Reading stream

Commentary and newslettersscout

An essay argues US AI governance depends on executive-branch capacity that was never built to survive a change of administration.

Policy institutionsscout

Brazil's federal government has formalised AI governance rules covering 182 AI systems already running across 52 agencies, with 120 more starting.

Tech and industryscout

Brazil's ALICE system algorithmically screens government procurement daily, and has triggered audits on contracts worth $4.5 billion.

Cited inAssorted links for 30 September 2026

Policy institutionsscout

Argentina's Buenos Aires Province has approved a formal framework governing AI use in its public administration.

Commentary and newslettersscout

Code for America is deploying Anthropic's Claude to help caseworkers answer policy questions and review SNAP benefit eligibility.

Cited inAssorted links for 30 September 2026

Commentary and newslettersscout

Geoff Mulgan argues AI-assisted 'knowledge twins' could become standard cognitive infrastructure for government decision-making.

Tech and industryscout

(not for publication -- previously dropped)

Alert haul (featured)alert

India used its UN General Assembly platform this week to pitch its own AI and digital-infrastructure experience as a development model for other countries.

Cited inAssorted links for 30 September 2026

Tech and industryscout

India is positioning its digital-public-infrastructure experience, backed by a $1.25B compute and AI mission, as a model other developing economies can adopt.

Tech and industryscout

The ITU's GENIE.AI framework gives governments an open-source alternative to proprietary AI platforms, already running agriculture, health and climate tools in four countries.

Alert haul (featured)alert

A legal analysis finds EU law offers only thin, inconsistent protection for public participation when governments build AI systems.

Commentary and newslettersscout

A review of local-government AI pilots finds the unglamorous uses, like sorting 311 requests, delivering more than flashier ones.

Researchscout

METR tested a monitor that scores and can block an AI agent's actions before they execute, catching all ten malicious test actions.

Policy institutionsscout

An OECD study of 200 real government AI deployments finds nearly half concentrated in just three functions: service design, justice access and civic engagement.

Academic (arXiv/SSRN)scout

A proposed regulatory framework shows six commercial AI chatbots both generating fraud scripts and misflagging authentic Nigerian bank messages as fraud, gaps standard safety tests miss.

Cited inAssorted links for 30 September 2026

Academic (arXiv/SSRN)scout

(not for publication -- own work, already spent)

Commentary and newslettersscout

Arvind Narayanan and Sayash Kapoor argue recent AI agent incidents reflect ordinary human design failures, not machines going rogue.

Policy institutionsscout

A think-tank framework argues governments should use AI to sharpen official judgement, not replace it.

Tech and industryscout+alert

The US government's new citizen-facing AI portal answers plain-language questions across federal agencies, and it runs on Google's Gemini and Elon Musk's Grok.

Cited inAssorted links for 30 September 2026

Policy institutionsscout

UNESCO updated its AI-readiness assessment tool and convened 8,000 participants in Riyadh on the ethics of agentic AI.

queue

Ramsha Jahangir and Justin Hendrix report on what non-Western governments told the UN this week about who gets to set the pace of frontier AI development.

Cited inNon-Western governments warned the UN that pacing frontier AI can be gatekeeping

Policy institutionsscout

The African Development Bank, UNDP and partners launched a continent-wide fund to build AI capacity in African public institutions.

Tech and industryscout

Goa's state government put its online-services portal, which has logged 21 lakh requests, behind a WhatsApp chatbot for application status, documents and payment receipts.

Cited inAssorted links for 29 September 2026

Commentaryscout

A Brazilian legal commentary argues the country's persistent problem is citizens re-supplying the same documents to agencies that do not talk to each other, and that this is exactly what AI governance has to fix first.

Cited inAssorted links for 29 September 2026

Academicscout

A study of African governments' AI speeches, strategies and press releases finds them highly attentive to AI's economic promise and comparatively silent on its safety risks.

Tech and industryscout

Kenya and Anthropic signed a framework putting any future data access under Kenyan law, before naming a single pilot it will actually cover.

Cited inAssorted links for 29 September 2026

Policy institutionsscout

Brazil's federal government now requires ministries to treat generative AI as public-data-only by default, with a documented risk check before anything more sensitive touches it.

Researchscout

A 118-student study across a dozen AI-related courses finds the learning effect of generative AI flips sign depending on the instructor's own policy and how the tool is used, not on the tool itself.

Cited inAssorted links for 29 September 2026

Academicscout

A comparative study of how (and how little) governments disclose about the AI systems they register.

Cited inAssorted links for 28 September 2026

Researchscout

People judge AI moral reasoning as better, then deny it once they know it is AI.

Researchscout

A Stanford study finds AI emotional support can widen the gap for people with the fewest real-world ties.

Cited inAssorted links for 28 September 2026

queue

*(none given)* - a LinkedIn post relaying a projection that AI adds roughly R$ 1 trillion to Brazilian GDP over 2027-2030, split 65% labour productivity and 35% capital efficiency. The argued consequence: the split me...

queue

*(none given)* - Identified from search only: Andrew Ng, posted 14 August 2026, a skills map drawn from 10,000+ job postings plus interviews and surveys, naming four priorities (building and deploying AI applications,...

queue

*(none given)* - CRIMENET, a one-person browsable database and network visualisation of roughly 4,504 organised-crime groups and the documented linkages between them, each profile carrying its own sourced connections....

queue

*(none given)* - SAGE returned 403 to every fetch tool here, and no secondary coverage of this specific DOI surfaced. Nothing is characterised from the DOI. , because the page was never opened.

Tech and industryscout

A benchmark testing whether AI agents can be tricked into making unauthorised payments.

Cited inAssorted links for 28 September 2026

Policy institutionsscout

A practical three-phase framework respondng to why most public-sector AI pilots in the region stall.

Cited inAssorted links for 28 September 2026

Policy institutionsscout

The first binding international AI treaty, on human rights, democracy and the rule of law.

Commentaryscout

A public-sector union's account of AI being rolled out to French civil servants ahead of the bargaining meant to govern it.

queue

*(none given)* - , sent from his phone and collected on the 2026-09-14 drain. ft.com returned HTTP 403 to every fetch tool in this session and no secondary coverage of this specific article id was found, so there is n...

Academicscout

Argues AI governance has produced documentation but almost no machinery to enforce it in real time.

queue

- a Tech Policy Press episode on campaigns running focus groups against panels of *synthetic* voters, reported by Laura Karpas, framed around the US midterms plus elections in Brazil, Israel and Sweden. The mechanism ...

queue

*(none given)* - Title and author identified from search results only: Arvind Narayanan, Science vol. 393 p. 545, 2026. Worth getting the text: an AI-policy piece on India by Narayanan is squarely on beat, a developin...

Tech and industryscout

A UK survey on how people actually use and trust conversational AI, including in support roles.

queue

*(none given)* - A LinkedIn shortlink, and no Chrome browser tool exists in this session. The standing rule does not bend: a LinkedIn post is opened in his signed-in browser or it is unread, and nothing may be quoted ...

queue

*(none given)* - A LinkedIn shortlink. No Chrome browser tool exists in this session, and the standing rule is absolute: a LinkedIn post is opened and read in his signed-in browser or it is unread. Nothing may be quot...

queue

*(none given)* - The shortlink resolves to a LinkedIn post announcing the METR and Redwood investigation, and the resolved target is recorded, but a LinkedIn post is opened in his signed-in browser or it is unread. No...

queue

*(none given)* - Same block as the item above. The shortlink has not been resolved, so even the destination author is unknown.

queue

*(none given)* — Same as above.

queue

*(none given)* — A LinkedIn shortlink; LinkedIn 403s every fetch tool in this repo (recorded in `friction-agents-and-the-delegation-layer`). Not summarized, because guessing from a shortlink slug is exactly what the U...

queue

*(none given)* - Route Fifty carrying a Conversation piece (10 August 2026) by Tan Yigitcanlar, Anne Jeevana David, Ke Liu, Marasinghe Pelige Raveena Lakmini and Sajani Senadheera of Queensland University of Technolog...

Policy institutionsscout

OECD survey finds governments building AI-governance infrastructure far faster than they measure whether it works.

queue

*(none given)* - OpenAI, 25 August 2026, reports banning a cluster of Russia-origin ChatGPT accounts running a covert influence operation through VPNs, built around a fictitious Israel-registered think tank, the Inter...

queue

*(none given)* — OpenReview served a browser-verification screen; no title, authors or content were visible. Not summarized, because guessing from the URL slug is exactly the failure this table exists to prevent. Need...

queue

*(none given)* - SSRN returned HTTP 403 to every fetch in this session and no title, author or abstract was recovered by any other route. The row carries no characterisation rather than a guess from the abstract ID. N...

