Latest insights

The debate about artificial intelligence is still too often framed as a choice between rapid innovation on one side and restrictive regulation on the other. That is a false choice. The real issue is how to capture the extraordinary benefits of AI while ensuring that the systems being deployed remain accountable, resilient and worthy of public trust.

Transforming Public Services with AI The debate about artificial intelligence is still too often framed as a choice between rapid innovation on one side and restrictive regulation on the other. That is a false choice. The real issue is how to capture the extraordinary benefits of AI while ensuring that the systems being deployed remain accountable, resilient and worthy of public trust.

The UK Government’s Extract programme is not particularly glamorous. It does not involve humanoid robots, superintelligence or another headline-grabbing frontier model. It tackles something much more useful: the extraordinary amount of planning information still locked away in paper files, scanned PDFs, old maps and microfiche. Extract uses AI to convert that material into structured, usable geospatial data. Early testing showed that key information could be extracted in arou

Kanishka Narayan, the UK Minister for Artificial Intelligence, opened his speech at Imperial College this week with an interesting historical analogy. Two centuries ago, industrialisation was transforming Britain. New technologies created wealth and extraordinary economic progress, but they also imposed heavy costs. Children worked in appalling conditions and early attempts at regulation were largely ineffective. The state eventually responded with the Factory Act of 1833 and

What happens when an AI system makes a mistake that affects someone’s healthcare, welfare benefits or access to public services? Who takes responsibility? On 29 September 2026, the White House introduced its Accord on Super Intelligence, bringing leading technology companies together around a voluntary commitment to AI safety and accountability. The agreement reflects growing recognition that increasingly powerful AI systems require stronger safeguards.

The predicted employment apocalypse has not arrived. But two major banking announcements suggest that something more significant is beginning to happen. For the past two years, predictions about the impact of artificial intelligence on employment have ranged from reassuring to apocalyptic. So far, the evidence has supported neither extreme.

Governments do not have an AI experimentation problem. They have a scaling problem. Across the world, public institutions have spent years launching pilots, testing tools and exploring what artificial intelligence might do for government. In 2026, the conversation is changing.

Nvidia is paying nearly $13 billion for Hugging Face, a company reportedly generating around $150 million in annualised revenue. The numbers are striking. But focusing on the revenue multiple risks missing the strategic significance of the deal. Nvidia is not simply acquiring another AI company. It is acquiring a strategic position in one of the ecosystems where more than 18 million developers, researchers and creators discover, evaluate, customise and deploy AI. And that sho

In forty years of financing technology companies I have learnt to pay close attention when a newcomer does not try to beat the market leader at its own game, but asks whether the game needs playing at all. TypeSafe AI, a small San Francisco company that emerged from stealth on 15 September, is asking that question of the large language models, and I suspect the answer will interest investors as much as developers.

Below is Amdodei’s important warning about the frightening pace of AI development. I have reproduced Dario’s letter in full after my own introductory comments. Victorian railways ran into a problem that will sound familiar. Trains had become faster than the brakes fitted to them. Inspectors had pressed for continuous automatic braking for years and the companies had resisted on grounds of cost. On 12 June 1889 a Sunday school excursion train stalled on a gradient outside Arma

Most discussion about AI in government takes place at the national level. Governments publish strategies, legislatures debate regulation and central departments announce major technology programmes. Yet this is not where most citizens experience government. They experience it through local services: applying for planning permission, reporting a damaged road, paying a local charge, seeking community support, using public transport or challenging an enforcement decision. Local

Molti governi stanno investendo nell’intelligenza artificiale con una promessa precisa: fare di più con meno. Meno tempo dedicato alle pratiche burocratiche. Elaborazione più rapida dei documenti. Costi operativi più bassi. Ma se accadesse il contrario? Se l’IA non riducesse il volume del lavoro pubblico, ma rendesse possibile svolgerne molto di più? Questa ipotesi può essere letta attraverso una teoria economica formulata oltre 160 anni fa: il Paradosso di Jevons.

Google has published one of the largest empirical studies yet of how generative AI is actually being used. Its new AI & Economy ATLAS report analyses nearly 15 million de-identified interactions across the Gemini App, Google AI Mode and the Gemini API. It maps usage across more than 800 occupations, 4,000 work tasks, 300 household activities, 150 countries and 140 languages.

This week, the UK’s new Prime Minister, Andy Burnham, announced that the Department for Science, Innovation and Technology (DSIT) would be disbanded, with its responsibilities redistributed across government. The principal organisations responsible for AI will be the Cabinet Office; the new Department for Business, Innovation, Science and Trade (DBIST); and the expanded Department for Digital, Culture, Media and Sport (DCMS), which will incorporate the Government Digital Serv

Picture this: a chatbot notices you mention self-harm – and under a new Rhode Island law, its maker must steer you toward crisis support or face penalties for noncompliance. That’s not hypothetical. It is part of a broader June 2026 wave of state action that made this one of the busiest months yet for AI and privacy regulation.

For years, the public sector’s relationship with artificial intelligence was defined by cautious pilots and basic chatbots. That phase is giving way to a more consequential shift: AI systems are beginning to support end-to-end service delivery by routing requests, validating information, and coordinating work across digital systems, while keeping humans in the loop for exceptions and high-stakes decisions.

I read a fascinating paper from the Hoover Institute written by Niall Ferguson “Artificial Intelligence and Global Security”. I offer a summary here however the full article is well worth reading. Ferguson’s central argument is that artificial intelligence is now developing through two barely controlled races. The first is a race between a small number of dominant technology companies. The second is a geopolitical race between the United States and China. His concern is not s

Health systems across the world are under pressure. Waiting lists are lengthening, costs are rising, clinicians are overstretched and public expectations continue to grow. In poorer countries, millions still lack access to basic services. In richer countries, universal health systems are struggling to deliver timely care. The underlying problem is structural: demand is rising faster than available money, workforce capacity and institutional productivity. Artificial intelligen

England’s education system is showing signs of structural fatigue. For decades the same policy tools have been used to address persistent problems such as the attainment gap and the growing difficulty of recruiting and retaining teachers.

Agentic AI shines where work is structured, multi-step, and slightly tedious, the kind of work that drains expensive human attention. The goal isn’t to replace judgement. It’s to industrialise the first draft, then let humans do what humans do best: decide.