Editor’s note: This article is longer than our usual posts — and intentionally so. We encourage you to read it in full.
In many parts of the world, public trust continues to face serious challenges: consensus is increasingly difficult to reach, and evidence-based rational discourse is periodically marginalised or manipulated in public debate. Against this backdrop, the think tank community’s discussions of “trust” carry immediate practical urgency. An article by Andrea Baertl, as early as 2018, underscored the centrality of credibility: credibility determines whether think tanks are consulted in policy processes, gain access to reputable networks, engage with the media as experts, or attract funders. It is the bedrock of every form of influence a think tank can wield and of the public trust it seeks to earn.
Think tanks have traditionally built trust through professional reputation, expert endorsement, research quality, and policy impact. As AI technology rapidly enters the knowledge-production chain, the question of trust has grown considerably more complex. Data from the Chinese Think Tank Index (CTTI) system maintained by the China Think Tank Research and Evaluation Center at Nanjing University show that the three categories of outputs recorded in the system — internal policy briefs, research reports, and newspaper/online articles — now total nearly 90,000 entries (See Figure 1 below): more than 30,000 internal briefs, more than 26,000 research reports, and more than 30,000 newspaper/online articles. After the General Office of the Communist Party of China (CPC) Central Committee and the General Office of the State Council issued Opinions on Strengthening the Constructionof New-type Think-tanks with Chinese Characteristics in January 2015, Chinese think tanks entered a period of vigorous growth, becoming important “external brains” – that is, sources of external intellectual support – for policy consultation, theoretical innovation, public opinion shaping, and public diplomacy. Over the decade since, CTTI-Indexed think tanks have produced an annual average of approximately 2,744 internal policy briefs, 2,249 research reports, and 2,975 newspaper and online articles — testifying to their formidable policy-research capacity and the vitality of Chinese think tanks as producers of policy knowledge.
In the AI era, think tank research efficiency has improved dramatically. But as the volume of policy research output continues to climb, how can quality and credibility be guaranteed in a systematic way?
On one hand, AI can help think tanks process information faster, scan the literature more broadly, and respond to policy questions more nimbly. On the other hand, it may also amplify the risks of faulty citation, data bias, algorithmic hallucination, methodological opacity, and blurred accountability. When any institution can rapidly generate apparently professional text, think tanks need to rethink their distinctive value. In think tank research, more careful attention must be paid to organising reliable evidence, forming considered judgements, and ensuring that the knowledge-production process can withstand scrutiny. We should therefore approach the question of trust in think tank research from a broader and more fundamental perspective. In this article, I refer to “AI trust infrastructure” as the full set of foundational conditions, organisational arrangements, and institutional mechanisms that think tanks need to build as AI becomes embedded in knowledge production to safeguard the credibility of their research. It comprises four dimensions:
- Culture and capability;
- Institutional norms;
- Knowledge-resource foundations; and
- Methods and tools.
Together, these dimensions aim to ensure that AI-assisted think tank research generates reliable evidence, transparent processes, auditable methods, manageable risks, and traceable accountability.
Figure 1: Outputs of CTTI-Indexed think tanks in three categories
Three relationships of trust in think tank research
Once AI becomes embedded in think tank research workflows, the question of trust involves at least three distinct relationships:
- External trust: The trust that policymakers, funders, partners, media, the public, and other stakeholders place in a think tank’s knowledge products. The persuasiveness of research conclusions matters, of course, but these audiences also care whether the process by which those conclusions were reached is reliable. Are the facts in a report verifiable? Are data and references clearly sourced? Was AI involved in the analysis, writing, or judgment? If so, was the AI-generated content reviewed by human experts? These questions are increasingly becoming central criteria by which external audiences assess a think tank’s credibility.
- Internal trust: The understanding, acceptance, and judicious use of AI tools by researchers, managers, and communications teams. A more apt concept here is “appropriate trust”. In human-machine collaborative research, scholars adjust their reliance on and use of automated systems based on their assessment of those systems’ reliability. Over-trust leads users to accept AI-generated content uncritically, overlooking potential flaws; under-trust may cause them to dismiss potentially valuable AI judgements. In the think tank research context, “appropriate trust” means that researchers need to judge the boundaries of AI reliability across different tasks: which tasks can be handled with AI assistance to improve efficiency, which steps require human verification, which outputs can only be treated as leads or drafts, and which judgements must be made by human experts.
- Process trust: Whether a think tank can make its knowledge-production workflow verifiable. A think tank’s credibility cannot rest solely on institutional brand or individual expert reputation – even if, as Andrea Baertl’s paper found, these matter! Once AI is involved in research, trust must increasingly be process-supported: Are data sources traceable? Are methods replicable? Is AI use recorded? Have critical judgments been reviewed? Are lines of accountability clear? Process trust requires a set of mechanisms that can be organised, implemented, and inspected.