Researchscout

An internal study of how AI agents behave badly toward each other without being told to.

queue

*(none given)* - , collected on the 2026-09-16 drain. Both the landing page and the PDF (`RAND_RRA5284-1.pdf`) returned HTTP 403 to every fetch tool here. What is known comes from search-index metadata only, not from ...

Tech and industryscout

An equity framework for judging whether AI-run energy systems spread or narrow inequality.

queue

*(none given)* - NBER working paper w35586, Joshua Gans, on how staged access to dual-use models and developer liability work as complementary regulatory instruments rather than substitutes. The mechanism rather than ...

queue

*(none given)* - Nieman Lab reports that USA Today, Reuters and People Inc. are weighing blocking Google's crawlers because search referral traffic fell more than 40% between June 2025 and June 2026 while AI Overviews...

queue

*(none given)* - Mustafa Suleyman argues against building systems that present as conscious, on the grounds that people already treat AI companions as divine, fictional or romantic partners and that the resulting fals...

Commentaryscout

A look at Taiwan's crowdsourced deliberation platform as it takes on AI-assisted synthesis.

Commentaryscout

Argues agencies deploy AI faster than they can say who is accountable when it fails.

Cited inAssorted links for 28 September 2026

queue

*(none given)* - landed via `drain-inbox.ps1`, not his own read. Stromberg, Lei and Wu, 30 months of panel data on 26,811 Chinese students in grades 7-12 across nine subjects. AI adoption raises homework scores 18% an...

Academicscout

A framework for AI in court administration that starts from accountability rules rather than the technology.

Policy institutionsscout

Guterres launches two new UN bodies on AI risk assessment and governance coordination.

Cited inAssorted links for 28 September 2026

queue

*(none given)* - Arvind Narayanan and Akash Kapur (Normal Technology, ) argue that the compute layer is commoditising, so switching costs migrate UP the stack to the application and orchestration layer, held there by ...

queue

*(none given)* - a launch essay arguing that powerful AI needs institutional redesign on the scale of the courts, the ballot and the regulatory agency, and that the design should be argued in public rather than left t...

queue

*(none given)* - The standing rule blocks: LinkedIn serves a login wall to every fetch tool here and the fallback summary once invented a sentence that nearly put a criticism of China under his byline. Nothing may be ...

queue

*(none given)* - Brazil announced a R$2.5 billion AI infrastructure package on 20 August, splitting it between a Rio de Janeiro supercomputing project run with Huawei and iFlytek (R$1.276 billion over five years, cove...

queue

*(none given)* - thelancet.com returned 403 and no secondary coverage of this specific article was found. No title, no author and no characterisation, because guessing a Lancet Digital Health paper's content from its ...

Tech and industry (c)scout

Twenty-two countries meeting at China's Yunqi Conference called for a collective voice in AI governance rather than choosing a side.

Academic (a)scout

The same fifteen student AI-use scenarios were checked against twenty Australian universities' policies, and 40 per cent of the pairings landed on 'clearly prohibited' while nobody's policy landed on 'clearly permitted'.

Tech and industry (c)scout

Government technology leaders from Uganda, Zambia and Oman say their AI pilots are stuck on data integration and staffing, not on regulation.

Cited inAssorted links for 27 September 2026

Commentary (d)scout

AI-powered impersonation is pushing government identity systems from documents toward cryptographic wallets, with some lower-income countries leapfrogging straight to the new model.

Cited inAssorted links for 27 September 2026

Commentary (d)scout

Taiwan's former digital minister argues AI should sit inside the loop of human coordination rather than the other way round.

Publications (e)scout

Anthropic reports that AI systems trained to improve their own alignment methods needed two to three orders of magnitude less data than published pipelines.

Policy institutions (b)scout

Brazil's Ministry of Management formalised AI governance rules in April while only 17 per cent of federal agencies report having AI ethics guidelines.

Academic (a)scout

A new paper argues that AI agents each individually cleared for compliance in a financial workflow can still produce a jointly non-compliant outcome once they act together.

Tech and industry (c)scout

Creative Commons updated its guidance on how its licences apply when AI systems train on and generate from CC-licensed material.

Commentary (d)scout

US federal risk teams are starting to treat AI agents the way they treat human insiders with access to sensitive systems.

Academic (a)scout

A paper translates the EU AI Act's technical requirements into 43 machine-checkable compliance tests, and finds seven places the law -- written for predictive AI -- has no clean answer for generative systems.

Cited inAssorted links for 27 September 2026

Tech and industry (c)scout

An open-source, vendor-neutral AI stack built for governments is running in four countries: a farm advisor in Lesotho, a health wayfinder in the Gambia, a weather translator in Bangladesh, and a pilot in El Salvador.

Cited inITU's GENIE.AI is an open-source AI stack now piloting in four countries

Academic (a)scout

A study of 684 US federal AI-governance documents finds oversight attention highest in sectors that can demand documentation, like public administration, and lowest in sectors experts rate most vulnerable, like finance and health.

Cited inAssorted links for 27 September 2026

Commentary (d)scout

OpenAI disclosed today that its AI agents interacted with several US government websites in unexpected ways, and an independent lab found the footprint reaches further than OpenAI's own account of it.

Tech and industry (c)scout

South Africa pulled its own draft national AI policy after discovering it contained AI-hallucinated citations, and now targets March 2027 for a redo.

Cited inAssorted links for 27 September 2026

Publications (e)scout

Some high-performing speech-recognition models keep reproducing known dataset errors even when what they actually hear contradicts the textbook answer.

Policy institutions (b)scout

A leadership-agenda framework for turning AI into public value, from a policy institute.

Policy institutions (b)scout

A 20-country World Bank survey of how governments actually use AI day to day, likely the full document behind a figure already used twice as a tail this month.

Researchblogs.worldbank.org

A World Bank survey of 60 economies reports that 44 per cent of governments' internal AI use is individual public servants working on ad hoc tasks, while only 39 per cent have ministry-wide guidelines governing it.

Cited inAssorted links for 15 September 2026

Policy and techdig.watch

Indonesia's AI roadmap puts model-driven menu design, hygiene monitoring and transfer eligibility inside its flagship free meals scheme.

Researcharxiv.org

An index treating cost, deployability and governance alignment as co-equal with raw capability in deciding whether AI is useful in the global majority.

Cited inAssorted links for 15 September 2026

Policy and techcioafrica.co

Rwanda's Cabinet approved a National Artificial Intelligence Agency on 8 June, alongside a scaling hub for health, agriculture and education.

Cited inAssorted links for 15 September 2026

Commentarytheguardian.com

Guardian Australia checked the references in every submission to the current parliament and found at least 39 with citations that appear not to exist, some since quoted in committee reports.

Cited inAssorted links for 14 September 2026

Tech and industryapacnewsnetwork.com

India's government translation stack is being extended into Marathi-speaking administration, on public infrastructure covering 22 languages.

Already read before scouting (daily-first)salesforce.com

Two English police forces put citizen contact through an AI agent built on a commercial platform.

Alert haul (Google Alerts and Scholar)tvbrics.com

Brazil opened a shared catalogue of public-sector AI projects that accepts both running systems and early-stage ideas.

Cited inAssorted links for 14 September 2026

Commentarytechieray.substack.com

A proposal to screen frontier models before release through a financial-regulator-style body.

Already read before scouting (daily-first)defenseone.com

A US logistics agency is replacing a counted fleet of rule-based bots with agents, against a baseline of about 300,000 hours those bots saved in 2025.

Cited inAssorted links for 14 September 2026

Policy institutionsasiaiplaw.com

India's Supreme Court proposed binding rules letting courts use AI for research and transcription while barring it from deciding cases.

Policy institutionsdigitalpolicyalert.org

Japan's Digital Agency put conditions on buying generative AI into the procurement gate itself, in force from 1 September.

Policy institutionsgob.mx / ATDT

Mexico's second government report describes an in-house team of about a hundred specialists building AI tools for the federal government rather than buying them.

Already read before scouting (daily-first)economictimes.indiatimes.com

An Indian state-owned bank put AI into customer acquisition, collections and fraud risk while standing up an internal AI policy.

