Government was the busiest sector in September's record, and several governments set limits on their own staff's use of AI.
Published 26 September 2026
September in the record
395 findings from 110 countries.
229
Deployments
74
Evidence of adoption
70
Regulations
22
Evidence of impact
From the editors
Government was the busiest sector in the record this month: 175 of the 395 findings, from 82 countries. Of those 175, 93 were public bodies putting AI to work.
The record also shows governments limiting their own staff. Argentina's Attorney General barred prosecutors from putting live case files and personal data into public AI tools such as ChatGPT. Pakistan barred officials from feeding those tools classified papers, work email, code or citizen records. Costa Rica's Institute on Drugs went furthest: it banned AI outright for its own data, revoking a 2025 directive that had allowed regulated use. If you work in a public body, find out whether yours has a rule like these, and what it covers.
Most of September's record is launches: 229 of 395 findings announce that something was switched on. Only 22 report what happened after. Organisations publish their launches and rarely publish their results. So if someone brings you an AI proposal, ask them what they will measure, and when.
Three of those results are the reporting body's own figures: Ghana's revenue authority, El Salvador's education ministry and a Taiwanese city's land bureau. We publish them and say whose figure each is, because a self-reported number is still evidence. Denmark's is independent: its central bank measured what firms did, from national records.
Three of the ten below record AI going wrong: an AI agent used to attack an organisation's systems, a court striking a brief for invented citations, and a survey in which workers report spending more of their AI time fixing errors than doing the work.
This month, three stories from the August dispatch moved on. They come first below, then the ten new findings, each with its source. The other 385 are in the record.
— The Industrial Research UnitStraits Institute for Applied AI
In the United Arab Emirates, Cleveland Clinic Abu Dhabi now runs Transpara, the AI software that trial used, to mark suspicious areas on mammograms before clinicians review them.
Rentosertib's Phase 3 trial enrolled its first patient on 9 September. It aims to enrol 320 people with idiopathic pulmonary fibrosis and treat them for 52 weeks against a placebo. Phase 3 is the last stage of trials before a drug can go to regulators for approval.
In the United States, the Department of Veterans Affairs has opened Abridge's AI scribe to clinicians at every VA medical centre, after a pilot in 75. Abridge is a different product from the one in August's story, and the report does not say how clinicians check the notes it drafts.
Chosen from all 395 because each records a result, a new rule, a first, a failure or a big organisation’s own use of AI.
NewEvidence of impactGovernment & Public SectorGhana
Ghana says AI valuation lifted customs revenue by up to US$100m a month
Monthly customs revenue went from roughly US$350 million to as much as US$450 million after AI valuation switched on in April, Commissioner-General Anthony Kwasi Sarpong told the West African Tax Administrators Forum. The authority's own figure; no independent evaluation has isolated the software from the reforms alongside it.
In Taiwan, Tainan City's Land Administration Bureau has run AI checks on 142,758 property-price declarations since it introduced the system. The software raised 2,558 warnings, and staff confirmed 74 needed correcting, such as a price typed with a digit too many or too few. The bureau says the AI identifies risk and its staff make the final call; the figures are its own.
Firms that adopt AI hire fewer people, Denmark's central bank finds
Danmarks Nationalbank compared firms that took up AI with similar firms that did not, using Danish administrative records. The adopters went on to recruit fewer people, most of all younger and better-educated candidates. Pay has not moved, and the bank expects no fall in national employment soon.
In force from 1 September, it governs development, deployment and use, and separates sovereign models — built, trained and stored entirely in Russia — from national ones that may carry foreign open-source components. A model enters either class only once the government confirms it fits.
NewEvidence of impactEducation & TrainingEl Salvador
El Salvador says its AI pilot schools scored level with Germany and Sweden
The education ministry says 171 pilot public schools outscored the national average in science, mathematics and reading, level with Germany and Sweden, crediting AI tutoring, teacher training and learning monitoring. OECD's PISA for Schools assessed 1,198 students in June; the schools were picked for internet access rather than past results, so they do not stand for the national system, and no independent evaluation has isolated the tutors' effect.
Kenya's central bank now requires approval before AI decides a loan
Lenders must get the Central Bank of Kenya's sign-off before a model settles a loan decision. The bank's own 2025 count put half of licensed institutions on AI, with one tier-one lender pushing Sh1.5 billion a day through machine-learning mobile lending.
NewEvidence of adoptionGovernment & Public SectorSaudi Arabia
Saudi Arabia's AI skills programme passed 1.2 million completions
Completions under the Samai programme passed 1.2 million against a target of one million, the head of the Saudi Data and Artificial Intelligence Authority said as the UNESCO ethics forum opened in Riyadh.
NewEvidence of impactLegal & Professional ServicesUnited States
A US appeals court threw out a brief over four invented citations
The District of Columbia Court of Appeals struck a Deutsche Bank unit's entire appellate brief in a foreclosure case and referred the matter to disciplinary counsel. The filer said Google's generative tool produced the citations; the panel held every firm lawyer who signed the brief shares responsibility.
Spain's data regulator logged its first breach carried out by an AI agent
Spain's data protection agency logged an autonomous attacker, built on a widely known language model, that signed in and explored an application unaided until exploitable flaws surfaced, then altered personal records and opened invoices. Neither the victim nor the model is named, and the account rests on the affected organisation's own report.
NewEvidence of impactBusiness OperationsUnited States
US workers say more of their AI time goes on fixing errors than on the work
US salaried workers put 42 per cent of their AI time into troubleshooting errors and reworking prompts, against 35 per cent on work that moves their job forward, a survey of more than 1,600 by HR software vendor BambooHR finds. Average use is 87 minutes a day, and the vendor puts the troubleshooting at roughly 20 workdays a year.
Industrial Research Unit, Straits Institute for Applied AI. (2026). The State of Applied AI — September 2026 Dispatch. https://www.straitsai.institute/state-of-applied-ai/september-2026
Ghana
Taiwan
Denmark
Russia
El Salvador
Kenya
Saudi Arabia
Spain