Building AI trust infrastructure in think tank research
1. Culture, mindset, and capability: competence for the appropriate use of AI
Whether technology can be used credibly depends, first of all, on whether researchers possess the capability to understand, evaluate, and direct AI. For think tanks, AI is entering the stages of literature retrieval, material organisation, data analysis, trend assessment, policy simulation, and output dissemination, and is thereby changing how researchers handle evidence, form judgements, and articulate the limits of their conclusions.
The Organisation for Economic Co-operation and Development (OECD) proposes a Framework for Trustworthy Use of AI in the Public Sector, listing “skills and talent” as one of the enabling elements, highlighting the importance of AI-related capabilities within public institutions. This insight applies equally to think tanks. For them, building the right culture and mindset, and cultivating AI literacy among staff, are both aimed at helping researchers develop a capacity for “appropriate use of AI” – neither instinctively resistant because of AI’s hallucinations, biases, and opacity, nor excessively reliant because of its powerful generative speed and breadth.
AI-generated content has no fixed status in itself. It can be treated as a lead, a draft, a piece of evidence, an analytical result, or even (mis)used directly as a conclusion. Different modes of use produce entirely different trust outcomes. If researchers lack basic AI literacy, fluently generated text is easily mistaken for reliable analysis; if they see AI only as a source of risk, they may miss its value in information integration and methodological innovation. Think tanks need to build an organisational culture and mindset that recognises AI’s strengths and limitations. Effective calibration depends on humans receiving accurate feedback about the reliability of AI tools, so that “appropriate dependence” behaviours can take shape.
One clear observation from Chinese think tank practice is that researchers’ acceptance of AI is uneven. Senior researchers typically possess stronger policy experience, domain knowledge, and judgment, and they place greater weight on research reliability, political sensitivity, and lines of accountability. Their caution towards AI is sometimes a product of unfamiliarity with the tools, and sometimes a clearer understanding of the costs of misjudgement in policy research. Younger researchers tend to be more familiar with digital tools and more willing to experiment with large language models and intelligent analysis methods, but they may not always be able to convert technical skill into high-quality policy research.
The goal of culture and capability infrastructure can therefore be summarised as fostering “appropriate adoption”: establishing cross-generational, interdisciplinary human-machine collaboration mechanisms; encouraging researchers to experiment with AI while retaining sensitivity to evidence, method, and accountability; and advancing technical capability into think tank research workflows while emphasising that policy judgement and value judgement must remain the responsibility of human beings. This is especially important for Chinese think tanks. Many have long conducted research by relying on expert experience, policy understanding, and contextual knowledge. The introduction of AI does not diminish the value of those capacities; on the contrary, when combined with data, modelling, and algorithmic capabilities, it can enable existing advantages to be leveraged to an even greater degree.
2. Institutional norms and risk governance: building organisational consensus
In many think tanks, AI use is already happening, but the relevant governance rules have not yet fully taken shape. Researchers use large language models to organise literature, generate summaries, translate materials, draft reports, analyse policy texts, and even conduct preliminary scenario exercises — practices that are often individual and implicit. Management may not know which projects use AI, which data have been input into external models, which judgments have been influenced by machine outputs, and which content has been verified by humans. This situation creates a new kind of organisational risk: AI may appear to enhance individual efficiency while reducing the institution’s visibility over the research process as a whole. For think tanks that depend on professional reputation, invisible processes mean untrackable accountability. If an output contains factual errors, inaccurate citations, biased judgments, or leaks of sensitive information, it becomes very difficult for the institution to pinpoint where the problem occurred or to repair trust in a timely manner.
The core of norms and governance is to convert dispersed individual research into organisational processes that can be discussed, reviewed, and accounted for. Think tanks need to clarify the boundaries of AI use, covering: task classification (which tasks may use AI assistance, which require caution, which prohibit external models); data protection (which materials may be input, which sensitive or unpublished information must be protected); traceability (whether AI-generated content is labelled, whether key steps are logged); and accountability (the division of responsibility among researchers, project leads, and the institution).
At the 2026 OTT Annual Conference, discussions on think tank AI use raised the idea of developing an “AI constitution”; Chinese think tanks are likewise actively exploring common principles around AI values and codes of conduct. On one side, China has in recent years developed a series of external regulatory frameworks for generative AI services and AI safety governance. The Interim Measures for the Administration of Generative Artificial Intelligence Services, which came into force in August 2023, explicitly requires that service providers “adopt effective measures, based on the characteristics of the service type, to improve the transparency of generative AI services and to enhance the accuracy and reliability of generated content”, while establishing a principle of balancing development and security and combining the promotion of innovation with law-based governance, and implementing classified and tiered regulation of AI services. Subsequently, the Measures for the Labelling of AI-Generated and Synthetic Content, issued in March 2025, further established a closed-loop management system of “label at generation, verify upon dissemination”, providing an institutional solution to the governance challenge of identifying and tracing AI-generated content. For think tanks, these frameworks provide an overall benchmark for formulating internal institutional rules – particularly when think tank research involves public policy, social governance, industrial development, international relations, and sensitive data, all areas in which national-level governance requirements shape how institutions define the boundaries of AI use.