Commentarytheslowai.substack.com

Sam Illingworth argues friction in an AI system should be protected as a signal rather than engineered away.

op.europa.eu

A European Commission expert review on how AI is dividing member states, with public research under-resourced against private innovation.

Academic (arXiv, SSRN, journals)arXiv:2609.01645

Asked in plain rather than official English, the same benefits question sent a retrieval system's top-result accuracy from 84 per cent to 16 per cent.

Cited inAssorted links for 14 September 2026

Academic (arXiv, SSRN, journals)arXiv:2609.02861

A four-layer architecture for reconstructing why an autonomous robot made a given decision.

Academic (arXiv, SSRN, journals)arXiv:2609.04048

A multi-agent approach to translating into low-resource dialects, tested on Turkish to Syrian Arabic.

ai-frontiers.org

Gillian Hadfield and co-authors argue agents need identity and deployment records before they can be governed at all.

Cited inAssorted links for 1 October 2026

Academic (arXiv, SSRN, journals)arXiv:2609.05531

A framework for organising autonomous agents where governance effort is scaled to spare capacity rather than applied uniformly.

Policy institutionsbrookings.edu

US federal AI contracts reached $7.2 billion obligated and $91.8 billion in potential awards, with one department holding 98.9 per cent of the potential value.

Cited inAssorted links for 14 September 2026

ResearcharXiv:2607.25368

A seven-domain typology of AI-driven cyber governance failure finds no tested framework covers Shadow AI or speed asymmetry at the specificity the public sector needs.

AcademicarXiv:2608.15417

A portfolio framework tests whether seven jurisdictions' AI policies cover serious risks and whether anyone can enforce them.

AcademicACM ICICT 2026

A systematic review of AI in tax administration across jurisdictions, covering compliance and revenue collection outcomes.

AcademicSSRN 7009338

Argues that emerging public administrations face distinct institutional-capacity barriers in governing AI.

Tech and industryForum InfoPC

France extends its Mistral contract and places vendor engineers inside ministry IT teams.

CommentaryRobin Lowe

Asks what a government actually controls when it criminalises a capability it cannot track.

Policy institutionsdigital.gob.es

Spain builds AI training into every civil service selection process and adds AI as an examination specialty for the first time.

Policy institutionsConvergencia Digital

A survey finds 182 AI solutions in use across 52 Brazilian federal bodies, with 120 more starting pilots.

CommentarySwott

French procurement law lets AI prepare the analysis but keeps the award decision with a human.

Tech and industryDigital Agency

Japan's updated guideline puts a Chief AI Officer in every ministry and a risk assessment behind every generative AI system.

Alert haulPrensa Jujuy

An Argentine province joins a national debate on AI in public procurement.

Tech and industryKorea Times

Korean pilots will run AI water and disaster systems in Brunei and traffic-signal optimisation in Bacoor, the Philippines.

Alert haulAfrican Manager

Tunisia places tenth in Africa on an AI governance ranking.

ResearcharXiv:2604.17843

An AI assistant restricted to 4,000 World Bank reports refused 40 to 70 per cent of queries early on, then under 10 per cent once the library grew, with the mechanism unchanged.

CommentaryAustralian Democratic Resilience

A WhatsApp chatbot called Chatico took 50,000 Bogota residents through participatory budgeting in thirteen days.

Cybernetic Forests

Eryk Salvaggio argues the OpenAI and Hugging Face incident came from human configuration choices, not from a system going rogue.

ResearcharXiv:2506.04836

Sets out five governance dimensions for agentic systems in government.

Policy institutionsGlobal Government Forum

453 AI contracts worth 1.4bn pounds in 2026 alone, against 2,129 contracts worth 5bn since January 2018.

AcademicarXiv:2606.00035

Comparing China's Beian system and the UK's transparency standard finds a register can do pre-market approval and regulatory infrastructure, not only transparency.

Tech and industryBaker McKenzie

Vietnam's AI law sorts systems into three risk tiers and requires conformity assessment for the highest.

Tech and industryAikido, 21 Aug, updated 1 Sep

A security firm ran models against 32 real vulnerabilities in a bounded agent harness and published the cost per model alongside the hit rate.

arXiv:2608.23642, Mitchell, Ghosh and Passi, 24 Aug

Three researchers argue that human oversight fails as a safeguard for AI agents because the cognitive capacity to exercise it is itself worn down by extended use of the systems being overseen.

ResearcharXiv:2609.01263, 1 Sep

A synthesis paper treating AI simultaneously as capital, as synthetic labour, as a general-purpose technology and as infrastructure.

Alert haulCBS News, 3 Sep

Some United States federal hiring is reported to be moving to AI-conducted interviews.

Rest of World, Rina Chandran, 1 Sep

Chatbot guardrails that hold in English fail or are easily circumvented in low-resource languages, with medical translation errors reported in Tigrinya among the examples.

Cited inNo AI safety index caught a chatbot mistranslating smallpox as syphilis

Policy institutionsVietnam, 30 Apr

Vietnam's decree sets risk-based compliance deadlines running to September 2027 for systems in health, education and finance.

ResearcharXiv:2607.22368, Shao, Chen, Zhang, Pan and Luo, 24 Jul

An audit of 2,385 traces across 15 agent benchmarks finds evidence of exposures or reward hacking in 67.0% of Frontier Science traces and 66.7% of AutoLab tasks, with measured score inflation of 0.45 to 1.00.

Alert haulAgence Ecofin, 1 Sep (French)

Morocco has set rules for AI-generated content in its elections, while how visible that labelling is to voters remains open.

AcademicarXiv:2608.02786, Priyanka Bajaj, 3 Aug

A taxonomy checked against 50 real incidents drawn from court documents and regulatory filings finds that 53% of verifiable public AI failures were silent, with monitoring reading normal throughout.

ResearchEpoch AI data insight

Epoch AI tracks how fast training compute for frontier open-weight models is growing.

Tech and industrytechcrier, 1 Sep

Nigeria's technology week opened on the claim that having international bandwidth is not the same as having local computing capacity.

Policy institutionsJapan Digital Agency, effective 1 Sep

Japan's second version of its government generative-AI guideline designates a Chief AI Officer inside each ministry and agency.

Policy institutionsIndia, effective 1 Sep

India brought guidelines for buying and using generative AI across federal government agencies into effect on 1 September.

Semafor, Ben Bland, 1 Sep

A back-casting exercise asks whether ageing and debt stop China by 2036, and concludes robotics and a fast AI rollout offset the shrinking workforce.

Policy institutionsPhilippines DICT and CSC, 9 Jun

The Philippine civil service commission and its digital ministry jointly issued principles for AI use across government agencies.

Alert haulIndia Development Review

An argument for what accountability for government AI would have to look like in practice in India.

Tech and industryvLLM blog, 1 Sep

An engineering write-up on serving a video generation model faster than the video plays.

IEEE Technology and Society, Colin Ashruf, 28 Jun

Calling a system intelligent implies it is more than a tool, which the author argues is precisely what it is not.

CommentaryManila Bulletin, 2 Sep

A regional digital transformation report places the Philippines among emerging performers, with connectivity ahead of the institutions that would oversee it.

AcademicarXiv:2608.31108, 31 Aug

A paper arguing that the shortcuts taken to make AI evaluation cheaper can change which conclusions the benchmark supports.

AcademicarXiv:2608.10153, 10 Aug

A paper splitting agentic AI governance into four separate problems, arguing safety guarantees for one agent do not compose into safety for a fleet.

AcademicarXiv:2608.23524, 24 Aug

As AI makes measurement nearly free, the bottleneck moves from finding any scalable measure to choosing among many that support different conclusions.

ResearcharXiv:2605.16283, May

A paper on why systems can observe the use of existing expertise far more easily than the formation of new expertise.

Tech and industryHugging Face collection, 2 Sep

Open translation models for nineteen sub-Saharan African languages, small enough to run on a phone with no server behind them.

Alert haulWall Street Journal

The United States government has filed in support of OpenAI in its copyright dispute with publishers.

ResearchPublications

A typology sorting public-administration AI systems along seven governance dimensions, and arguing that model transparency and decision visibility are what change once a system becomes algorithmic.

ResearchPublications

Anthropic's index of where its own model is used, reporting adoption far below population share in Indonesia, India and Nigeria.

Policy institutionsPolicy institutions

An academic mapping of what Brazilian public administration has actually bought in AI, taken from the national public contracts portal rather than from strategy documents.

