The knowledge in this room: What comes next

[This article is based on Erica Schoder’s keynote speech at the On Think Tanks  Conference in Rabat, Morocco in May, 2026]

At the last OTT conference in Johannesburg, Senzi Bengu opened a workshop with something I have thought about often since. She said: All the knowledge we need is in this room. At the time, it felt like permission to leverage practitioner experience. Over the past year, it has become more profound. The knowledge we need is in this room because it does not exist yet. It will be produced here, in the encounter between us.

That insight is what I built my OTT keynote around. Think about what happens when a colleague pushes back on your framing, and the findings get sharper. Or when a stakeholder tells you the problem looks different from where they sit, and you realise they are right. Something was produced in that exchange that did not exist before. The most important knowledge we produce is constituted in encounter. 

The question of what produces that knowledge started as a practitioner’s puzzle, and I was surprised to find convergent evidence across multiple research traditions. It is produced when humans face decisions where other people’s stakes are embedded in the outcome. Encounters can be thick or thin. A market is among the cleanest examples of impersonal encounter producing socially useful knowledge. The institution is what channels it into something others can act on. For think tanks, the highest-stakes cases are those in which people work through hard problems together and arrive at a shared understanding that did not exist before. But encounter knowledge also runs through everyday work: the peer critique that sharpens a finding and the stakeholder whose perspective changes what you investigate next.

AI cannot stand in for humans in the encounters that produce knowledge. But it is remarkably good at working with the exhaust of human knowledge, the information that already exists. It is the most powerful research partner most of us have ever had. It can pull everything published on a policy issue across five countries and organise it into a comparison that would have taken weeks to build. For a small think tank, it is the analytical horsepower of a much bigger one. Consider what Stanford’s RegLab accomplished: AI analysed 5.2 million property deeds for racial covenants in six days at a cost of $258. The same work would have cost $1.4 million and taken one person 160 years to complete manually. That analysis enabled California legislation requiring the redaction of covenants. The work was impossible before AI. But humans built the category of “racial covenant” over decades of civil rights encounters before there was anything to hunt for. Humans decided the project mattered and worked out what to do with the findings. AI did the complexity work.

But the most useful thing AI is doing for our sector is showing us what becomes scarce. AI can produce the white paper. What that reveals is that the deliverable was never the only place the value lived. Value was also being produced along the way. The judgment and shared understanding built through the encounters that shaped the work, and the scholars formed by doing it.

That was invisible. AI is what makes it legible.

Once you see that, the question changes. It moves from deciding whether to adopt AI to understanding where AI fits into what we do. And it lives in more places than most people think.

Encounter is in the evidence, not just the convening

It would be easy to read this as a story about convening. AI handles the research, humans handle the relationships. But the encounter is inside the production of the evidence itself.

The question we decide to ask. The way we frame the problem. The evidence we choose to foreground and the evidence we set aside. The colleague who pushes back on a draft and sharpens the findings. All of that is an encounter. It happens during the production of the work, long before any convening takes place.

Some of those encounters warrant a particular piece of evidence. The colleague who says the methodology does not hold and forces a revision. The reviewer whose reputation relies on whether they let a weak claim through. They are building the work’s credibility through consequence-bearing engagement right now. Other encounters happened years ago and built the judgment the analyst brings to the page. Which framing gets traction? Which objection to preempt? The brief looks like a solo analytical product, but it is saturated with encounter. The room is bigger than you think.

OTT founder and director Enrique Mendizabal has argued that think tanks’ real value lies in their usefulness, and he has mapped what that usefulness looks like. The question I want to add is what gets produced through those activities, and why it requires human encounter.

AI can displace both phases of what we do: shaping the evidence and carrying it into the room where decisions get made. The second displacement is visible: a convening that did not happen, a consultation that got skipped. The first displacement is the one we may not notice. The framing decision that never got worked through, the analyst whose judgment did not develop because the hard cognitive work was handed off. The output might look the same, but what is underneath it is thinner.

The whole arc is where encounter knowledge lives. From the first conversation that shaped the research question through to the room where the findings are weighed by people who have to act on it. The work at each stage of the encounter is what makes the final product consequential.

Multiple research communities have arrived at similar conclusions without coordinating their work. Democratic theorists studying deliberation, economists working in the Austrian tradition on dispersed knowledge, and empirical researchers measuring what well-designed interaction produces all describe the same structure. Knowledge of this kind is produced in encounters at stake. Something specific is lost when the encounter is bypassed.

What happens when someone runs the experiment

David Caswell ran an experiment that many of us should be imagining for our own work. He conducted the same scenario-planning exercise twice: first with 60 humans over six months at roughly £250,000, then with AI agents in under two weeks at roughly £2,500. According to Caswell, the AI output was more imaginative and more complex. It was also less narratively compelling and less likely to produce organisational commitment, which is itself an encounter product: people commit to what they worked through together. The artefact was better in some dimensions, but the encounter knowledge was not there. 

