development
Assorted links for 1 October 2026
Seven links: a new complexity index finds Japan and China lead different halves of the AI-export supply chain, a 10-country survey finds most public servants already use AI while few trust their government's use of it, Uruguay's AI-readiness ranking, a Canadian policy magazine on AI flooding public consultations, Argentina and Delaware's competing proposals for AI legal personhood, a critical read of Ghana's AI strategy, and content-authenticity standards reaching consumer phones.
An AI model now reads Brazil's messy $75 billion medicine-procurement records
Brazil's public health system buys $75 billion of medicine a year through 263 different local systems with unstandardised records. A large language model now matches each purchase to a federal product code, cutting a 90-minute manual task to five seconds and lifting the match rate from 5% to 80%.
Assorted links for 30 September 2026
Five links: a Nigerian fintech AI-evaluation framework finds standard safety tests miss local fraud risks in both directions, India pitches its digital-infrastructure experience to the Global South at the UN, the Trump administration launches an AI front door for US federal services, Code for America and Anthropic pilot Claude for SNAP caseworkers, and Rondonia's state government adapts a wildfire-physics model for an AI early-warning system in the Amazon.
Non-Western governments warned the UN that pacing frontier AI can be gatekeeping
At the UN General Assembly's 81st session in September 2026, Pakistan, Nigeria, Argentina and other non-Western governments said proposals to pace frontier AI development for safety can function as gatekeeping favouring whichever countries already lead, while pressing their own proposals for a real hand in the rules.
Assorted links for 29 September 2026
Four links: Kenya and Anthropic sign a data-sovereignty AI framework at UNGA, Goa puts its state government portal behind a WhatsApp chatbot, a 118-student study finds generative AI's learning effect depends on how it is used, and a Brazilian legal commentary traces today's AI-governance gap to a 2017 audit of un-shared government data.
No AI safety index caught a chatbot mistranslating smallpox as syphilis
AI safety evaluation budgets concentrate in a handful of Western labs testing for model-level risks, while deployment failures in low-resource languages go untested until a user reports the harm, including a Tigrinya medical chatbot that rendered smallpox as syphilis and gonorrhea as diabetes.
Generative AI raised homework scores and lowered exam scores for 26,811 Chinese students
A CEPR working paper tracked 26,811 Chinese secondary students for 30 months and found generative AI raised homework scores while lowering exam and entrance-exam performance, concentrated among the students whose usage pattern looked like outsourcing the work.
Assorted links for 27 September 2026
Five links: South Africa's AI policy rewrite after an AI-hallucinated citation scandal, a mapping of US federal AI governance against sector vulnerability, African diabetic-retinopathy screening running on imperfect data, record US scam losses driving a shift to cryptographic digital identity, and an executable Governance-as-Code compliance pipeline for the EU AI Act.
ITU's GENIE.AI is an open-source AI stack now piloting in four countries
ITU assembled GENIE.AI from existing open-source tools -- OPEA, Docling, IBM Granite and Google Gemma -- so governments can run AI services without depending on one vendor. Pilots are already live in Lesotho, the Gambia and Bangladesh, with a fourth under way in El Salvador.
Assorted links for 15 September 2026
Six links: survey collectors across 3.4 million households systematically dropping the people who cost most to reach, a World Bank survey finding 44 per cent of government AI use is individual staff working ad hoc, nominal AI GDP put at roughly $250 billion, Rwanda approving a national AI agency, Africa holding under 1 per cent of the world's data centres, and an index that scores deployability rather than capability.
Assorted links for 14 September 2026
Five links: at least 39 submissions to Australian parliamentary inquiries carrying apparently hallucinated references, a retrieval system whose recall falls from 84 to 16 per cent when a benefits question is asked in plain English, federal AI obligations reaching $7.2 billion, 300,000 hours saved by bots at the Defense Logistics Agency, and Brazil's open platform for reusable public-sector AI.
AI trials involve governments less often than trials in general
Five researchers tracked AI interventions across two trial registries from 2019 into a partial 2026. AI is on track to make up one in five registered trials, up from one in thirty in under four years, the average planned AI trial runs 11.3 months, and AI trials in the AEA registry are government-related 10.7 per cent of the time against 14.1 per cent for trials overall.