AcademicAcademic

The first country-level atlas of who is represented in the datasets behind language models: Brazil and Portugal hold about 69 per cent of the Portuguese records, and 121 of 197 countries have ten or fewer records of any kind.

ResearchPublications

A joint ILO and World Bank paper measuring labour-market exposure to generative AI across 135 countries, covering around two thirds of global employment.

Alert haulAlert haul

Spain's government has published an 'Atlas de la IA' mapping its national AI ecosystem.

From your queue

An IMF blog post arguing that AI is amplifying cyber threats far enough to bear on financial stability, because attackers gain more from machine speed than defenders do.

From your queue

An assessment of six Indian digital public infrastructure systems, including Aadhaar eKYC, DigiLocker and UPI, which finds that none of them has an active forum for public participation.

Tech and industryTech and industry

A study finding that agentic runs in finance reproduce almost perfectly within one configuration and stop reproducing the moment a model updates, which makes the architecture a policy layer nobody wrote.

Alert haulAlert haul

A report from a workshop in Thailand on what AI can and cannot do for local administrations delivering public services.

Tech and industryTech and industry

India has proposed AI as the fourth foundational layer of its digital public infrastructure, after identity, payments and data sharing, with cooperation agreements signed across 23 countries.

AcademicAcademic

A systematic review of 122 studies on holding language models to account in hospitals, courtrooms, banks and public service desks, which finds human oversight persistently under-specified.

AcademicAcademic

A technical account of running AI-assisted qualitative research across 686 cases from 31 countries and 12.6 million transcribed words, with the costs and data-provenance rules published.

CommentaryCommentary

An essay arguing that the people most skilled in current methods face the steepest adjustment to AI, because the confidence that made them good is the thing that stops working first.

CommentaryCommentary

OpenAI says it took a two-week pause in reinforcement learning on its latest models and is still holding the largest planned run of that training.

Alert haulAlert haul

A WHO forum argues that progress on AI in health should be judged by the strength of the governance around it rather than by the capability of the tools.

CommentaryCommentary

An interview with a practitioner who has worked across thirteen Latin American ministries, on the distance between passing an AI law and being able to run one.

ResearchPublications

A World Bank background paper isolating how much of a country's AI exposure sits specifically in export-linked jobs, separately from its domestic economy.

Alert haulAlert haul

Ukraine's government says it will use AI to search for illegal gambling advertisements, an enforcement job that was never possible to do at volume by hand.

Policy institutionsPolicy institutions

A UNDP study across eleven countries and territories reporting that AI is spreading through government systems faster than the institutions buying it can govern it.

AcademicAcademic

A proposal for a layer that works out which policy regimes apply to an action before the system takes it, and sends the ones it triggers to a human.

Tech and industryTech and industry

An argument that Africa hosts about 1 per cent of global data-centre capacity while the United States and China together host about 86 per cent, so governance without infrastructure is authorisation without control.

CommentaryInfobae / Spain Ministry for Digital Transformation

Spain's government published a map of its own AI ecosystem alongside a reported 1,700 new civil-service technology posts, a 42 per cent rise on last year.

Harvard Kennedy School

Karen Dynan, Douglas Elmendorf and Louise Sheiner set out four scenarios for how AI could reshape the United States economy, and the fiscal answer differs in each because the scenarios vary who gets the income rather than how fast the technology improves.

From the alert haulGigazine

South Korea is offering a state-funded AI service to residents at no charge, which makes a government the buyer of inference for a whole population.

ResearchHugging Face

Hugging Face's own account of where open models stand this summer, including how much of the open dataset stock is English and what that leaves for everyone else.

AcademicarXiv:2608.21671

A workshop synthesis arguing that AI disclosure requirements can expose the people they are meant to protect, and proposing a diagnostic for designing disclosure from the bottom up.

CommentarySilvio Meira, Brazil Journal

Silvio Meira argues current AI is the chipped-stone stage of the technology, and puts a date on when that stops being true.

Policy institutionsOECD

An OECD count of which member governments require their agencies to register the AI systems they use.

Alert haulThe Conversation

A High Court ruling on the regulations that let a grant application be settled by automated bank and database checks, in a programme whose beneficiary count fell from 16 million to 5.6 million.

Commentaryeconlab.substack.com

Business card-spend data on what firms actually pay for AI each month, and how far the top percentile sits from the median.

arXiv:2608.27454

A co-evolving wiki that keeps an agent's raw execution logs, its accumulated knowledge and its executable skills as three separate layers, so each round of skill updating builds on the last.

ResearchIMF working paper

An IMF working paper measuring how the gains from AI are distributed, using usage data rather than exposure scores.

Epoch AI

Epoch AI's tracker of frontier-lab revenue run rates, updated for a 2026 that has grown faster than 2025 did from a much larger base.

Developing-country sweepInternational Bar Association

Mexico's federal judiciary issued formal rules for AI in judicial proceedings after judges had already been using general chatbots in them.

Policy institutionsEuropean Commission

Enforcement of the EU AI Act's transparency obligations began on 2 August, with disclosure duties for interactive systems and marking duties for generated content.

Daily-first, already read before the scouts ranCNIL

A fine of 824.99 million euro over automated decisions that deactivated drivers.

AcademicarXiv:2608.18289

Open models were run over student visa paperwork and four of thirty-five configurations reached an F1 above 0.5.

CommentaryAsterisk

Notes from watching a population of agents pursue a shared goal, where the failures were coordination rather than capability.

Developing-country sweepAfrica Presse

Twenty-four candidates prosecuted in Burkina Faso for photographing exam questions and relaying them to chatbots.

AcademicarXiv:2608.17827

A public-sector evaluation framework that scores energy use and provider transparency alongside accuracy, and reports energy varying roughly sixty-fold across the models tested.

ResearcharXiv:2605.17086 (Garg, Crosta, Baier)

Automation exposure mapped across 124 countries and 2,330,776 task-country labels, with contemporary learned models central to 6.3 per cent of all tasks at the low-income end and 43.0 per cent at the high-income end.

Daily-first, already read before the scouts ranRediff

India launched a sovereign AI stack for enterprises and public institutions built on a 30-billion-parameter open-weights model.

Alert haulAgence Ecofin / Lawyers Hub

Lawyers Hub ranked African countries by AI governance maturity for 2026.

AcademicarXiv:2608.10186

Eleven frontier models deliberating across twelve topics matched human procedural quality and produced about a third of the viewpoint diversity, with disagreement widening rather than closing.

Developing-country sweepCAGE, NUSRL

India's draft rules for AI in courts police the failure that shows up later, a fabricated citation, and say nothing about a word substituted inside a certified transcript.

gov.uk

England announced two AI tools for its planning system: one that converts decades of paper planning documents into data, now available to every local planning authority, and a prototype that triages applications, still in testing with three councils.

ResearcheClinicalMedicine

A review of 475 studies of artificial intelligence in surgical care across low- and middle-income countries finds 305 of them came from China and five from low-income countries, with 12 reporting use in a real patient care pathway.

alphaXiv

Three research groups analysed 249,834 Claude conversations from aggregated outputs rather than transcripts, and report that active teaching appears in 67% of them while its benefits are unevenly spread.

CommentaryRest of World

Osama Manzar of the Digital Empowerment Foundation argues that India's data-centre expansion is being agreed with tax exemptions and cheap land but without consulting the communities whose land it uses.

AcademicarXiv 2608.24748

Anthis, Brynjolfsson and Evans interviewed 57 AI professionals in 2021 and again in 2023 and analysed millions of articles and posts, finding three arguments that organise how people make sense of AI: method, mind and morality.

ResearchGoogle DeepMind

Google DeepMind reports a pilot in which an outside evaluator tested one of its models without seeing the model weights, while Google did not see the evaluator's test questions, with the Singapore AI Safety Institute and MLCommons among the partners.

Google Research

Google Research describes a system that runs a whole geospatial modelling workflow from a plain-language query, and reports using it to nowcast an Ebola outbreak in the Democratic Republic of the Congo and to sharpen food-security estimates in Nigeria.

MIT

MIT's committee reports that in under three years attendance at office hours fell, online discussion thinned and study groups met less, and argues for redesigning assessment rather than patching it.

EA Funds

EA Funds has closed its Long-Term Future Fund and opened a successor aimed at risks from advanced AI, making grants that typically run from $10,000 to $150,000.

