For the last few years, think tanks have been talking about artificial intelligence as if it were either a threat to be contained or a productivity hack to be quickly adopted. At first, this made sense. ChatGPT arrived suddenly. I played with it over a holiday break, asking it to define a think tank, draft a policy brief, or outline a course, and was both impressed and worried by what came back. I wrote in January 2023 that my first experiments with ChatGPT left me excited and concerned: the answers were often good enough to force a serious conversation about what, exactly, remains distinctive about our work in a world where machines can produce competent text.
That conversation has now moved on.
At the 2026 OTT Conference in Rabat, AI appeared again and again in keynotes, parallel session and conversations. But it was no longer discussed simply as a tool. Instead, it took centre stage in our discussions on think tanks and trust. The Conference opened with a public event on visibility, credibility and policy influence, and the programme returned repeatedly to the question of how think tanks earn, protect and demonstrate trust in an information ecosystem transformed by AI.
AI is not just changing how think tanks work, but also how others judge the work think tanks do.
If AI can produce a policy brief, summarise a report, generate a stakeholder map, translate findings, draft a media pitch, scrape a parliament’s website, and answer a policymaker’s question faster than any research team, then the value of a think tank cannot rest on output alone. It has to rest on other contributions: judgment, relationships, accountability, quality assurance, public purpose, etc. In her keynote, Erica Schoder argued this point by arguing that the distinctive contribution of think tanks lies not only in analytical products, but in the knowledge that comes into being when people with real stakes work through hard problems together.
AI does not make think tanks irrelevant, but it does make vague claims of credibility less convincing. “Trust us” is not enough. We need to show our workings.
We already knew this was coming
On Think Tanks has been tracking AI’s implications for the sector for several years. In 2024, I wrote about the promise and perils of AI in shaping tomorrow’s think tanks, policymakers and foundations. Aidan Muller argued that think tanks’ engagement with AI had been more opportunistic than strategic and proposed a preparedness framework that included experimentation, leadership, internal policy, AI/data strategy, online footprint audits and thought leadership on AI policy. His central point was not “use AI more”. It was: bring order to the chaos.
That same year, Keith Burnet and Nick Scott shared WonkComms survey findings showing that think tank communicators were already using AI at work, but that organisational support, training and policies lagged behind actual practice. Their survey found high AI usage, strong demand for training, and reputational risks around undisclosed AI-generated content. Their warning was direct: in a sector where credibility is key, high AI use combined with little training is a risk.
Joscha Wirtz, writing about Sociopúblico’s first AI for Think Tanks cohort, made a complementary argument: think tanks need to move from FOMO to intentionality. The programme he described was built around practical deliverables such as an AI landscape analysis, prompt library, ethical guidelines, readiness framework and the role of certified AI stewards inside organisations.
More recently, Tony Bader made the credibility problem explicit. In his article on why every think tank needs an internal “AI constitution”, he argued that think tanks cannot credibly advise governments on AI governance while leaving their own internal AI practices undefined. An AI constitution, in his framing, is not a legal document or a technical manual, but a shared institutional framework for how AI can and cannot be used.
And for the last few years we have been tracking think tank adoption of AI through the OTT State of the Sector Report: increasing but uneven – higher for comms and deeper in larger think tanks and in higher-income contexts.
Taken together, these pieces show a progression in OTT’s own thinking. First, we asked: What can AI do? Then: Are think tanks ready? Then: How do we govern AI internally? At Rabat, the next question became unavoidable: how do we make responsible AI use visible and trustworthy to others?
Trust is no longer only human-to-human
For all of their history, think tanks built trust through people: personal relationships with policymakers, reputations of researchers, repeat interactions with journalists, invitations to closed rooms, citations by peers, and the slow accumulation of institutional credibility.
Those still matter. In fact, according to this year’s keynote speakers, they may matter more than ever. But they are no longer enough.
Think tank outputs now circulate through platforms, search engines, AI agents, databases, repositories, automated summaries and chat interfaces. Increasingly, a policymaker may encounter a think tank’s work not by visiting its website, but through a generated answer. A student may ask ChatGPT for sources. A journalist may rely on a search summary. A funder may use AI to scan hundreds of reports. A civil servant may ask a chatbot to compare competing policy options.
Our own content at OTT is increasingly accessed through LLMs!
