The programme set out to tackle three interlinked challenges (could I ever use bullet points without wondering if I sound like an AI?):

· Getting data AI-ready: for AI to be reliable and fair, the data it uses must be well governed, clearly structured, and accompanied by information about its provenance and content.

· Making AI-ready data available: AI can only deliver on its promise if access to data in an AI-ready format is not restricted to a few organisations. Open data remains a driver of AI progress and innovation.

· Ensuring responsible data practices: transparency about training and other forms of data is essential for oversight and public trust. Governments and organisations need clearer information about which datasets AI systems use and how well those systems perform.

Today, I think the same message is not only still valid but considerably more urgent than it was then. The AI landscape is changing and there’s specific data-adjacent points that are overlooked and could have grave consequences for the next wave of AI innovation. As I a techno-optimist, I see them creating a renewed opportunity to build data ecosystems that serve broad public interests rather than narrow commercial ones as it’s the case for consumer-facing AI.

Enterprise AI needs data foundations, not just foundation models

AI has matured rapidly from a research curiosity into something closer to a general-purpose technology. This means it now needs to deliver in specific, often messy contexts: enterprise processes, regulatory environments, legacy systems, real business constraints. Companies that try to adopt AI without first getting their data in order tend to find this out the hard way, and spend lots of money to do so.

This must sound familiar! In the early 2000s, the vision of semantic web services promised that software components would discover each other, negotiate terms, and compose themselves into workflows automatically, provided the right metadata, ontologies and reasoning/planning capabilities were in place. I co-authored a book on this topic with my mentor, the late Dieter Fensel, whom I still thoroughly miss.

Business process management, workflow languages, and enterprise application integration essentially say the same thing: describe your data and services richly enough, and machines, today called AI agents, can do the rest. Those ideas were ahead of their time and adoption was limited, partly because the investment in data foundations that would have been needed to make it work never materialised at scale.

Now, as AI is increasingly being adopted in enterprise settings, there is a genuine chance that organisations may finally build the semantic data layer they have needed for decades: rich metadata, transparent provenance, proper governance. Not because these things are nice to have, but because without them, AI will fail. The models will hallucinate, the agents will make poor decisions, and the compliance risks will be unmanageable. Atanas Kiryakov gave a fabulous talk about this at the European Semantic Web Conference.

This is also the focus of IDEA, a programme of research, community building, and peer learning that the ODI is delivering in partnership with SAP. IDEA stands for “Interchange for Data and Enterprise AI” and it is designed to help organisations of all sizes make their data infrastructure ready for AI. It is vendor neutral and community led, and we are actively looking for more partners who want to help shape the standards and frameworks that will define what AI-ready enterprise data looks like in practice.

Embodied AI needs new kinds of data (and new kinds of care)

This one has been on my mind a lot recently. AI is entering the physical world through robotics and embodied intelligence, and the data requirements for this are acute. There’s no public web to create foundational models for embodied AI for everyone, including smaller players, to use and build on.

China has understood this at an extraordinary scale. Over the past year, a series of government-backed humanoid robot training centres have opened across the country, in Suzhou, Wuhan, Sichuan, and elsewhere, where human trainers wearing VR equipment physically teach robots to fold clothes, clear tables, pour tea, and sort materials. These centres are generating millions of data points per year and sharing them across the robotics industry. In parallel, Chinese companies are paying thousands of residents and factory workers to film themselves doing everyday tasks at home to create first person training data for robot learning. MIT Technology Review recently described data infrastructure as the defining story of 2026 in robotics.

As far as I know, we do not have anything comparable in Europe or the UK. This matters not only for competitiveness but also for tech sovereignty (not my favourite term, but that’s a story for another post!). Robotics training data is deeply contextual. It captures how people move, live, and work in specific cultural and home settings. Collecting it at scale raises questions about consent, working conditions, representation, and who benefits from the resulting systems.

These are questions I’d love to work more on again. I spent years researching microtask crowdsourcing and the people who do data work for AI. For instance, in a paper I co-authored in Data & Policy, we explored whether data workers could be included as stakeholders in data trust design, giving them a say in how the datasets they help create are governed and used. The conclusion was that current governance models almost entirely ignore the people who produce the data, even though they are directly affected by how it is exploited. Others have gone beyond abstract frameworks to drive change in the sector e.g. Mark Graham’s Fairwork project that rates companies across 39 countries, slowly and steadily driving policy changes. They have recently extended their methodology to AI data supply chains, examining the conditions under which the data powering AI systems is actually produced. As new forms of physical, embodied training data become central to the next wave of AI, we could do things better from the start. When I’m asked about this topic, I always say that if AI were a sandwich or a t-shirt, people would not touch it if they knew its supply chain.