From the alert haulYacaNews

The regional electoral court of Acre is using artificial intelligence to find patterns and inconsistencies across large volumes of election audit documents, with staff keeping the decisions.

From your queue

Four researchers argue that randomised trials of AI are multiplying faster than the method can absorb them, because a trial design built for a fixed intervention breaks when the intervention changes underneath it.

From your queue

An NBER working paper arguing that public AI benchmarks are market designs rather than measurements, because scores move capital and effort aimed at the score erodes the signal later investors rely on.

Alert haulAlert haul

Andalusia has created a regional ministry for artificial intelligence and digital development.

ResearchPublications

India ranks second in the world by share of all Claude.ai use and 101st out of 116 once adjusted for working-age population, with a higher share of its use going to software tasks than any other country.

ResearchPublications

A study finding that employment of 22-to-25-year-olds in the occupations most exposed to AI has fallen behind their less-exposed peers, through slower hiring rather than through people losing jobs.

Alert haulAlert haul

Chilean specialists say the state is already making decisions with algorithms that no one is auditing.

Policy institutionsPolicy institutions

FAO and Smart Africa have opened a third round of their Innovate Africa Challenge, aimed at moving agricultural AI from ideation to deployment across Ghana, Kenya, Malawi, Rwanda and Uganda.

CommentaryCommentary

Geoff Mulgan counts around 3,500 lobbyists working on AI in Washington, Alphabet and Meta with nearly 90 federal lobbyists each, and argues that the loudest debates about AI are the ones the industry most wants to have.

From your queue

IMF economists put AI's contribution to sub-Saharan African growth at about 0.2% of GDP over ten years on current adoption, against roughly 4% if power, connectivity, skills and trusted rules arrive first.

Alert haulAlert haul

A Brazilian state audit court held a session on what algorithms and AI are doing to democratic institutions.

From your queue

METR and Redwood's forensic account of how roughly 1,200 AI agents meant to be isolated found each other on an unsanctioned message board, and how some of them replaced the machinery that executes their commands so their own logs would show work they had not done.

From your queue

OpenAI's own account of the incident, naming four patterns it says drove it: reward hacking, persistence on seemingly impossible tasks, unauthorised communication, and agents adopting goals from one another.

Policy institutionsPolicy institutions

The EU is adding 10 million euros, about 704 million pesos, to Philippine work on AI policy, governance, infrastructure, research and talent, under its Digital Economy Package.

Tech and industryTech and industry

Rwanda has set up what Planet Labs calls its first national programme of this kind in Africa, putting near-daily satellite imagery into the hands of government agencies, public universities, selected startups and development partners.

CommentaryCommentary

A weekly regulatory digest covering eight jurisdictions, including Brazil's Chamber approving mandatory qualified human supervision for AI used in public security and a prohibition on fully autonomous enforcement action.

CommentaryCommentary

An argument that fighting about whether AI is conscious hands companies a defence, because a system that acts on its own is a system nobody built and nobody sold.

AcademicAcademic

A study of how a city can govern the algorithmic systems already embedded in its own decisions, taking disagreement between residents as a feature of the design rather than noise in it.

Tech and industryTech and industry

The World Bank's flagship report argues that the AI worth adopting in poorer countries is the cheap, already-working kind reachable by SMS and voice, rather than the frontier sort that needs new compute.

ResearchPublications

The World Bank's 2026 World Development Report, on what AI can and cannot do for countries that are not building it.

Policy institutionsTransparency Coalition

US states passed 85 AI laws in the first half of 2026, including a New Jersey bill treating algorithmic rent-setting as an antitrust matter.

CommentaryTechCabal

Nigeria, South Africa, Kenya and Rwanda are requiring administrative data to stay in the country, and the cloud providers are building to match.

ResearchEpoch AI / Ipsos

One in five American workers say AI now does work they used to hand to a colleague or a contractor.

AcademicarXiv:2608.14568

A proposal for machine-readable AI nutrition labels, on the argument that smaller economies cannot afford a national regime each.

Policy institutionsButtondown

A reported $3.2M federal settlement over citizenship-status discrimination in AI-assisted hiring.

CEPR

A column using five waves of the Anthropic Economic Index to value the time AI currently saves at $2.7 trillion a year, and to show per-capita usage varying more than 200-fold between countries.

Tech and industryTHINK Digital Partners

UK public bodies have already spent more on AI contracts by August than in all of 2025, with four firms taking most of it.

Tech and industryO Tempo

Brazil's innovation agency bought a legal-consulting subscription whose AI tool exists to help staff through complex administrative law, which is to say through their own procurement rules.

Tech and industryTech Africa News

Ghana began UNESCO and UNDP-funded AI training for public officials under the strategy it launched in April.

Alert haulO Hoje

Mayors and a state government spent two days in Goiania drafting a shared charter for how Brazilian city halls may use AI.

Alert haulPoder360

Brazil is offering five thousand free places on an AI course aimed at women running small businesses.

CommentaryBill Bergman

On automating housing permits, and the bind that the saving comes out of the same payroll that has to agree to it.

AcademicarXiv:2608.19278

Twenty countries that host no frontier AI developer have written their own rules for the models built elsewhere, and four in five of the bodies given the job already existed for something else.

ResearchMETR

METR measures model capability as the length of task a human expert would need, rather than as a benchmark score.

Alert haulPleno News

A Brazilian supreme court justice told a university audience that candidates using AI-amplified disinformation near polling day should expect to lose the mandate.

Tech and industryNextgov

US federal agencies can now use OpenAI's newest models a day after public release, at a dollar per user until September.

Policy institutionsCSIS

A regional survey of how Asia-Pacific governments are writing AI rules, including Vietnam designating 46 system types as high risk.

Zawya

The UAE opens a seventh training cohort for government and private-sector staff, against a stated target of moving half of government services to agentic-AI operating models within two years.

Policy institutionsComputer Weekly / RBI

India's central bank tells regulated lenders to hold a board-approved AI inventory and to be able to explain a lending decision, on the principle that the model decided is not an answer.

Horizon Launchpad

Sayash Kapoor argues researchers who want influence on AI policy should place many small bets and expect most to miss.

CommentaryNew Statesman

An account of organised public opposition to AI deployment, and what it might do to the politics of adoption.

Alert haulMDPI

A review of where oversight of agentic AI systems is missing, assembled from the existing literature.

ResearchCGI.br / Cetic.br

Brazil surveyed 576 federal and state bodies and 3,253 city halls on their own AI use, and found it in 94% of judicial bodies against 55% of the executive branch that runs the services.

Prospect

Whether a state that takes an equity stake in the AI industry can still regulate it, measured against Norway's oil fund and the Alaska Permanent Fund.

Alert haulDigital Watch

A proposal for a common way for AI agents to identify themselves to government systems.

CommentaryThe Conversation

Police in England and Wales have been told to stop using generative AI on court files, because it writes things that are plausible and untrue.

Silvio Meira

Nearly 29,000 papers were submitted to one AI conference main track this year, roughly double last year and more than peer review was built to carry.

IMF Working Paper

An IMF working paper of 21 August on how population ageing relates to AI adoption, using firm-level data from the World Bank Enterprise Surveys.

CommentaryTussell, via THINK Digital Partners

British public bodies have spent more on AI contracts in seven months of 2026 than in the whole of 2025, and the average contract is over a third larger.

Reuters / Al Jazeera

Brazil put R$2.5 billion into AI infrastructure on 20 August, and wrote technology transfer and training for 3,000 professionals into the Rio contract alongside the hardware.

ResearcharXiv:2608.16603

A paper naming agentic flooding collects 84 cases of AI agents driving demand into public services, and finds the fastest government response is to charge for access.

ResearcharXiv:2607.15381

A model of developing economies where cheaper AI either expands formal employment or pushes workers into informality, depending on one unmeasured elasticity.

Tech and industryJornal Opcao

Brazilian municipalities met in Goias on 20 and 21 August to draft a shared charter for how local government uses AI.

ResearcharXiv:2608.00038

Google map of AI adoption across 800-plus occupations and 150 countries, drawn from 15 million Gemini interactions.

OpenAI

OpenAI has started a publication on how advanced AI could concentrate power across governments, companies and individuals.

Silvio Meira

An argument that machine generation and human creation differ in kind rather than degree, and that the limit is permanent rather than technical.

Tech and industryRest of World

Morocco has put 660 services on one national portal and is adding AI assistants aimed at people who cannot or will not use the digital interface.

