked over time, and an outlet that never corrects anything is not more accurate — only less honest about it.