What does good data look like for agentic AI?

The third development is of course agentic AI. The original pitch for data-centric AI was: we really need to pay attention to well-governed AI data to train or fine-tune machine-learning models. In that context, the data was read and maybe occasionally updated. Agentic data flows are more dynamic, as agents need to discover data, assess its quality and provenance, combine it with other sources, and increasingly, write back to it, update it, and act on the basis of what it says.

Hence data needs to be well structured, richly described, and transparently governed. At the ODI, we have been working on what this means in practice through our Framework for AI-ready Data, which I co-authored with Neil Majithia and Thomas Carey-Wilson. The framework, now in its second version, sets out actionable recommendations across four areas: the design of datasets, metadata, infrastructure, and governance. It was published as an academic paper in the AI Magazine earlier this year. What I find most interesting about the agentic use case is that it exposes the inadequacy of data that was designed only for human consumption. Agents operating autonomously in the world need machine-readable provenance, clear access protocols, and reliable quality signals. If the data is not there, the agent cannot be trusted, and rightly so.

We need to build new data foundations fast, but can we also build better?

Each of these developments is generating new and distinctive forms of data: robotics training data, business ontologies, agent interaction logs, multimodal datasets that capture movement, context, and intent. These are not existing data repurposed for AI. They are new data being created right now, and that gives us something valuable: a chance to get the foundations right from the start.

For these new ecosystems, governments, the private sector, and civic society have a real opportunity to come together and establish governance and stewardship mechanisms that create genuine public data utilities, supporting broad innovation rather than concentrating it among a few large players. The Data Labs established in Europe need to make this a priority, building on the common data spaces. In the UK we don’t have anything equivalent that I’m aware of, as the National Data Library is growing closer to data.gov.uk. There are some encouraging models to draw on. Wikimedia Enterprise, for instance, offers a way for AI companies that depend on Wikipedia’s human curated knowledge to access it through a sustainable, structured arrangement rather than simply scraping it. It is a practical example of how commons based resources can be maintained and fairly financed in the generative AI age, ensuring that the communities who create the knowledge continue to benefit from it.

No data, no AI does sound like a slogan and yet it’s a call to take the data side of AI seriously and, crucially, to do so with an eye on the entire ecosystem, not just a handful of organisations.

What else has been happening

· I finally managed to visit V&A East Storehouse. Amazing experience, the way they are clustering their artefacts was refreshing for a classically trained knowledge engineer. Almost no temporal or categorical element, which made me think about how to design serendipitous knowledge organisation. No big ideas yet, but I’ll keep thinking.

· Most speaking and writing I’ve been doing lately has been on AI and the workforce, which made me wonder about how my and other data and computational professions are changing. More about this hopefully soon.

· I was honoured to be invited on Friday to the Wikipedia’s 25th anniversary, which also marked 15 years of Wikimedia UK being a charity. Such a rich, diverse community. Looking forward to Wikimania later this month.

Knowledge engineering is the practice of creating and maintaining knowledge-based systems. It sits between software engineering, which involves representing knowledge in code, and AI, where software can reason over those representations in ways that begin to resemble human thought. Like much of knowledge work, the field is being reshaped by generative AI. This is of course exciting as it opens up real opportunities to address long-standing challenges of scale and inclusivity. But we also need to be mindful of familiar concerns about accuracy, bias, and the responsible use of automation. My team at King’s College London works at the intersection of AI, human-computer interaction, and social computing, trying to design better knowledge engineering tools and assistants.

Doing that well requires knowledge engineers, who have traditionally been trained in symbolic AI, to work closely with HCI and human-centred AI researchers, who have published hundreds, if not thousands of papers on human factors in AI over the past years.

Here is my take on the things that caught my attention at the workshop.

Automation and augmentation

Most of the research applying generative and conversational AI to knowledge engineering has focused on automating specific knowledge base construction tasks. In my lab, students have recently looked at taxonomy building and schema generation, but the list of relevant tasks is long, as explored in a recent review (link to be added). Others have tackled particular components of the knowledge engineering pipeline. For instance, OntoScope is an LLM-based system that helps ontology engineers scope their domain and requirements using a divergent-convergent interaction framework. All of this work seeks automation to achieve scale, while acknowledging the importance of keeping a human in the loop. When the task turns out to be harder than expected, as in our recent ESWC 2026 paper, the obvious conclusion tends to be that we need to push harder with AI. But that’s not always the only valid conclusion.