This changes the trust problem. It is not enough to ask if the reader trusts us. We may now need to ask if the system through which the reader accesses us recognises what we are, how we work, who funded us, what was reviewed, what was generated, what was human-led, and whether the object it is presenting is authentic?
A session convened by Toby Green of Policy Commons and Nick Cleveland-Stout of the Quincy Institute — “Your Research Is Good, But Does Anyone Know That?” — tackled a common challenge: many produce strong work without visible, structured signals that help people, platforms and AI systems distinguish it from everything else.
This is especially serious for smaller, newer, regional or non-English-language think tanks. If AI systems learn to privilege already-visible institutions, then AI may deepen existing inequalities in the global policy knowledge ecosystem. Good work may become even harder to find unless it carries recognisable signals of provenance, review, funding transparency and institutional accountability.
The old transparency agenda is back — but now it is digital
This is not entirely new. We have been arguing for transparency in the think tank sector for more than a decade. In 2014, OTT published a piece asking whether transparency for think tanks was “the latest fashion or an urgent reform”, noting that without transparency about funding and research processes, think tanks provide ammunition to those who distrust their independence and credibility.
Nick Cleveland-Stout’s 2025 article on US think tanks brought this debate into sharper relief. Drawing on the State of the Sector 2025 and Quincy Institute research, he argued that the US think tank sector has a financial disclosure problem: North American think tanks had a funding disclosure rate of only 35% in the survey, and 18 of the top 50 US think tanks did not disclose any information about their funding sources.
Trust does not collapse only when research is poor. It collapses when audiences suspect hidden interests, unclear methods, weak review, undeclared conflicts or undisclosed use of AI. This is how think tanks are often attached: Orazio Bellettini, founder of FARO, often recalled how Ecuadorian President Rafael Correa would seek to undermine them: “Who is behind FARP”, he would ask in his public broadcasts, implying that the think tank was funded by opaque sources (it wasn’t, and FARO was very transparent about its funding).
In Rabat, the conversation moved from transparency as a moral commitment to transparency as infrastructure. The Trustmarc proposal discussed by Toby and Nick in their session imagines a machine-readable, clickable digital token that could be incorporated into think tank outputs — reports, blogs, podcasts, datasets — and reveal an independent certificate covering key signals of trustworthiness: governance, funding, review level, authorship, provenance and stewardship. The proposal also included persistent identifiers, such as a Content Object Identifier, and tombstone pages to preserve citation integrity when content is removed or updated.
The future – and I would argue, the present – will require structured, interoperable, verifiable signals. A PDF with information about a think tank’s governance, funding and quality control processes buried on a website is not enough.
The Trustmarc suggested is not a silver bullet. No mark can guarantee the truth or remove politics, judgement or bias from knowledge production. But it could help users ask better questions, such as: who produced this? Who funded it? What process did it go through? Has it been altered? Is it archived? What level of review did it receive? Can I cite it with confidence?
AI disclosure should be reflective, not punitive
The second practical solution came from Busara. In the session convened by Mareike Schomerus, Busara introduced sIfA — the Statement of Intellectual Fellowship and Accountability — developed by Engy Saleh and Mareike Schomerus. The sIfA tool is an open-source, browser-based tool designed to help researchers document how humans and AI interact in producing knowledge. It is built around the 14 CRediT contributor roles and allows users to declare AI involvement across different stages of research, generate a contribution table and produce a visual “sIfA figure”.
sIfA avoids the trap of false precision. It does not ask researchers to declare that “AI contributed 37%” to a report. I think that we would all agree that would be absurd. Instead, it invites reflection: where was AI used? Was it light or extensive? Was it generative or non-generative? Did it shape core decisions or merely assist with editing? What tools were used? What audit trail exists?
The session discussion captured the cultural challenge well. Many organisations still operate with a “school teacher” attitude where AI use is either forbidden, hidden or treated as something shameful. But this leads to secrecy, not integrity. Participants argued for a different culture in which AI is used as a creative companion, disclosure becomes a habit, and transparency is focused on learning.
If AI disclosure becomes a compliance ritual, people will game it. If it becomes punitive, people will hide. But if it becomes a normal part of the research process — like declaring authorship, funders, methods or conflicts of interest — then it can help think tanks learn how AI is shaping their work.
There is also a historical reason to do this now. The period between 2024 and 2027 may be one of the most opaque transitions in the history of knowledge production. AI is being adopted quickly, unevenly and often quietly. If we do not document how it is being used today, future researchers may struggle to understand how policy knowledge changed in these years.
