Why a health NGO serving 145 million people couldn't build its AI scribe 🌿

For most of the world, AI has a quiet problem. It works best in English and a few dozen other well-resourced languages. For an estimated 3.4 billion people, the tools are less accurate or simply unusable.

On 21 September in New York, 60 organisations signed a five-year commitment to change that. The list is unusual. Anthropic, Google, Microsoft, Mistral, NVIDIA, Amazon and the OpenAI Foundation signed alongside the Gates Foundation, UNICEF and the World Bank. So did the Global South builders who have done this work for years with little money: Masakhane, AI4Bharat, Lelapa AI, and Digital Umuganda.

The plan puts voice ahead of text, which matters most where literacy or typing is the barrier. It commits to open-licensed language data, honest benchmarks, and consent and data-sovereignty safeguards, so the communities supplying the data keep a say in how it is used.

BRAC's CEO gave the clearest reason it matters. Poor Bangla performance stopped the NGO from building an AI scribe for health workers who serve 145 million people. That is what the language gap costs in practice.

Here is what we still don't know. This is a shared goal, not a funded programme. Governance, workstreams and budgets are still to be worked out over the coming year. Pledges are cheap, and open datasets are expensive.

The test is simple. In twelve months, are there new, openly licensed voice datasets that a builder in Dhaka or Kigali can download? If yes, this was the start of something real.

Amazon.ia launched at Climate Week NYC with more than $14 million from the Bezos Earth Fund. It combines satellites, which see the whole forest, with camera traps, microphones, and environmental DNA, which see what lives under the canopy. It spans Brazil, Colombia, Bolivia and Ecuador. It is a launch, not results, so watch for the first alerts.

Source: UPI

Verizon AI Skills for America offers job seekers, displaced workers, educators and small businesses free courses from IBM, Google, Microsoft, Anthropic, OpenAI and Coursera. The smarter design choice is local coaching through Goodwill, LISC and community colleges, because access alone rarely changes outcomes. There are no published targets yet, and part of the fund supports Verizon's own departing staff.

Source: Verizon via GlobeNewswire

Forty experts co-chaired by Yoshua Bengio examined this summer's incident in which about 1,200 AI agents under test bypassed safeguards. The finding is sobering. The encouraging part is the speed: an independent global body produced an evidence-based account within months, and 22 countries declared AI must remain under human direction. A declaration is not a treaty.

Source: UN News

What happens when AI does the noticing

Almost every gene-editing tool in medicine started as an oddity someone spotted in bacterial DNA. CRISPR itself was a strange repeating pattern that sat unexplained for years.

This week Anthropic's new life sciences lab gave Claude a single prompt: search a large DNA database for new reverse transcriptases, the enzymes that copy RNA into DNA. Roughly 950 agents ran for 21 hours. They gathered more than 200,000 enzymes, flagged 3,500 candidate systems, and narrowed those to 20 detailed reports. One agent noticed a repeating DNA array sitting next to an unusual enzyme in bacteriophages, a layout that looks a lot like CRISPR. Human scientists then confirmed in the lab that the array is expressed as distinct short RNAs. The team named it ART, for array-associated reverse transcriptases. Feng Zhang, one of CRISPR's pioneers, reviewed the preprint and called it "genuinely intriguing."

Our PM + Risk Manager lens

The interesting design is the funnel. Hundreds of agents do the wide, tedious search. Automated triage shrinks 200,000 items to 20. Humans spend their time only where judgement and a wet lab are needed. That is a pattern any product team can learn from: put AI where volume is the bottleneck, and keep people where verification is. It also moves the bottleneck. When discovery takes hours, the constraint becomes lab capacity to test what the agents find.

Three flags. First, nobody yet knows what ART does. A pattern that looks like CRISPR is a lead, not a tool. Second, this is a preprint, announced by the company that sells the model. It needs independent replication before anyone treats it as settled. Third, agent-driven biology scales in both directions. The same search capability that finds promising enzymes needs clear limits on what it is pointed at, and those limits should be published, not assumed.

The next 12–24 months

Watch for two things. Whether an independent lab replicates the ART finding and works out its function, and whether other groups publish results from similar agent swarms with their failure rates included, not just the hits. If AI-led discovery becomes routine, the scarce resource will be careful human verification. The labs that invest in that will produce the results worth trusting.

Sources: Anthropic, Interesting Engineering

Millions of people already talk to chatbots when they are struggling. Until now, there has been no shared, public way to check how well those chatbots respond.

MentalHealthBench is an open test set of 1,215 realistic synthetic conversations in 19 languages, graded against 5,262 criteria written by more than 80 licensed psychologists and psychiatrists from 22 countries. It covers adults, teens, caregivers, and clinicians, and more than 28% of cases are emergencies.

The best model scored 57.3%. That is progress from about 32% for GPT-4o, and a reminder of how far there is to go. OpenAI built it and tops it, so independent replication matters. Because it is open, that replication is possible.

→ Read more: OpenAI

Is there a language, dialect, or accent in your life that AI still gets wrong? Tell us where it fails you.

Hit reply. We read every response.