I ran a version of it on myself. Late last year, when agent capabilities matured, I started building teams of AI agents who could find information, contest each other’s findings, and assemble analytical positions. They were genuinely impressive. For a split second, I thought: we might not need people for this. The split second passed. What followed were months of working to bring my colleagues into what I had learned, and realising along the way that what the agents produced was only as useful as the human judgment and perspective connected to it. The goal now is agents connected to the organisation that extend what our people can see and reason through. But that split second of recognition is worth taking seriously. The pull is structural, not personal. Someone who has spent a career bringing diverse voices into decision-making should have been the last person to feel it. The fact that it was instinctive is the unnerving part.

The futures studies community has started drawing a useful distinction between augmentation and mediation. AI augments what individuals can do while simultaneously reshaping the environment in which they do it. For think tanks, the mediation side is the one that should keep us up at night. Search interfaces now synthesise answers directly for the user. Policymakers can ask an AI agent to assemble the strongest case for a position without ever encountering the organisation that produced the evidence. Recent analysis of this shift in journalism describes the AI intermediary as gaining, in a functional sense, editorial agency over what reaches the user.

In a zero-click world, the deliverable may never open the door to policy change. Which means the encounter work underneath it becomes even more of what we are actually for.

What we stake our names on

AI cannot stake its name on what information means. We can. There are signs that trust is becoming more conditional and context-specific, and the channels through which institutions can build credibility are narrowing. For think tanks, this should sharpen the question: What are we for? The meaning-making has to happen in the people responsible for the consequences of what comes out. But institutional credibility alone may carry less weight than it once did.

Think tanks are, at their best, spaces where the trust that matters for governance is produced. The trust that emerges when people with different stakes encounter each other’s reasoning and find they can work together. That trust is produced in an encounter. Think tanks do that in Rabat, Nairobi, and Budapest, not just in Washington. The political contexts differ radically. The deliberative function is the same.

That does not mean the institutions that channel this work will look like the ones we have now. They may not. The norms and structures through which encounter gets held accountable and contestable might emerge in forms we do not yet recognise. What matters is that they exist. The function is what endures: holding people with stakes in processes rigorous enough to warrant the knowledge that comes out.

Designing what we keep

I am bullish on AI that augments human encounters. The question is what we delegate by design and what we delegate by default. Treating AI as an adoption question extends the current model: more papers, faster, with the same underlying structure. As a design question, it builds toward the next one. Cognitive offloading has been with us as long as thinking has. What changed is that AI offloads reasoning, not just memory or computation. 

And AI keeps raising the baseline of what is analytically possible. As the baseline rises, the bar on what counts as a useful encounter with a think tank rises with it, faster than in the last transition. The policy brief feels essential now. The liquid analysis that updates itself as conditions change will feel essential soon, and then it too will feel inadequate.

AI is making human work more designable. There are three sites worth being deliberate about. The first is the knowledge produced through encounter with the evidence and with the people who have to live with it. The second is the judgment that forms through cognitive work that no one can do on our behalf. And the third is the trust built in rooms where people with different stakes find they can work together.

One of the biggest risks is that we unwittingly lose diverse sites of knowledge production woven throughout our work. An algorithm can optimise for what we already know is good. It cannot replace the sites where people exercise their judgment at the point of decision, because the knowledge those sites produce emerges only when someone with a stake faces a choice. No one could have specified it in advance.

The clearest case may be the formation of judgment in junior staff. An analyst spends years doing lit reviews, drafting memos, sitting in on stakeholder meetings, and arguing with colleagues about what the evidence means. That work is how someone becomes a person whose judgment others can rely on. If AI does the lit review and the first draft, the formation path that produced the next generation of judgment does not happen by default. The design question is what replaces it. But somewhere in the chain, there has to be a feedback loop to actual humans with actual stakes. The training encounter prepares for the real encounter. It does not substitute for it.

What I want to learn from you

Here is a diagnostic exercise. Take your last completed project. Imagine AI had produced the final deliverable overnight. Ask yourself: what would not exist if AI had done it alone? That list is the invisible product. The judgment and shared understanding that the encounter produced along the way. Once you have the list, sort it: which items are knowledge produced, which are judgment formed, and which are trust built? The sort reveals where the encounter was doing the work. Then map the decision points. Where in the workflow did someone with stakes face a choice that shaped the outcome? The editorial pushback, the stakeholder consultation that changed the question. Those are the sites where knowledge about encounters was produced. Those are what you design for.

Show your funders that list. General operating support funders already intuit that the value is in the capacity, not any single deliverable. They were right. The diagnostic gives them the language to make the case to the rest of the funding community. And the diagnostic runs in both directions. Funders have their own invisible product, the judgment and relationship knowledge that no algorithm can replicate. AI is about to make that legible too. The encounter was the product, and the deliverable is what it left behind.

Once you see the encounter infrastructure underneath the deliverable, the design questions follow. Which encounters do we preserve, and which do we redesign? Where does the new landscape call for encounters we have not had before, with stakeholders we have not worked with? Some organisations will find that optimising for AI as an audience is itself a serious endeavour: bringing in affected communities so their perspectives are in the evidence AI systems process, working with technologists to understand how knowledge gets structured for machine-readable environments. Those are encounters with real stakes. They just are not the familiar ones.

I do not have a template for what the answers look like. What I do know is that the knowledge in this room, the knowledge Senzi named in Johannesburg, does not get produced once. It gets produced again every time people with stakes work through something together and arrive at something they can live with. You have been doing that for years. The work ahead is designing for it on purpose.