Pew Research Center pulled nearly 490,000 English-language pages out of the Common Crawl archive, ran them through a detection model called Open Pangram, and published the domain split on 20 August. In samples from 2026, around one in ten pages on a .com domain shows signs of AI authorship, against 4.6% on .org domains and about 1% on .edu and .gov, which Pew reports as ten times below the commercial rate.
The government figure is the one that will travel, and it is the one the study can least support. Pew publishes the numbers behind the chart (the objection is not the detector’s accuracy, which Pew addresses itself, but how few pages the government line is drawn from). The 2026 government point is 0.76%, drawn from 669 sampled .gov pages, so it stands on about five flagged pages. Every point on that line is a single-digit count.
The line behaves the way a small sample behaves. Commercial pages sit near 1% until ChatGPT and then rise at every reading after it, 1.07% in mid-2022 to 9.35% in 2026. Government pages move between 0.23% and 0.41% across the same pre-ChatGPT period, peak at 1.72% in mid-2024, and then fall by more than half while .com, .org and .edu all rise. Read as adoption, that says American government websites have been carrying less machine-written text each year since 2024.
Nobody believes that. The government line is reporting its own sample rather than the state, and a sample built to answer this question about government domains would be a different study from this one.
Assorted links
-
Parents will pay more than 60% more for a premium AI subscription when told four-fifths of their child’s peers use one rather than a fifth. Telling them what the long-run risks are moves their stated beliefs a long way, raises their support for banning AI in education for everyone, and does little to curb their own demand.
-
The Global AI Regulation Tracker has put its data behind a paid interface for machines, an API and a Model Context Protocol server that lets an assistant query the tracker directly, on its author’s own account of why: the site was paid for by ad revenue and therefore by search traffic, and getting cited by a chatbot instead means letting it take the content, often for nothing.
-
A case for spending the AI dividend on care work rather than on headcount reductions, priced entirely in American data: health care and social assistance employing 23.9 million people in June 2026, against a median hourly wage of $17.21 for home health and personal care aides in May 2025.
-
Nine authors take the standard three-part map of a political party, its people in public office, its central staff and its members on the ground, and ask what AI is doing to each. They separate what can already be observed from what is merely prospective.