The people in the loop

There will always be scenarios where a person needs to validate what AI has produced. High-stakes domains require it. But beyond accuracy and safety, a lot of my work has focused on what motivates people to contribute to a knowledge base in the first place. If everything routine is automated, what happens to those motivations?

Advait Sarkar made a point that has stayed with me: the question of what and how to automate a knowledge task is not as simple as some studies suggest. He reminded us that doing simpler tasks has real value, as it helps people build the readiness and understanding they need to take on harder, more complex tasks when those arrive. In online communities, this is captured by the theory of legitimate peripheral participation (Lave and Wenger, 1991), which describes how newcomers learn and gain confidence by starting with small, low-risk contributions that are nonetheless genuinely useful. Entry-level jobs work the same way, and those are precisely the roles most exposed to AI automation right now. Getting this right matters.

Designing for better thinking

Assuming we have some sense of what to automate or augment, there is still the question of how to design dialogue agents that engage people meaningfully in building an ontology or filling gaps in a knowledge graph. Advait spoke very compellingly about AI as a provocateur, and presented experiments showing that adding deliberate provocations to AI-assisted knowledge work, brief critiques and alternatives surfaced at the right moment, has measurable positive effects on critical thinking and leads to more diverse outputs. That last point matters. If social media has pushed us into filter bubbles, AI tools risk doing something similar but at the level of ideas: collapsing the range of things people think and say towards a generic middle. In ontology engineering, a shared conceptualisation should not be achieved because the AI we use mode-collapse but because the parties involved reach an agreement, supported, but not overrelying on AI.

This raises some follow-up questions for me. What does the knowledge engineering experience of this era actually look like? Is it a Protégé-style tool with AI capabilities layered on top? A conversational agent that understands what people are trying to build? Some combination of both? And how do we design tools that genuinely help a people build a shared understanding of a domain, meeting real requirements, promoting good modelling practice, supporting critical thinking, and surfacing potential biases? There is rich HCI work on exactly these questions, including this paper on prompt middleware and this scoping review on designing for curiosity in cultural heritage.

What does good look like?

Hanna Wallach spoke about the evaluation and measurement of generative AI systems, and the need for social science methods alongside technical ones. I could not agree more. My view is that there is a real need to benchmark foundational AI not just on technical knowledge tasks such as link prediction and class representation learning, but on user-centred ones that mirror how people actually use downstream systems built on a knowledge base. This is not always straightforward, as we have generally been taught to design knowledge bases for reuse across application domains. But knowledge engineering projects need to at least be aware of this tension and plan for user-centred evaluation once applications become known.

There is also a much bigger question about participation in the evaluation and auditing of knowledge graphs. This connects to work I am involved in through the PHAWM project, which is exploring participatory approaches to auditing AI for harm and bias. I hope to bring some of those ideas back to the knowledge graph world.

Participatory knowledge graphs

Participation was also the theme of an excellent presentation by Anja Thieme, who talked about how to involve local communities in creating AI datasets in ways that are genuinely respectful and thoughtful. The paper focuses on disability representation in AI image generation, and it made me realise that I cannot think of a comparable example in knowledge engineering where knowledge curation has been approached with the same level of care. I suspect examples exist, perhaps in cultural heritage, and I intend to look for them. Knowledge graphs play a significant role in modern AI systems, and my instinct is that the localisation of knowledge is a significantly underexplored area.

Second-order effects

The impact of generative AI on knowledge work is still largely unmapped. I came across a poster at the workshop that looked at second-order effects, specifically what happens to people who interact with people who use AI, not just those who use it directly. Imagine a team made up of domain experts, knowledge engineers, perhaps a few ML engineers, and end-users. Would they trust a knowledge base they knew had been co-generated with AI? This is a genuinely open question and I don’t think dilligently recording knowledge provenance will be enough.

Final thoughts

One small thing I appreciated: the event was called “people-centric” rather than “human-centric” AI. As someone who worked on social computing and collective intelligence, this framing felt right to me. I hope to see more research into AI in teams and communities, not just the interaction between an individual and an AI agent. It’s time for a refreshed research agenda for collaborative knowledge engineering - human and AI agents working together, but also people collaborating more effectively supported by AI tools.

