This week made the trade-off legible. The same systems being deployed to read mammograms, triage refugees, and tutor rural students are also drawing down water tables, spiking residential power bills, and exporting their waste to countries that never asked for it. The technology is doing real good and real harm in the same week, often funded by the same checks. The honest question isn't whether AI helps — it does — but who absorbs the cost when it doesn't.
The Big Story
ChatGPT for Clinicians launches free, with HealthBench Professional benchmark · April 22 · [OpenAI]
→ OpenAI made a verified-clinician version of ChatGPT free for US physicians, NPs, PAs, and pharmacists, with clinical search, documentation flows, and journal-grade deep research baked in. The AMA already says 72% of US doctors use AI in practice; this drops the price floor on those tools to zero. The catch: "free for US clinicians" is a narrow definition of access when most of the world's doctor shortage is in countries OpenAI hasn't verified into.
Also This Week
Fortune: data centres cost the US $25B/year in hidden health and environmental damages · April 21 · [Fortune]
→ A single Northern Virginia facility running on-site generation racks up $53-99M a year in health damages alone, per a Piedmont Environmental Council study; PJM auction prices are forecast to push residential rates up 20% across 13 states this summer. The cost of inference is being socialised onto the grid.
India's eSanjeevani has now processed 282 million AI-assisted teleconsultations · April 2026 · [ICTworks / SAHI Framework]
→ India's new SAHI strategy formalises governance for systems already running at population scale — including AI-enabled TB detection that has cut adverse outcomes by 27% inside the National TB Elimination Programme.
Google.org closes $30M AI for Government Innovation Challenge · April 3 deadline; AI for Science deadline April 17 · [Google]
→ Two challenges totalling $60M for nonprofits, social enterprises, and academics; the AI Opportunity Fund adds $17M for nonprofits in the US, EMEA, and Asia Pacific to actually learn the tools. Cash is no longer the bottleneck.
Watershed launches sustainability AI agents and a fellowship at SF Climate Week · April 21 · [GlobeNewswire]
→ Data-cleaning agents that the company says cut time-to-actionable-data by 80% across pilot customers — a useful tell that "AI for climate" is mostly tedious accounting work, not moonshots.
Meta opens Teen Account topic visibility to parents across US, UK, AU, CA, BR · April 23 · [TechCrunch]
→ Topic-level, not conversation-level — a deliberate compromise that gives parents enough signal to start a conversation without handing over the transcript. The interesting test will be whether teens just route around it.
SK Group signs MOUs with Vietnam for AI ecosystem and a 1,500 MW LNG plant · April 24 · [Korea Times]
→ The "AI for development" pitch increasingly arrives bundled with fossil-fired baseload. Vietnam gets the data centres and the gas plant; the emissions stay onshore.
From the Lab
BRAIx mammography risk score, validated across ~500K women in Australia and Sweden · [The Lancet Digital Health]
→ An AI scoring tool flags women at four-year breast cancer risk more accurately than density, family history, or country of birth — almost 1 in 10 women in the top 2% by score were diagnosed within four years despite getting an "all clear" from standard screening. The kind of result that justifies the hype, if it survives prospective deployment outside the original cohorts.
NEJM AI: Therabot RCT shows generative-AI chatbot reduced symptoms in MDD, GAD, and high-risk eating disorders · [NEJM AI]
→ The first RCT of a fully generative therapy bot to show clinical-level effect — paired this month with new safety work showing voice-mode chatbots increased loneliness with longer use. Both findings are real; both will be cited selectively.
Worth Reading
- [Why AI alone cannot fix social problems] — Rest of World, April 21: a sober field-tested rebuttal to the "deploy a model, fix a system" pitch. AI keeps failing where institutions are weak, not where models are weak.
- [AI is about to make the global e-waste crisis much worse] — Ananya Bhattacharya on where the GPUs go to die — disproportionately in non-Western countries that bore none of the upside.
- [CNN: fixes for AI's grid problem exist but aren't being implemented] — Re-conductoring, storage, and fusion all exist on paper; deployment lags 7-10 years behind demand. OpenAI shuttered Sora partly because the compute wasn't there.
The good news this week is that AI is finally working on hard problems. The bad news is the same.
— Alexis