Two AI tools for England’s planning system were announced together and only one of them is live. Extract, which turns decades of paper planning documents and handwritten notes into usable data, went out to all local planning authorities in England that day after trials in 20 of them. The other, a prototype that triages an application and hands an officer a first assessment, is in early-stage testing with three councils. The 250,000 hours a year in the headline belongs to the document conversion, not to the triage.
MIT’s committee on AI in teaching reports that campus culture moved in under three years: fewer students at office hours, less participation in online discussions and, anecdotally, fewer study groups meeting in dorms and libraries. Its recommendation is not to patch the existing system but to redesign learning experiences, assessments and curricula. The undergraduate research programme it wants to build on engages 93% of undergraduates and 58% of faculty.
Three research groups analysed 249,834 Claude conversations without reading any of them, working from aggregated outputs Anthropic produced rather than from transcripts. They report that active teaching appears in 67% of conversations and that its benefits are unevenly realised. The figure circulating from this paper, that 56% of actionable tasks were consequential, is not in it: that sentence comes from the host site’s machine-written summary sitting above the abstract.
Google’s Earth AI prediction engine nowcast an Ebola outbreak in the Democratic Republic of the Congo and downscaled Nigerian food-security estimates from province to local government area, identifying 15 of 18 newly invaded health zones across five weekly forecasts, about ten points above the published Bayesian baseline. The work was done with the World Food Programme’s vulnerability analysis team and the Institut National de Recherche Biomédicale. Google calls it an early-stage research project and offers no way to use it.
EA Funds has closed the Long-Term Future Fund and opened a successor in its place, aimed at risks from advanced AI, making grants of typically $10,000 to $150,000 and rarely above $300,000. The page publishes no total moved and no count of grants made.
An outside evaluator has tested a Google model without seeing its weights, while Google never saw the test questions. Both sides’ assets stayed sealed inside Confidential Space, with the Singapore AI Safety Institute, OpenMined, AVERI and MLCommons taking part. It ran on one Gemini Flash Lite model, DeepMind calls it a pilot, and the post contains no numbers at all.
Osama Manzar says India’s data-centre build-out is being negotiated without the people whose land it uses: tax exemptions and land at low cost on one side, and on his account no consultation with any of the affected groups, for facilities automated enough that he doubts local people will be employed in them. His counter-proposal is 250,000 data centres at village-council level.
Social workers meet AI systems as users of them, as subjects in their datasets, and as first responders to what those systems deploy, and study them from outside the settings where the decisions are made. Five authors map five groups of technology decision roles the profession could hold instead, running from product work through to policy.
Fifty-seven AI professionals were interviewed twice, in 2021 and in 2023, before and after the surge of public interest, alongside a text analysis of millions of news articles and social posts. The three arguments organising how people make sense of AI turn out to be about method, top-down expert systems against bottom-up emergent capabilities, about mind, a passive tool against a humanlike digital mind, and about morality, whether to slow development down or speed it up.
Today’s argued piece: Five of 475 studies on surgical AI in poorer countries came from low-income countries.