Assorted links for 7 September 2026
Seven links: a European Commission expert review on AI's divides across member states, infrastructure for governing AI agents, a paediatric chest X-ray model whose sensitivity falls from 95.1 to 6.2 per cent across countries, smartphone oral cancer screening in India, AI policy priorities across Africa, a benchmark for simulating governance policy, and Nigeria's sovereign AI agenda at GITEX.
Access to an AI assistant built on World Bank reports saved no significant time
Researchers deployed an AI assistant restricted to a curated library of World Bank reports to more than 2,200 professionals across 116 countries. Its refusal rate fell from between 40 and 70 percent in its early weeks to under 10 percent once the library grew from about 50 reports to over 4,000, and the broader experiment found no significant time savings for the overall group given access.
Assorted links for 5 September 2026
Nine links: safety frameworks that miss dialects and mistranslate a diagnosis, an argument about the words the field uses, a scenario for China to 2036, agent design against human oversight, what 33 parliaments say about AI and work, a New York audit, Austrian plans to automate official notices, a Colombian official on digitalised procedures, and how unevenly frontier AI exposure falls across economies.
Assorted links for 2 September 2026
Six links: six Indian digital public infrastructure systems with no forum for public participation, the IMF on cyber losses and financial stability, OpenAI pausing reinforcement learning for two weeks, where the world's data centre capacity sits, a review of 122 studies on holding language models to account, and a proposal to work out which rules apply before a system acts.
Brazil and Portugal account for about 69% of Portuguese dataset records
Twenty researchers built the first country-level atlas of who is represented in the datasets behind language models. Brazil and Portugal hold about 69% of the Portuguese records, France, Switzerland and Canada about 63% of the French, and 121 of 197 countries have ten or fewer records attributed to them at all.
Assorted links for 1 September 2026
Five links: four scenarios for an AI economy, the 10 trillion won South Korea put behind a free public AI service, where the largest open models are built, what six months of use does to a working pattern, and what workshop participants feared disclosure would cost them.
Half these jurisdictions have no AI law of their own
Five researchers pulled 382 provisions out of 101 legal instruments across twenty jurisdictions that build no frontier AI model. Only 22 per cent sit in binding law, half the jurisdictions hold no domestic instrument of their own, and two provisions in the whole sample reach the party that built the model.
Assorted links for 29 August 2026
Nine links: which of England's two planning AI tools is actually live, what MIT's committee found had changed on campus in under three years, an Ebola nowcast in the Democratic Republic of the Congo, and the first evaluation of a proprietary model by someone who never saw its weights.
Five of 475 studies on surgical AI in poorer countries came from low-income countries
A scoping review gathered 475 studies of artificial intelligence in surgical care across low-income and middle-income countries. 305 came from China, five came from low-income countries, and 12 reported reaching a real patient care pathway.
Assorted links for 27 August 2026
Eleven links: what METR found inside OpenAI's agent incident, what an IMF paper puts on the table for African growth, where India sits per head rather than in total, and a Brazilian committee text that requires human supervision and authorises facial recognition in the same breath.
Brazil found AI in 94% of its judicial bodies and 55% of the branch that runs the services
The same survey asked where the technology sits, and at every level of government it is about twice as likely to be working on the administration's own processes as on anything delivered to a citizen.
Kenya's draft AI policy promises a pay benchmark for data work indexed to rates abroad
A row in the implementation matrix commits to a fair-pay reference framework for data annotation and content moderation, calibrated to international rates. The document gives no rate and does not say whose. What is scheduled first is a publication, and publishing it needs nobody else's signature.
Vietnam's AI scoring sheet puts sixty-five of a hundred marks on doing the work
Vietnam's criteria for public-service AI platforms, in force since 19 June, mark a platform out of 100. Thirty-five marks are for compliance and sixty-five are labelled for the sector's own work. What those marks reward is not published, so the split between the two blocks is all the ministry has disclosed.
A register of government algorithms can only hold what the state bought
Almost every country now has an AI policy, so that count has stopped telling anyone apart. The measure that has not saturated is whether a government must disclose the algorithms it runs itself, and procurement caps how much of it a register could ever reach.
A rule written for one market became a property of the tool everyone else uses
Brussels wrote AI transparency rules for its own market. The mark now travels with the tool everywhere, and the people it reaches had no part in either decision. The objection is jurisdictional, not to the rule.
Indonesia's welfare redesign lets people dispute what the state's records say they own
In one Indonesian district, more than 9,000 people used a new appeals channel to correct what the government's records said about them. That number counts errors found, not yet errors fixed.
