Amelia Arsenault and Sarah Kreps asked 386 Americans to comment on the 2019 proposal to repeal the Clean Power Plan. Half got ChatGPT. The rest wrote on their own. The ones with the tool found it easier, and 43 evaluators rated their comments about 46 per cent better. They understood the rule no better than those without it.
On one measure they did improve: the share who could not describe accurately what it did fell from 31.6 to 21.2 per cent. But the gap between better- and less-educated writers held, because both improved about equally.
An agency need not weigh submissions by number, and they were never scarce: Pew found 94 per cent of the 21.7 million comments on the 2017 net-neutrality repeal were duplicates.
Agencies now use the same models to read what arrives. Sola Kim and colleagues took 182 real comments and changed only the job of the supposed author. The same words, from a street vendor rather than a financial analyst, came back summarised with more of their meaning gone. The effect is small, and differs by which company built the model, so the choice of supplier changes how much of a street vendor’s meaning survives. Federal procurement rules do not check for that.
Assorted links
-
Nepal’s identity system was built in 2018 for up to 6,000 verifications an hour and now manages 600 to 700 a day. Land registration has been down for weeks in all 77 districts.
-
Raghvendra Kumar Chaudhary argues that India’s constitution assumed a human official would decide, and finds a right to human decision-making in Articles 14 and 21.
-
Rystrøm and colleagues split AI in government into five kinds: hand-coded, glass-box, black-box, general-purpose and agentic. Public-administration research, they argue, treats them as one.
-
Ghana, UNESCO and the European Union have trained 240 civil servants as AI trainers. Gleb Tsipursky asks what counting certificates shows that watching the work would not.
-
Algeria’s AI roadmap, from a 5 August meeting, sets out sovereign open-source models, national computing centres and public-sector data held at home.