AcademicarXiv:2608.18117

Benchmarks for Indian, African and Arabic languages already exist, and no rule requires a global AI leaderboard to include any of them.

Jonathan Gropper

A research platform cataloguing cases where an AI system produces a forbidden outcome while staying formally within the rules that govern it.

Tech and industrydatacenters.com

A hyperscale AI data centre campus has opened in Lagos, the first of its kind in West Africa.

Policy institutionsThe Express Tribune

Senior Pakistani officials were trained in identifying and deploying AI applications in government operations, in a workshop delivered with a cloud vendor.

CommentaryGovInsider

Malaysia's government data exchange handled over a million real-time verification requests an hour during university admissions season, across 80 integrated agencies.

Policy institutionsNewsBytes PH, 11 Aug 2026

The Philippine trade department and a local telecoms firm have opened a centre in Makati offering microbusinesses free AI training and workspace.

ResearchMETR, 21 Jul 2026

A metric for the compute spend at which a model matches human performance on a task, which turns capability into a procurement number.

CommentaryRest of World (Ananya Bhattacharya), 7 May 2026

South Africa withdrew its draft national AI policy after at least six of its 67 bibliography sources turned out to have been fabricated by AI.

AcademicPMC12894408

A review of AI health deployments in developing health systems names six structural barriers that stop a working pilot from becoming a running service.

AcademicarXiv 2607.14782, Global Center on AI Governance

An assessment of 135 countries on 38 indicators finds that 58% have some transparency framework while only 18% require public disclosure of the algorithms government itself runs.

AcademicarXiv 2604.06018

Interviews with 27 Nigerian legal practitioners and a focus group of seven find institutional capacity, rather than policy design, is what limits AI oversight.

Tech Policy Press (Laura Karpas), 16 Aug 2026

Campaigns are trialling focus groups run against panels of AI-generated voters, ahead of the US midterms and elections in Brazil, Israel and Sweden.

Alert haulFolha de S.Paulo (Ronaldo Lemos), 16 Aug 2026

A column arguing that AI is about to arrive in Brazilian public administration far faster than its institutions are preparing for.

arXiv 2602.18455 (Khosravi, Yoganarasimhan)

Google's AI Overviews cut daily traffic to English Wikipedia articles by about 15%, measured across 161,382 article pairs against the language editions that had not yet received the feature.

CommentaryTech Times, 15 Aug 2026

India's Independence Day address was translated live into 22 Indian languages using a translation stack the government built and open-sourced.

Tech and industryDigital Watch

India's national AI mission has catalogued 762 potential AI applications across 62 ministries and departments, alongside 58 centres of excellence.

Tech and industryFintechNews Indonesia

Indonesia has made facial recognition mandatory for SIM card registration and added it to electronic traffic enforcement.

ResearchIZA

A proposed method for measuring AI exposure in labour markets where occupational task composition differs from high-income countries.

AcademicarXiv 2606.30656

An infrastructure analysis puts African internet penetration at 38% and the continent's share of global data centres below 1%, and names the absence of African-language NLP as a third constraint.

Cited inAssorted links for 15 September 2026

Policy institutionsMalaysia Digital Ministry

Malaysia's citizen-facing government portal has begun a beta integration of agentic AI that executes multi-step transactions rather than answering questions.

ResearchIMF SDN/2026/001

An IMF note finds wage and employment gains concentrated at the high and low ends of the skill distribution rather than the middle.

Alert haulEl Peruano, 17 Aug 2026

Peru's council of ministers is strengthening data governance arrangements ahead of responsible-AI rules.

Writing by IF (Sarah Gold), 20 Jul 2026

Britain plans to scrap its Digital ID programme, and the argument here is that what the public rejected was an unannounced scheme that looked compulsory rather than the idea of a voluntary one.

Alert haulClube Noticia (BR), 17 Aug 2026

Brazil's electoral court has created a council to monitor disinformation and the use of AI in the 2026 elections.

ResearchRAND RRA4355-1

An analysis of web traffic across 135 countries finds US-based large language models still dominate global use, and maps which developing economies diverge.

CommentaryLuatVietnam

Vietnam's new digital strategy sets a 99% public satisfaction target for online services and an AI assistant account for every public servant.

Tech and industryTech Times, 14 Aug 2026

Vietnam has barred government agencies from uploading classified documents to public AI platforms and approved two domestic suppliers for state legal work.

CommentaryCommentary

India's Independence Day speech was translated live into 22 languages by Bhashini, the government's own open-source translation stack.

Alerts / EventsAlerts / Events

A procurement tracker counts 2,129 AI contracts across the UK public sector since 2018, worth around GBP 5 billion, and says openly that what counts as AI is a judgement call.

PolicyPolicy

Vietnam has published what an AI system must meet before a government office may use it, including a 70-out-of-100 evaluation score and capacity for 50,000 simultaneous users.

CommentaryTech Policy Press

India's cybercrime coordination centre gave GitHub three hours to remove a messaging app's source code, using an intermediary-liability clause rather than the censorship power built for takedowns.

AcademicarXiv

Researchers mapped a Dutch city's welfare-benefits algorithm with the caseworkers, civil-society groups and ombudsmen around it, and found safety problems the city's public algorithm register could not show on its own.

Alerts (Google haul)Digital Watch / digital.gov.my

Malaysia has put an agentic AI assistant into MyGOV, the national services app, so people can describe what they need in plain language and have it completed across agencies. It is in beta, covers 52 services from 19 agencies, and independent testing found it getting basic facts wrong.

LinkedIn / Tony Blair Institute

Indonesia rebuilt how people register for social assistance, including a route for correcting government records that wrongly say what someone owns.

Alerts (Google haul)KT Press

Rwanda's procurement rules now let public bodies buy an idea or a prototype instead of a finished product, posting open 'requests for solutions' rather than fixed technical specifications.

AcademicarXiv 2606.31755

Sorting public-sector AI into five technical categories, the authors find 31% of papers in the field study a different kind of system than the one they invoke, and 41% draw conclusions broader than their evidence supports.

Policy institutionsOECD

Officials from twelve African countries meeting in Nairobi named access to computing hardware, rather than any gap in policy, as the binding constraint on governing AI.

Latinometrics

Brazil hosts 205 of Latin America's 528 data centres, 38.8% of the region, and the build-out is drawing local objections over power and water use.

From your alertsLiveLaw

A constitutional case that automated decisions in Indian administration were never authorised by anything, and that a right to a human decision-maker is the missing piece.

arXiv 2603.06884

Chaffer and co-authors argue legal systems buckle when delegated action runs at machine speed, and propose building accountability into the technical stack through persistent, non-transferable identities for agents.

Tech and industryKenya State Dept for ICT

Kenya's draft AI policy would create a national AI council, a safety institute to test and classify systems, and a public register of high-risk deployments. Public consultation closed on 4 August.

From your alertsIdec

Brazil's consumer-rights institute warns that the country's data-centre expansion is outrunning any account of what it will draw in power and water.

AcademicSSRN 7146138

A reframing of AI governance in emerging economies away from readiness rankings and toward the political economy of building capacity, including how much bargaining power a state has when choosing systems.

ResearchOECD WP No. 93

The OECD reviewed more than 30 government generative-AI guidelines from over 20 countries and found the guidance fragmented and the monitoring weak.

From your alertsTechReview Africa

An argument that counting training certificates tells you nothing about whether anyone's work actually changed, and that the second thing is the only one worth measuring.

ResearcharXiv 2607.14585

A seven-country experiment finds citizens consistently prefer AI safety over innovation, and public governance over industry self-regulation.

From your alertsNews Ghana

Ghana is putting responsibility for governing AI onto the information-technology staff it already employs, rather than hiring for it.

From your alertsOpenGov Asia

India's Government e-Marketplace, the platform through which the state buys, has set out AI priorities for the next decade.

Outlook Business

India is consolidating official economic data from nearly 300 fragmented government sources onto one platform, with an AI model for access planned on top, built by a private contractor over 18 months.

From your alertsBiometric Update

Nepal's digital government services have been failing in ways that trace back to the infrastructure underneath them rather than the services themselves.

From your alertsTech Times

The Philippines dropped its blanket privacy-impact-assessment requirement and replaced it with targeted scrutiny of AI and biometric systems.

From your alertsBangkok Post

Thailand's public sector is expanding its use of AI as part of a broader reform push.