At the same time, the Chinese think tank community is actively exploring responsible think tank research. In September 2025, the organising committee of the Sixth International Symposium on Think Tank Science and Engineering formally issued the Initiative on Responsible Think Tank Research (informally known as the “Ten Principles for Think Tank Research”). Article 6 of Part II of the initiative, “Codes of Conduct for Think Tank Research,” explicitly states: ‘Think tanks should establish management systems and standard procedures covering every stage of research; the research process should have internal review mechanisms, with external experts introduced as necessary, and research ethics must be upheld.’ The credibility of think tank research must be guaranteed through standardised processes, transparent use of technology, and rigorous internal review.
On the other side, external regulatory frameworks cannot substitute for a think tank’s own research norms. Regulatory documents and initiatives typically provide a “floor”; the formulation of internal rules requires further tailoring to the specific characteristics and details of each institution’s knowledge-production process – on such core questions as the human-machine division of labour, process traceability, data protection tiers, review mechanisms, risk-tiered authorisation, and accountability. These judgements must be made in the context of each institution’s mission, data types, research subjects, and public responsibilities.
3. Data-resource foundations: turning knowledge assets into trust assets
When AI participates in think tank research, an easily overlooked issue is the provenance of knowledge. AI can rapidly produce apparently perfect answers, yet may be unable to explain its basis; it may summarise large amounts of information but include outdated material, erroneous citations, and logically unsupported inferences. The quality of the data resource foundation directly determines the trustworthiness of think tank research results. Research institutes, think tanks, and industry research institutions in China have accumulated, over years of sustained work, large volumes of policy texts, literature, statistical data, industrial data, case databases, expert outputs, and research archives. These resources constitute an important stock of original knowledge assets. But in the AI era, whether assets can be converted into capability depends on whether they can be effectively organised, governed, and called upon. Dispersed materials, non-standardised data, texts whose sources cannot be traced, and literature without metadata are ill-suited to support credible AI-assisted research.
At the Chinese think tanks session at the 2026 OTT Conference, Senior Researcher Liu Xiwen from the National Science Library, Chinese Academy of Sciences, elaborated on the concept of building an “AI-ready” knowledge-resource repository. Using scientific literature and data resources as examples, he proposed that AI-ready data resources should include: high-quality base data; value-added data with multi-dimensional tags; standardised data processed and governed by specialists; semantic data composed of multi-granular semantic knowledge units; and graph data forming heterogeneous information association networks.
Only when data and literature resources are clearly traceable, structurally stable and unified, quality-assessable, semantically recognisable, permission-manageable, sustainably updated, and effectively linkable to actual research questions can the knowledge resources a think tank has accumulated over the years become trusted assets in the AI era.
The National Science Library (NSL) of the Chinese Academy of Sciences (CAS) serves as a vital information hub, providing literature support, strategic intelligence, and public platforms that drive independent innovation in the natural sciences, interdisciplinary fields, and advanced technologies.
To support these objectives, the NSL has developed nearly 30 specialised databases encompassing intelligence monitoring, research output, innovation activity, and domain-specific knowledge, totalling nearly 1 billion records. By leveraging scientific literature data through advanced content organisation, mining, and ontology-based knowledge extraction, the NSL has created intelligent service engines. These engines power diverse applications—ranging from knowledge retrieval and Q&A to evidence synthesis, reasoning, prediction, and situational awareness—enabling deep mining and value creation from scientific literature.
Once a think tank possesses large data resources, the more important task is to convert this dispersed information into a searchable, analysable, and explicable knowledge system, and to establish a knowledge-service system that integrates think tank-wide collaboration organically into research and decision-making processes. Think tank data-resource construction, therefore, needs to shift from material accumulation to knowledge organisation, from literature preservation to intelligent usability, and from standalone databases to a knowledge foundation that supports human-machine collaborative analysis.
Future differences between think tanks will increasingly depend on two capabilities: whether they can effectively use advanced AI tools, and whether they possess a reliable, traceable knowledge foundation suited to policy research. The former is readily available and achievable; the latter requires long-term accumulation and sustained governance. What is truly difficult to replicate are the literature resources, domain databases, expert knowledge networks, and knowledge-organisation capabilities that an institution has built up over many years. From this perspective, building the data-resource foundation is not only a technical matter but also a matter of trust. How a think tank manages its materials, labels data sources, updates its knowledge base, and handles conflicts among different sources affects the credibility of AI-assisted research.