I want to end with something from Neil Lawrence’s keynote. He reminded us that intelligence is not a problem to solve, as he argues in The Atomic Human. And neither, I would add, is knowledge. Access to knowledge, yes. Creating and maintaining machine- and human-readable knowledge artefacts that document and organise what we know, yes. But even when we use AI to build those artefacts, we should not lose sight of their social fabric, which is just as essential, and in my view considerably more interesting to study :-)

More to come. I have, perhaps unwisely given my workload, promised to give a few talks on these topics over the coming months.

And a sincere thank you to Denny Vrandecic for answering my questions from last week.

She generously invited me along to make some concluding remarks, which I thought would make for a fitting first post. If I ever were to write a book, nearly not as eloquently and eruditely as Giota, I could probably call it the same - well, not anymore, I couldn’t! But it would be about similar topics, tackled with the mindset, methods and tools of an AI scientist.

The Web of Knowledge offers something genuinely valuable to those of us working in computing: a rigorous historical and analytical account of a system we often take entirely for granted.

As computer scientists, we build the infrastructure upon which encyclopedic knowledge now rests. We design the platforms, the APIs, the governance algorithms, and the recommendation systems. This book serves as a timely reminder that what flows through that infrastructure is not neutral data; it is contested, culturally situated, and politically charged knowledge.

This resonates with my own scientific journey. Since my PhD in 2001, I have researched semantic technologies, a field sparked by the vision of Tim Berners-Lee, James Hendler, and Ora Lassila for a “semantic web.” Their goal was a technology stack that created web content meaningful to computer agents, where human knowledge was augmented with machine-readable information structured through shared ontologies.

For the first fifteen years after the paper came out, these ontologies were built manually, meaning they could only scale through mass participation. This led me, and others, to peer-production systems like Wikipedia and Wikidata. Over time, I have experienced firsthand exactly what this book describes: the “web” is not merely technical; it is social, economic, and epistemic.

Several threads in the book deserve particular attention from our community:

The question of governance

Wikipedia’s model, characterized by open contribution and distributed moderation, is essentially a distributed-systems problem with human nodes. The successes and failures documented in managing authority and bias are fundamentally problems of consensus, trust, and incentive design: challenges I have sought to address in my own research on crowdsourcing and Wikidata.

The question of labour

The creative and editorial work sustaining these platforms is often invisible within our system architectures. Giota makes a compelling case that it should not be.

The question of access

As encyclopedias increasingly mediate how we understand the world, our design choices regarding ranking and availability carry real weight. We see this today in the disruption brought by generative AI. Answer engines and chatbots are the new emerging forms of algorithmically mediated knowledge. As an educator, I believe it is vital to provide students and researchers with the most advanced tools to ensure we narrow the “AI divide.”

What does this mean for we AI we build

Wikipedia’s evolution offers a direct lesson for those of us building AI. Large language models are trained on the very encyclopedic knowledge this book examines, which is knowledge shaped by editorial gatekeeping, cultural bias, and uneven representation. Wikipedia’s decades-long struggle with neutrality is not a footnote to AI development; it is a preview of the challenges we face.

If we want AI systems that are trustworthy and equitable, we must study how Wikipedia succeeded and failed in governing its own knowledge commons. We should hold AI governance to the same standards, asking of every new system the same fundamental question: Would it break Wikipedia? And conversely: Would Wikipedia break if it were governed like this? We must continue to teach the technical AI community to operationalize these questions, ensuring that the web of knowledge of today and tomorrow maintains the human qualities we cherish.

I am so grateful to Giota for inviting me to share my thoughts at her book launch and for the inspiring conversations that followed. Her work provided a wonderful opportunity to step back and reflect on my own journey, prompting me to start my journey here.

I plan to share my thoughts here a few times a month as inspiration strikes.

Among others, I’ll conclude each time with some of the things I wonder about at the moment in case you can point me to some answers or feedback. Right now, I’m thinking about sonification as a way to communicate how a digital artifact, like a Wikipedia article or a dataset, comes about. I haven’t done any serious work on sonification, but it has fascinated me for a long time - maybe a subject for another day.

Giota spoke about a sonification project built on Wikipedia. Essentially you can “hear” how different articles across the encyclopedia are being written. I wonder: Do edits by people and bots sound the same? Do articles where a large share of edits is automated sound the same? If readers were to experience that information in a accessible audio form, would they engage with the content differently? Are there any lessons we could learn about algorithmic transparency more widely?

Subscribe now