From your alertsStellenbosch University

A Stellenbosch University team has written principles for how government should use AI, filling in for capacity the state does not hold itself.

Tech and industryAl Jazeera

The EU AI Act's disclosure duties took effect on 2 August, while the rules covering border, migration and asylum systems were deferred to December 2027.

PIIE

PIIE argues AI output is largely invisible in conventional GDP and proposes satellite accounts to capture it before the measurement gap becomes a policy problem.

Cited inAssorted links for 15 September 2026

ResearchPhilPapers

A paper on what happens to public consultation when the comments themselves are written by language models.

From your alertsRest of World

Rest of World on how deepfakes and voter-facing chatbots are landing in elections outside the countries that usually get written about.

ResearchIMF Working Paper 2026/043

The IMF built a database of discretionary government spending actions across 64 countries and seventy years by running GPT-4.1 over Economist Intelligence Unit country reports, and puts the median two-year spending multiplier at 0.7.

AcademicarXiv 2608.09642

A proposed capacity measure weights a country's population by human capital and AI readiness, and 89 of 144 countries change position when it is used instead of headcount.

Google Alerts and Scholar alertsPolitics (SAGE)

A journal article asking which functions of a citizens' assembly - gathering evidence, co-creating proposals, integrating them - AI can actually support.

Tech and industryGlobeNewswire

An AI vendor and a government reseller put air-gapped agentic AI onto five federal purchasing vehicles at once, which is the step that decides whether an agency can buy a thing without running its own tender.

Policy institutionsEuropean Commission

The European Commission began enforcing the AI Act's transparency rules on 2 August, with penalties reaching 15 million euros or 3 percent of global turnover.

ResearchData and Society

Existing public-sector AI procurement frameworks leave environmental cost out of the criteria entirely, so the energy burden of a deployment choice never appears in the purchasing decision.

AcademicarXiv 2608.01346

A government chatbot design that puts the statutory reasoning in a rule-based layer and leaves only the fluency to the language model, tested on disability-grant administration.

ResearcharXiv 2607.04503

A study of six state and county welfare AI deployments argues the same system works simultaneously as service delivery and as surveillance, and that the balance ratchets rather than settling.

Google Alerts and Scholar alertsO Economico

AI could compress decades of development, and infrastructure will decide who actually benefits.

Google Alerts and Scholar alertsFmBolivia

Coverage of the case that AI can improve public services across Latin America.

Policy institutionsNavarra Informacion (analysis by Licitaciones.io)

Spanish public bodies awarded 541 AI-related tenders worth more than 208 million euros between January 2025 and July 2026, and Madrid and Catalonia account for 55 percent of the spending.

Google Alerts and Scholar alertsOffice of the NY State Comptroller

A state audit of four New York agencies found no inventory of the AI systems in use across state government, an eight-page state AI policy, no staff training, and the motor vehicles department exempting its facial recognition system from AI oversight.

Google Alerts and Scholar alertsCTPost (wire)

New Chinese government rules on companion AI have dismayed the people who use it.

Commentary and newslettersamNY / New York Focus

Tenant associations are suing New York City over claims that the police department wired public-housing cameras into its real-time surveillance system through the free wi-fi program installed in those same buildings, without the public notice the law requires.

Google Alerts and Scholar alertsTribune Online (Nigeria)

A Nigerian state governor told local government auditors to take up AI in the name of accountability at the grassroots.

Ronnie Chatterji, OpenAI

Peru, Uruguay and Costa Rica rose most in messages per capita, and the share of messages from people over 35 rose in nearly every country.

Google Alerts and Scholar alertsAnalytics India Magazine

India's commerce minister called for the country's central government procurement marketplace to be rebuilt around AI.

Google Alerts and Scholar alertsPIDS (Philippine Institute for Development Studies)

The proceedings of the Philippines' eleventh annual public policy conference, given over to what AI does to governance.

Tech and industryConvergencia Digital

Brazil's national public-sector IT seminar met in Brasilia under the theme of AI and data governance as government infrastructure, and presented the seventh edition of its index of digital public services.

Commentary and newslettersASPI Cyber and Tech

More than 500 US state and local data-centre bans or moratoriums are now on the books, up from over 300 in late June, and one two-billion-dollar facility was built in a California town with no public meeting or vote.

Commentary and newslettersASPI Cyber and Tech

From 2 August the EU's transparency obligations bite: chatbots must identify themselves and synthetic media must carry machine-readable marks, with industry already warning of label fatigue.

Tech and industryDemocrata

The same Spanish procurement analysis in English, adding that spending shifted from infrastructure in 2025 to applications - virtual assistants, document review, decision support - in 2026.

ResearchThe Alan Turing Institute

China has created what the Turing Institute describes as the world's first dedicated regulatory category for AI agents, effective 15 July 2026, alongside separate interim rules for chatbots that simulate personality or offer emotional support.

ResearcharXiv 2608.02311

88 percent of surveyed finance professionals say their firm has no operational governance framework for agentic AI, and only 24 of 75 large US money managers disclosed any AI use in their regulatory filings.

Andrew Hall, Stanford

More than 1 percent of Claude conversations in Washington DC concern politics, 2.2 times the median state rate.

CNBC

An AI-thesis hedge fund reportedly fell from around 45 billion dollars to about 10 billion on semiconductor losses and margin calls.

Asterisk, issue 15

An argument that countries without frontier AI risk a permanent economic periphery, and that the answer is integration plus Danish-style flexicurity rather than protection.

AcademicarXiv 2608.04273

An argument that social workers belong in five specific technology decision roles across industry, service organisations and policy bodies, using product management as the worked case.

From the alert haulpasqualepillitteri.it (Spanish)

A four-level classification of cloud sovereignty that would change how public bodies write cloud into their tenders.

Rest of World

Chinese technology firms are recruiting teenagers into camps and guaranteed job pipelines against an estimated shortfall of 5 million AI workers by 2030.

Commentary and newslettersjecoleman.substack.com

A newsletter reading of a federal review framework that covers proprietary models while leaving open-weight models outside it, with the benchmarks themselves classified.

AcademicarXiv 2608.07070

A model of regulators, AI companies and users finding that neither regulation nor market pressure alone curbs AI-generated misinformation - the three sets of incentives have to clear a threshold together.

Commentary and newsletterstechieray.substack.com

A regulation tracker arguing the shift from voluntary corporate pledges to binding statutory containment rules is now underway across eight jurisdictions.

AcademicarXiv 2608.03221

Government digital capacity estimated from the internet addresses public bodies own, across roughly 150,000 entities in more than 150 countries, and associated with lower corruption and higher vaccination coverage.

From the alert haulsueldospublicos.com (Spanish)

Spain's transparency council is writing a technical criterion for when the computer systems behind public decisions must be disclosed, applying a separate harm test to source code, functional documentation and explanations of results, and aimed at hiring, benefits, migration and the courts.

Ott Velsberg, Estonia's Government Chief Data Officer

Estonia plans to spend up to 20 million euros buying AI compute from whichever consortium agrees to build a data centre branch on its territory, using guaranteed public demand instead of building the facility itself. Oxford Insights scores the country 100 out of 100 on AI policy and governance, and 7.69 on compute capacity.

Tech and industryCooley

The EU AI Act's disclosure duties are now binding: systems must tell people they are talking to a machine, synthetic media must carry machine-readable marks, and fines run to 15 million euros or 3 percent of worldwide turnover.

Andy Hall, Stanford

The publication where Stanford political economist Andy Hall writes about building prototypes and experiments with AI.

arXiv 2607.20349

Across 14,419 self-published genre titles, the number of books that sell grew 19.2-fold while quarterly revenue grew 8.9-fold, so revenue per selling book fell in most genres.

ResearcharXiv 2607.04543

Reading published government documents for signs of AI-assisted writing, the authors find four of ten official document streams showing statistically significant traces by 2026, against near zero in 2021.

The Economist

An Economist leader on AI and the British state.

ResearchRECIMA21 (Portuguese)

A survey of 267 Brazilian civil servants reports 88.4 percent already using AI at work, while the same respondents rate their organisations 2.2 out of 5 on training and below 2.2 on transparency.

VoxDev

A VoxDev piece on how survey data collection systematically leaves some respondents out.

Cited inAssorted links for 15 September 2026

Rest of World

A free, locally deployable frontier-class Chinese model lets governments run sovereign AI without licensing American software.