4. Methods and tools: reshaping think tank research and service workflows
Many think tanks’ adoption of AI tools typically begins with a focus on localised efficiency gains. Writing, summarising, translating, polishing, and compiling meeting notes are among the earliest and most accessible AI use cases for most think tanks. As AI further penetrates core research stages – information sourcing, problem identification, trend assessment, policy simulation, quality review, and output dissemination – the changes it brings begin to exceed localised efficiency gains and move towards a foundational iteration of the think tank’s knowledge-production paradigm.
The policy problems think tanks face are typically complex, open-ended, and highly uncertain. A single policy issue often simultaneously involves multiple interests, institutional constraints, data gaps, value conflicts, and future risks. In building a methods and tools system, think tanks also need to adopt systemic thinking rather than adopting any given technology in isolation.
Researcher Liu Xiwen has proposed that clusters of intelligent agents can include four types of intelligence agents, each responsible for data production, information organisation, knowledge representation, and intelligence computation, respectively. Different agents can take on tasks such as information retrieval, literature analysis, knowledge extraction, data processing, trend assessment, report generation, dissemination, and conversion, forming modular collaboration within the research workflow. From the Chinese experience, in terms of research tools, some think tanks and research institutions have already begun integrating large language models, intelligent agents, and knowledge bases to conduct policy analysis, science and technology intelligence research, and industrial assessment. The National Academy for Development and Strategy at the Chinese Academy of Sciences launched work on a “Strategic Competition Decision Support System” at the end of 2022, which decomposes complex problems into modules for intelligent collaborative simulation. Guided by the “Double Helix Methodology in Think Tanks,” the system breaks down the complex and systemic issue of ‘strategic competition in science and technology among major powers’ into three components – ‘science and technology profiling of major powers,’ ‘science and technology strategic-competition countermeasures,’ and ‘decision support for strategic competition’ – establishing an intelligent big-data platform and an auxiliary decision-making prototype system with a strategic-competition core. It provides decision-makers and think tank researchers with a one-stop integrated solution for dynamically and accurately assessing the state of science and technology competition, addressing three major research challenges in major-power strategic competition in science and technology: policy evaluation, situation analysis, and strategic simulation. These tasks are undertaken collaboratively by various research teams and intelligent technology modules, thereby integrating expert wisdom with an intelligent platform.
In terms of research methods, some Chinese think tanks are currently using digital technology to optimise their organisational structures and operating mechanisms. In the domain of field survey methodology, for example, the Shanghai Academy of Social Sciences, building on a long foundation of offline field research, has progressively developed a dedicated sample database for public opinion surveys, collecting 2.37 million data samples from offline field surveys covering different regions and social groups across the country, and on this basis establishing a cross-provincial field survey network. On top of this, the research team has introduced intelligent questionnaire systems, mobile terminal data collection tools, and analysis mechanisms linked to online public opinion data, enabling survey design, data collection, quality control, and results analysis to operate in an integrated platform.
But the more complex the methods and tools system, the higher the demands it places on trust. In knowledge production: Is AI’s trend assessment supported by evidence? What assumptions does policy simulation rely on? Has the model output been reviewed by experts? In system operation: How do agents transmit information between themselves? How does the system handle conflicts when different data sources disagree? In accountability attribution: If a step in the tool chain goes wrong, can the institution trace the source of the problem and define lines of accountability? These questions together determine whether an AI tool system is actually enhancing research credibility or creating new black boxes.
AI can make research more agile, richer, and more dynamic – but the capacity to make judgements amid value conflicts, to draw boundaries under uncertainty, and to find workable solutions among competing interests can only be shouldered by human researchers. Posing questions, interpreting results, and taking responsibility for final recommendations: these steps form the non-delegable core of human judgment in think tank knowledge production. AI’s real value depends on whether it is embedded in a research system that is human-led, evidence-constrained, methodologically transparent, and quality-controlled.
From AI application to trust infrastructure
Trust is developed and experienced within specific institutional environments, regulatory cultures, and histories of technology adoption. Once AI has entered think tank research, an institution’s value no longer rests solely on whether it can produce more reports, respond to more issues, or disseminate more content. More importantly, it rests on whether a think tank can make its knowledge-production process more transparent, more robust, and more accountable.
Observations from China indicate that AI is driving multi-level change in think tank research: it affects the capability composition of researchers, drives updates to organisational rules, depends on high-quality data and knowledge resources, and is impelling research methods to evolve towards greater intelligence, modularity, and collaboration.
AI will accelerate knowledge production. But it will also accelerate the exposure of trust risks. The truly influential think tanks of the future may not be those that adopted AI earliest, but those most able to make AI a credible knowledge-production mechanism. Moving from AI-application to AI-trust infrastructure may be precisely the key path by which think tanks can rebuild public trust amid technological transformation.
Read more: From AI anxiety to trust infrastructure.