Counting where migrants work moves Tajikistan from the 21st to the 75th percentile of AI exposure
A new country-level measure follows AI exposure through migrant work and remittance income. Tajikistan shows why a national labour market does not stop at the border.
Claude's content mark shows that a model touched a text, not whether it supplied the argument
Claude's new mark can show that a model touched a text. It cannot tell whether the model supplied the argument, or only the English through which the argument had to travel.
African developers chose Chinese models before any government signed anything
Ten African countries signed an AI pact with China in Shanghai in July. The adoption it looks like it caused had already happened, in the hands of developers who could download the weights while the American onboarding queue was still running.
OpenAI published two datasets on ChatGPT use and neither covers the public sector
OpenAI publishes two datasets on how ChatGPT is used, one for individuals and one for enterprises, and the public sector that it counts in the millions when selling appears in neither.
An AI exposure estimate was checked against vacancy data, and it moved the same way
AI exposure scores are hypotheses, and their value lies in predicting correctly. Here is one that was checked against vacancy data, and moved the same way.
The World Development Report 2026 is out
The World Bank's flagship on artificial intelligence has published. One of its numbers, what the nine chapters cover, and what I will be writing about here.
China's 29-country AI alliance, and what it is buying
China closed the World AI Conference in Shanghai by launching a 29-country AI alliance. A few weeks earlier, at the G7, Anthropic and Google DeepMind called for an American-led coalition.
What closing a digital governance programme in Serbia actually showed
This week, World Bank leadership flew to Belgrade for the closure meeting of Serbia's Enabling Digital Governance (EDGE) project.
A place where every claim about AI and work points somewhere
Johanna Einsiedler just launched what the argument about AI and work has been missing: a place where every claim points back to a number on a page.
The hard part of AI infrastructure is not the hardware
In two hours I am joining the MCDF workshop on AI Infrastructure to make a case that often gets skipped: the hard part of the Agentic State is not the models, it is the plumbing underneath.
Whose data trains the model, and who benefits from it
As one X user bluntly summarized: 'Claude is winning because rich people are providing the training data. Poor Meta has to train AI with the peasants of the internet.' Crude, and not quite right about the mechanism.
Theoretical AI capability against observed usage
This figure from Anthropic comparing “theoretical AI capability” with observed usage across occupations has been circulating widely in the AI policy bubble.
ChatGPT is splitting into work and everything else
OpenAI recently released data on how people use ChatGPT. One number stood out. In December 2025, about 75% of messages from paid Pro users were work-related.
A technical milestone for the region, and the harder question underneath
A regional public-AI launch is a real technical achievement, and 'sovereign' is doing a lot of unexamined work in how it is described. Ownership, hosting, control and capability are four different claims.
What development research keeps missing about AI
A few thoughts as I join the Global Development Conference 2025 in Clermont-Ferrand, promoted by the Global Development Network.
The inference divide: the inequality no one is talking about
Two people with identical devices, connections and skills, using the same model, can receive fundamentally different intelligence augmentation. One subscription lets the model think for minutes. The other gets seconds.
Where generative AI is actually hitting labour markets
Most studies of generative AI and jobs rest on exposure estimates rather than observed effects. New World Bank research asks where the impact on labour markets is actually landing.
Unwritten 2025
'We don't know where this technology is going' sounds thoughtful and feels responsible. Increasingly I am convinced it is neither, and that waiting is a decision that may cost us.
The overlooked upside of AI for developing nations
While most discussions about AI and developing nations fixate on risks and challenges, we often overlook the glaringly obvious opportunities.
AI's coming data saturation is an opportunity for the countries left out of the corpus
This recent Nature article projecting AI data saturation in the near future inadvertently highlights a significant opportunity for developing economies.
Inclusive AI infrastructure, and who is in the room when it is designed
Still energized from moderating this panel at the Tallinn Digital Summit on inclusive AI infrastructure - born from a growing collaboration between Estonia’s Government and The World Bank Group.
The link between open data and trust in government is weaker than we assumed
Two rounds of survey evidence suggesting the link between open data and institutional trust is weaker than open-government advocates assume, and in places runs the other way.
The hidden risks of AI: how linguistic diversity can make or break collective intelligence
Diverse groups solve problems better. Models trained mostly on English inherit a narrower collective intelligence, and a subtler digital divide follows.