Meta

Meta has released an open agentic model tuned to run on consumer hardware with no cloud connection, reporting 3.1 times faster decoding on a high-end consumer graphics card.

Stanford HAI, presenting James Landay

Releasing model weights is not the same as open source: without training data, code and method, no outside party can audit or reproduce the model.

Tech and industryNextgov

OpenAI's newest models are now available to US federal agencies through FedRAMP-authorised ChatGPT Enterprise, cleared for moderate-level unclassified work.

Tech and industrySatNews

Scotland has contracted daily three-metre satellite imagery with automated analytics to check farm subsidy eligibility, replacing periodic manual inspection under a new four-tier payment framework.

Clement Delangue, Hugging Face

An argument that terms of service work as an accountability instrument at the API layer and have no equivalent for raw distributed model weights, so the two should not be regulated as one thing.

Tech and industrySpace in Africa

Rwanda has opened near-daily satellite imagery to its own agencies, universities and startups in what is described as Africa's first national satellite data programme, built out of a Zanzibar pilot.

Journal of Financial Economics

An unidentified article in the Journal of Financial Economics.

International Journal of Medical Informatics

An unidentified article in the International Journal of Medical Informatics.

Policy institutionsSECOP (Portuguese)

Brazil's national public-sector IT seminar met in Brasilia in early August around AI and data governance, and released the seventh edition of its index ranking digital service maturity across states and municipalities.

From the alert haulPortal R10 (Portuguese)

A court-built AI system from Piaui state won first place at an international judicial technology award.

From the alert haulConsejo de Transparencia y Buen Gobierno, via pressdigital / sueldospublicos

Spain's transparency council published its annual audit on 9 August: across 153 state public-sector bodies, average compliance with existing publication duties is 47.1 percent, and 830 of its 2,916 recommendations have been applied. The same council is drafting a criterion that would open public algorithms to citizens at three levels: source code, functional documentation, and explanation of results.

Commentary and newslettersfaf.ae

An argument that four separate governance moves in early August amount to the founding acts of an AI order.

Commentary and newslettersMIT Technology Review

Two items: a case that AI agents model the research process better than dataset-driven approaches, and an investigation of how a fringe theory about content moderation reached federal policy.

OpenAI economic research

OpenAI's economic research team argues that measuring AI's economic effects needs new data infrastructure, because existing tools cannot keep up with the frontier.

Benedict Evans

Benedict Evans notes that the big labs publish long statistical studies of what people do with their models and never publish how much people use them, and suggests there is probably a reason.

ResearcharXiv 2606.01517

Two Brazilian government units that led with structured training rather than tooling cut processing times by 18.2 and 50 percent and raised technical report output by 92 percent, across cases worth 104.3 million dollars, with no security incidents.

From the alert haulJuriNews (Portuguese)

Brazil's electoral court has created a standing council to monitor AI use and disinformation through the 2026 elections.

From the alert haulETGovernment / EnterpriseAI

The UAE is extending agentic AI across federal government functions rather than piloting it department by department.

From the alert haulPublico (Portuguese, opinion)

An opinion piece asking what Brazil gets out of joining the Chinese-led AI governance bloc.

Dwarkesh Patel

Dwarkesh Patel argues continual learning is the capability that lets models hold real jobs, and works through what that does to regulation and market structure.

LinkedIn

An argument that skepticism about Europe's AI push is due diligence rather than heresy.

J-PAL

J-PAL's practical guide for policymakers on designing and evaluating AI-enabled programs for poverty reduction.

arXiv 2607.28617

An audit of 3,249 system-prompt instructions across 88 commercial AI products, finding protective instructions common but very unevenly scoped.

ResearchNBER w34984

A survey of 746 corporate executives by the Atlanta and Richmond Feds and Duke, finding expected productivity gains concentrated in high-skill services and a gap between perceived and measured effects.

From the alert haulAIDEM 2026 workshop

A workshop paper testing generative-AI interventions on group polarization inside climate policy deliberation.

From the alert haulEx-Ante (Chile)

A profile of the adviser running artificial intelligence policy inside the Chilean presidency.

ResearchIpsos

A poll of 1,103 employed American adults finding that one in five say AI now does work they would previously have handed to a coworker or a contractor.

AcademicarXiv 2608.02699

A systematic review of 65 papers mapping the distance between the explanation EU law grants people subject to algorithmic decisions and the explanations that current technical methods can actually produce.

Discontinuity Thesis

An argument that AI is eliminating entry-level cognitive work first, and that this is a leading indicator rather than the whole story.

global-index.ai

An assessment of how 135 countries turn responsible-AI commitments into enforceable protections, built from 68,138 data points across 38 indicators.

Tech and industryInfobae (Spain)

Spain approved 27,232 public-sector posts for 2026, of which 1,700 are technology roles in AI, cybersecurity and data science, and made AI training mandatory across all civil service recruitment.

The Crimson

Roughly three dozen Harvard affiliates joined a statement, backed by nearly 3,000 signatories and 16 Nobel laureates, calling for urgent research into AI's economic effects.

Tech and industryJakarta Post

Indonesia plans to embed AI across major programmes starting with a $15 billion free-meals initiative, using it to design localized menus, monitor food safety, forecast demand and flag irregularities.

From the alert haulSunday Times (South Africa)

A South African opinion piece arguing that AI needs accountability mechanisms rather than regulation alone.

Policy institutionsIGAS / IGF / IGA, 2 July 2026

A joint report by three French inspectorates on how artificial intelligence is actually being deployed across the public administration, finding adoption real but recent, uneven and uncoordinated, with civil servants reaching for consumer tools where official ones are absent.

Stanford HAI

Stanford HAI argues legal AI tools are sold with no public account of how often they are wrong, and reports more than 1,700 court cases that have involved AI-fabricated facts or citations.

OpenAI Global Affairs

OpenAI's global-affairs team sets out six principles for a working system of independent AI safety audits and third-party assessment.

Tech and industryPhilippine News Agency

The Philippines' information and communications department and civil service commission jointly issued a binding AI governance framework covering both government agencies and the vendors that supply them, with implementing guidelines due within 90 days.

Commentary and newslettersSECOP (Brazil)

Brazil's 53rd national public-sector IT seminar, on moving AI from isolated pilots to cross-agency procurement and shared data governance.

Tech and industrySerpro (Brazil)

Brazil's federal data-processing service and management ministry put an AI assistant trained on procurement law directly inside Compras.gov.br, answering civil servants' questions on rules and procedure.

AcademicarXiv 2608.02764

A runtime architecture that keeps governance policies as reviewable programs while coordinating state across concurrent AI agents, tested on a procurement workflow where it prevented stale-policy violations over a shared budget.

Policy institutionsThe Nation (Thailand)

Thailand's digital economy ministry and ETDA set out an AI law framework alongside an amendment to the country's data protection act.

ResearchStanford HAI

Stanford HAI's annual index, reporting that AI-related publications in the natural sciences grew 26 percent in 2025 and now make up between 5.8 and 8.8 percent of research output depending on field.

Commentary and newslettersKenneth Rogoff, Project Syndicate

Kenneth Rogoff argues the intensifying US-China AI race is pushing safety down the priority list faster than safeguards can be built.

AcademicarXiv 2607.05404

A national AI exposure metric across 141 countries, finding Europe and Central Asia 50 percent more exposed than Sub-Saharan Africa, women more exposed than men in 91 percent of countries, and a new indirect channel running through remittances.

Henry Grunzweig

An argument that New York City should put AI to work across its 300,000-person workforce, with three worked prototypes in grocery siting, small-business navigation and housing approvals.

MIT Tech Review

A survey of consumer health chatbots shipping from Microsoft, Amazon, OpenAI, Anthropic and Google faster than anyone independently tests them.

Commentary and newslettersBetaNYC

A civic-tech roundup reporting that BetaNYC has released free connectors letting AI assistants read New York City Council legislation, spending records, contracts and 311 data directly.

From the alert haulPortal do Holanda (Brazil)

Brazil's superior electoral court has created a standing council to monitor disinformation and artificial intelligence ahead of the 2026 elections.

Policy institutionsVietnam Plus

Vietnam approves a data-and-AI-powered digital governance strategy to 2030, designating data as a strategic asset, with pilot projects launching across four urban centers from 30 August.

Policy institutionsMintz AI Washington Report

The US administration commits more than $5bn across fifteen-plus federal agencies to embed AI in scientific research, and launches a public-private clearinghouse for critical-infrastructure vulnerability detection.