Stefaan Verhulst

Article by Hetan Shah: “…Given AI’s exponentially growing ability to solve well-defined, data-rich problems, it is plausible that employers will place more of a premium on the skills the humanities offer, such as critical and creative thinking. Anthropic’s co-founder Daniela Amodei recently said: “I actually think studying the humanities is going to be more important than ever . . . this idea that there are things that make us uniquely human — understanding ourselves, understanding history, understanding what makes us tick — I think that will always be really, really important.”

It would be ironic if universities cut these subjects just as demand turned a corner — switching from literature to AI-prompt engineering may turn out to have been a poor trade. And it would be bad for society if we made these subjects the preserve of the wealthy. Our research at the British Academy shows emerging regional “cold spots” where it is no longer possible to study particular humanities courses at a reasonable distance from home. This is increasingly important as nearly a third of prospective students — often the poorest — intend to stay at home while they study.

Cutbacks in universities are also affecting humanities research, where the UK is world-leading, which contributes in a broad way to building a thoughtful and prosperous society. Historians can guide us on why our policymaking is the way it is, so we can try to learn from the past. Philosophers are thinking about responsible use of AI in multiple contexts, including the creation and preservation of artworks. Academics feed into the development of video games such as the globally successful Hellblade, which had a reader in Scandinavian history as its historical adviser to provide an accurate account of Viking society and beliefs.

You might think these research cutbacks in universities are the exercise of market forces, with the strongest departments surviving. But it is more complicated than that because of the complex cross-subsidies within higher education. Financial instability is largely driven by the drop in high-paying international students and the erosion in the value of tuition fees by inflation over the past decade. Caught up in the crossfire of this instability and through no fault of their own, some of the top-ranked humanities research departments are closing, while our analysis shows the number of early-career academics in the humanities has fallen by 20 per cent over the past decade…(More)”.

Book by Kent Anderson and Joy Moore: “Scientific evidence affects policies, business, health outcomes, and economies worldwide. But is that scientific claim you just read reliable? Or nonsense? More and more often, it’s unreliable. The global system managing scientific claims has been hacked by bad ideas, big money, and bad incentives, and is being flooded by sketchy papers.

How the Internet Disrupted Science reveals the untold story of how science has been corrupted by digital information, academic and professional incentives, strange political ideologies, and big money interests. In this explosive expose, Kent Anderson and Joy Moore uncover how notions from the Big Tech world such as ‘move fast and break things’ and ‘information wants to be free’ have corrupted a scientific endeavor that once prided itself on truth-seeking, accountability, and transparency. Soon after that, scientific publishers were abdicating their responsibilities to practitioners and the public, while organized crime rings and conspiracy theorists took over.

In How the Internet Disrupted Science, two experts who witnessed this shift firsthand throughout their decades of experience in scientific publishing share a sprawling, endlessly fascinating tale decades in the making— one that is more relevant with each passing day, as we face new outbreaks, uncertainty around what information we can trust, a gutted scientific infrastructure, and concerns about centralized information in Large Language Models and AI systems. There is a way out of this mess, but only if we return to the self-correcting practices and core values that made science a reliable engine of progress for more than 500 years…(More)”.

Article by James Evans, Casey Petroff, and Gary King: “…Information theory, pioneered by mathematician Claude Shannon, offers a different perspective: only surprises carry new information. If people in an American town have conservative political preferences, and someone demonstrates that they are more likely to vote Republican, little is learned. A paper that confirms what readers already believe tells them little they did not already know and perhaps nothing that changes their worldview. Yet, the incentives of modern science reward this type of uninformative communication. Peer reviewers are drawn from one’s own field and publication norms favor deference to established paradigms. Scientists cite sources in their immediate field roughly 500% more than those in distant fields. This is performance art designed for reviewers rather than a search for the most informative surprise designed to advance knowledge. Thus, most papers are written to be unsurprising, and consequently, have minimal impact. These consequences are linked. The ultimate value that a paper adds to collective knowledge depends not only on what it shows, but on whose worldview it updates, and by how much.

To understand why surprise is the currency of impact requires understanding what science is. Science is not merely a method. It is an emergent, socially organized system that has produced more knowledge over the last 400 years than at any time in human history, by any means. The architecture deserves credit, and within it, the papers that contribute the most are those that update knowledge the most. This structure is governed by rules of cooperation (sharing methods, data, and goals) and decorum (validating empirical claims solely with empirical evidence, and justifying the importance of these claims based solely on whether other scientists will go along). But its primary driver is competition. Progress requires persuading the scientific community to accept new empirical claims. What is less appreciated is that the farther a scientist can move the predominant position of a community, the more people will acknowledge the finding, and the more importance will be attributed to the work. The predominant view is often supported by considerable evidence, which is why moving it demands extraordinary evidence and why the system offers its largest rewards for successes and largest penalties for failures, the farther a claim sits from that view. A researcher who convinces the scientific community, with robust evidence, that objects can travel faster than the speed of light would become famous, and the researcher who presses such claims without convincing evidence would be ignored or sanctioned…(More)”

Report by the OECD: “..explores the opportunities and challenges of the adoption of artificial intelligence (AI) tools to improve citizen participation. To do so, it builds on desk research and analysis of 50 AI use cases in citizen participation processes from 22 OECD Member and partner countries.