If anything, sIfA offers one practical way to leave a record.
AI tools must be designed for legitimacy
The third practical contribution came from the session led by Sonia Jalfin of Sociopúblico and Noel Gruber of CIGI: “Can AI Become Trust Infrastructure?” The session asked whether AI itself can be designed to support trust.
As with most think tank affairs, the answer was: it depends.
Noel Gruber introduced a framework for responsible AI use in policy research built around three pillars: human leadership, transparency about AI’s role (see sIfA), and responsible analysis using credible sources (see the TrustMarc).
I found the phrase “human leadership” especially useful. It is stronger than “human in the loop”. Too often, “human in the loop” means the machine does the work and a person signs off at the end. Human leadership means researchers define the problem, set the boundaries, choose the sources, interpret the outputs, and remain accountable for the final judgement.
Sonia invited participants to prototype AI agents for real think tank tasks: media monitoring, legislative tracking, internal research synthesis, conference travel preparation, lessons-learning interviews and collaboration mapping. These are not futuristic abstractions. They are exactly the kind of tasks think tanks already struggle to do consistently because they are labour-intensive, fragmented or poorly resourced.
But the session also made clear that usefulness is not the same as trustworthiness. By their design, AI agents can make mistakes at scale: they can trigger error cascades, rely on weak sources, obscure judgement behind a friendly interface, and appear more competent than they are.
This session introduced the idea of “trust stickers” to encourage teams to assess their agents against principles such as transparency, security, identity and voice, governance, auditability and human ownership. In other words, trust as a design requirement from the beginning.
Could the TrustMarc promoted by Policy Commons help design AI agents?
From principles to practical infrastructure
What, then, should think tanks do?
The discussions in Rabat suggest a practical agenda:
- Think tanks could develop an internal AI policy or constitution. Tony Bader’s proposal is a good starting point: keep it simple, values-led and operational. It should clarify what tools can be used, for what purposes, with what data, under what review conditions and with what disclosure expectations.
- Think tanks could make the disclosure of quality assurance and AI contribution a habit. If an output has been internally reviewed, externally reviewed, fact-checked, methods-reviewed or discussed in a seminar, say so. Policy Commons’ proposal for review metadata and trust marks points to a future in which this information can be standardised and machine-readable. And Busara’s sIfA offers a practical way to do this without pretending that AI use can be measured precisely. It can be used in publications, teaching, institutional reporting and research workflows.
- Think tanks should disclose funding, conflicts of interest and governance arrangements more clearly. Nick Cleveland-Stout’s work is a reminder that donor opacity is not a marginal issue. It directly affects credibility.
- AI agents should be designed with trust requirements from the start. Sonia Jalfin’s trust stickers or Policy Commons’ Trustmarc are simple but powerful ways to force design teams to ask: who owns this tool, what sources does it use, how is it audited, how are errors corrected, and what will users be told about AI involvement?
- The sector needs shared standards. None of the tools presented at the Conference – and there are others – should become the single answer, but together they suggest the outline of a sector-wide trust infrastructure.
The opportunity
AI has made it easier to produce content. I have used AI to piece together notes, audio recordings and presentations from all the sessions included in this article.
But it has not made it easier to produce trust. In fact, I think it has had the opposite effect. As information becomes easier to generate, trust becomes more important.
This is an opportunity for think tanks.
We have spent years explaining that evidence is not enough. Browse through the On Think Tanks archive – it is easier with AI – and you will find that: context, relationships, values, legitimacy, and politics matter too – if not more!
Now we must apply that same lesson to ourselves. It is not enough to produce good research; we need to make visible why it should be trusted.
I was worried the Conference’s main theme – “Trust” – would be too abstract. I was wrong. At the 2026 OTT Conference, the trust conversation became more concrete. sIfA gave us a way to document human–AI collaboration. Trustmarcs gave us a way to imagine verifiable content provenance. Trust stickers gave us a way to design AI tools responsibly. AI constitutions gave us a way to govern internal practice. AI-enhanced search reminded us that discovery is now part of credibility. Financial disclosure reminded us that old transparency problems have not gone away.
The challenge now is to put these pieces together.
The next phase of AI adoption in think tanks should not be about who can generate more content faster. It should be about who can build the most trustworthy knowledge systems — human-led, transparent, accountable, useful and open to scrutiny.