It proposes a typology of applications to help government officials and practitioners navigate the landscape of AI tools for participation based on their needs and the challenges they face. The report also provides insights on emerging trends in the adoption of AI tools for participation, analyses the related risks, and outlines relevant mitigation strategies that allow governments to steer the trustworthy adoption of the technology…(More)“

Article by Anthony Vargas: “For publishers, getting cited in AI chatbot responses isn’t enough. They also want those answers to reflect their editorial judgment and the culture of the communities they cover.

A new initiative called SAIL – which stands for Standardized Agentic Intelligence Ledger – aims to do both, compensating publishers when AI scrapes their content and guaranteeing the outputs adhere to the same cultural standards they apply to their own coverage.

The framework was designed by AI licensing platform Next Net in partnership with Sundial Media & Technology Group, the publisher of Essence, Refinery29 and Afropunk, among others.

SAIL is a digital record-keeping system that tracks how AI solutions use publisher content and mix it with other sources, said Sundial CEO Kirk McDonald. It’s meant to protect the value of high-quality publisher content when it’s cited by AI alongside less rigorous – but still culturally relevant – user-generated content.

Think of the framework as a more collaborative alternative to striking one-off licensing deals with AI vendors, McDonald said, or to the nuclear option of suing them over unauthorized content scraping…(More)”.

Paper by the Europeana Initiative: “…places Public AI within the realities of the cultural heritage sector, outlining four concrete ways for the sector supported by the data space to contribute to making it happen in practical operational terms.

The four contributions to Public AI from our sector are to:

• Provide high-quality data and the knowledge needed to keep it traceable, interpretable and reliable. In doing so, cultural heritage institutions help make AI outputs more robust, contextualised and open to scrutiny.
• Manage data access and reuse in fair and reciprocal ways. Cultural heritage institutions are reliable intermediaries and should influence the way cultural information is accessed and reused by AI developers, helping to prevent value extraction without reciprocity and the depletion of public infrastructure.
• Shape how smaller, domain-relevant AI systems are developed and used. This includes helping to shape the governance of their development and deployment, ensuring that AI remains transparent, contestable and aligned with democratic values and regulatory frameworks.
• Strengthen Public AI literacy across institutions and society. Institutions must upskill professionals while helping audiences understand how AI works and engage critically with AI systems they encounter…(More)”.

Article by Sarah O’Connor: “In the early 20th century, a gifted engineer called Frederick Winslow Taylor embarked on an ambitious task: to extract knowledge from inside the heads of workers on America’s factory floors.

In the eyes of Taylor, who would go on to become one of the world’s first management consultants, factory workers possessed a “mass of rule-of-thumb or traditional knowledge” which had been “handed down from man to man by word of mouth” or “almost unconsciously learnt through personal observation”. Taylor thought it was about time this knowledge was “codified or systematically analysed or described”. To that end, he sent managers with stopwatches and notebooks on to shop floors to observe, time and record every stage of every job.

More than a century later, employers of white-collar professionals are beginning to confront a similar challenge. It is becoming increasingly clear that the knowledge required to make AI models genuinely powerful in a swath of workplaces is currently locked inside employees’ heads.

This isn’t true in every workplace. AI models have transformed the software profession, for example, because the task of writing code is testable and rules-based and there were vast reams of training data publicly available, thanks to online forums like Stack Overflow.

But for many other jobs, that sort of data just does not exist on the web. Indeed, some subtle but important skills are very hard to codify at all, which is why they are often learnt through experience and osmosis. This sort of tacit knowledge was famously summed up by the scientist and philosopher Michael Polanyi with the phrase: “we can know more than we can tell”.

As a result, general-purpose AI models are simply not very good at many specific tasks which require both domain and institutional knowledge. Investment firm Bridgewater Associates recently experimented, for example, with using LLMs to do something their human professionals do all the time: parsing reams of news stories and financial documents for information that might be relevant to their investment decisions.

While this could be a useful timesaver, Bridgewater found that variants of Gemini, Claude and GPT only tended to get it right about 50 per cent of the time..(More)”.

Article by The Economist: “WHEN THE leaders of the biggest Western economies gathered last month in the French Alps for the G7 summit, it was not just presidents and prime ministers in attendance. The bosses of America’s leading artificial-intelligence firms—Sam Altman of OpenAI, Sir Demis Hassabis of Google DeepMind and Dario Amodei of Anthropic—were seated alongside the political bigwigs. Even the stewards of the mightiest economies, it turns out, are buttering up the potentates of AI.

Days before the summit, America’s government had barred Anthropic from making its most advanced model available to foreigners. OpenAI soon imposed similar restrictions on some of its frontier systems. The move turned a simmering concern into a burning one: might countries without their own, home-grown AI firms end up excluded from the world-changing technology? Emmanuel Macron, France’s president, warned that no one would buy American AI if it could “turn off the switch” at will. The American AI labs’ Chinese rivals offer only faint reassurance. On July 7th Reuters reported that Chinese officials, too, were considering restricting foreigners’ access to leading AI models.

Sovereign chic

Caught between the world’s two AI superpowers, governments are increasingly pinning their hopes on “sovereign AI”: the idea that countries should limit their dependence on foreign providers of AI by actively developing domestic AI firms and infrastructure. Announcements of such initiatives are proliferating. The European Union has unveiled a technology-sovereignty package encompassing semiconductors, AI and cloud computing. Canada has launched its own plan to reduce reliance on American technology providers. India, Japan and Singapore are backing domestic AI infrastructure and models. The Centre for a New American Security (CNAS), a think-tank in Washington, estimates that the number of state-backed AI projects outside America and China grew fivefold over the course of 2024 and 2025 (see chart 1). Governments have announced investments of $70bn-plus in such initiatives, according to CNAS’s tally.

Yet turning these ambitions into reality will be extremely difficult. No matter what, most countries will still rely on American chips, Chinese open-source models or both to build their AI systems. Even then, most governments will find it prohibitively expensive to achieve anything remotely resembling AI sovereignty.

The incentive to seek it, however, is clear and compelling. Pablo Chavez of CNAS puts the restrictions on Anthropic in a “long line of events” demonstrating America’s willingness to use its technological dominion as a weapon. In the dying days of Joe Biden’s presidency in 2025, America proposed sweeping export controls that would have rationed access to advanced chips through a complex licensing regime. When Donald Trump became president a few weeks later he dropped the plan. But he has not hesitated to use America’s lead in AI as leverage in trade negotiations. The Department of Commerce is considering rules that would allow it to vet all sales of AI chips designed by American firms anywhere in the world. It might also insist that buyers allow inspections or monitoring to make sure chips are being used as promised. Another possible requirement could be to invest in American AI infrastructure in exchange for access…(More)”.

Policy Backgrounder by The Conference Board: “On the eve of the 2024 presidential election, while polls showed a race that was too close to call, online platforms called “prediction markets” told a different story. Contracts trading on the largest prediction market, Polymarket, placed the President’s odds of victory as high as 67% in the closing week of the race. For supporters of prediction markets, the President’s eventual victory – with nearly 58% of electoral votes – was evidence of their power to more accurately predict future events than existing methods, a phenomenon that has led Polymarket’s CEO to call them “global truth machines.” The outcome also helped shift what had once been a relatively obscure interest of some economists and traders into a mass market phenomenon – monthly volume on Polymarket and Kalshi jumped from less than $1 billion in mid-2024 to nearly $24 billion by early-2026.

Fundamentally, prediction markets are simple – users trade contracts tied to the outcome of a future event. In the most common format, an “event contract” might pay $1 if a candidate wins an election, if a hurricane makes landfall in a specified region, if inflation exceeds a stated level, or if a sports team wins a championship. This functionality could have both entertainment value – in the case of a contract that hinges on the length of the Super Bowl halftime show, for example – and real economic value – such as contracts predicting US gas prices.

For events with significant trading volume, supporters argue that these markets can translate large amounts of complicated real-world information into a market price that more accurately predicts events than existing methods such as polls or expert analysis. However, prediction markets have also raised significant regulatory and legal concerns related to alleged insider trading, outcome manipulation, regulatory arbitrage, and other issues. These questions have prompted policymakers and stakeholders to debate whether legal or regulatory action may be needed to protect users, clarify the boundary between trading event contracts and gambling, and preserve the potential economic value of prediction markets…(More)”.

Paper by Sachit Mahajan: “AI-assisted consultation can speed large-scale public engagement, but concise summaries may reflect some submissions more closely than others. This paper introduces participatory provenance, a framework for auditing how semantic coverage is distributed from submissions to summary sentences. Applied to two topics in Canada’s 2025 AI Strategy consultation (5,253 records; 2,861 participants), official summaries had higher observed mean coverage than exact-length random text, although statistical significance depended on the embedding model. Low coverage concentrated in semantic regions, especially those centered on criticism of educational technology and distrust of technology and oversight, whereas few or no records crossed the operational threshold in several better-covered regions. Same-budget, cross-fitted extractive benchmarks improved mean and lower-tail coverage on held-out submissions, showing that better semantic coverage was feasible without longer summaries. Consultation summaries should be evaluated not only for coherence and factual support, but also for how coverage is distributed across the range of submitted views…(More)”.