The Challenge
The rise of generative artificial intelligence (AI) in research communities is making the research pipeline more efficient but also risks a credibility rupture. AI can support valuable work for researchers, but it also enables the rapid production of credible-sounding content that can be misleading or false. At a time of declining trust and growing misinformation, this raises a fundamental question for think tanks: How can audiences identify credible research?
As think tanks incorporate AI into research and dissemination workflows, the absence of clear norms creates risk for both producers and users of research. It becomes harder to assess what is evidence based, what has been responsibly produced, and what deserves trust from policy makers and publics alike.
CIGI’s Human Analysis Standard
CIGI has developed a standard to establish a common baseline for responsible use of generative AI in policy research, support more consistent practice around AI use, and provide a practical standard for think tanks and other knowledge-producing organizations.
The Human Analysis Standard (HAS) can guide and communicate a think tank’s commitment to the responsible use of AI in research and policy engagement. The proposed standard has a three-pronged approach to ensure that human oversight and analysis remain at the forefront of policy research.
HAS Principles
- Human in the lead: Human guidance, review and sign-off on research content where AI was involved.
- Transparency: Disclosure of AI use on research products and public access to the organization’s internal AI use policy.
- Responsible analysis: AI-assisted content anchored in credible sources and fact-checking.
HAS Wordmark
Visual elements accompany HAS on a publication or content page to communicate CIGI’s standard and commitment to responsible AI use.
We invite feedback and participation from other organizations that would like to be involved in the continual development of this standard and/or incorporate it into their publications.
Read CIGI’s Generative AI Policy for Research
CIGI’s policy sets out expectations for responsible use of generative AI in research, including transparency, accountability, research integrity, protection of confidential information and human review of AI-assisted outputs.
This policy reflects CIGI’s recognition that, while generative AI tools can enhance research productivity and innovation, they also introduce certain risks. This policy sets out the expectations and requirements for acceptable uses of generative AI in CIGI research and writing, and outlines the required disclosure framework for use of AI tools.
Read the Rationale for HAS Development
AI and Human Verification: A Rationale Behind the Development of CIGI’s Human Analysis Standard
Paul Samson, CIGI President
Until very recently, the analysis and writing required to publish a policy paper was usually time-consuming. A researcher might take days, weeks or even months gathering evidence, judging sources, testing arguments and writing. Review and editing were essential but typically represented a small share of the overall effort. Generative artificial intelligence (AI) is turning that ratio on its head. AI can now produce a polished article or brief, complete with analytics, data and recommendations, almost as quickly as one can type out the prompt.
But efficiently producing an analysis and knowing whether it can be trusted are very different things. In many cases, AI output can be excellent, but it comes with caveats. A case in point is OpenAI, which has long noted on its interface window that “ChatGPT can make mistakes. Check important info.” In contrast, you might see the opposite on an expert’s publication, along the lines of: “All errors and omissions remain the responsibility of the author alone.” It is an important difference. With AI, there is a heads-up that the output may contain questionable assumptions and contrived facts, or subtle errors. With human-produced work, authors take full responsibility for the content and put their reputation behind it.
As humans integrate the use of AI in their intellectual work and “share” the writing process, one constant emerges: regardless of the length of time it takes to generate the information, it must be subjected to an appropriate review process. To preserve trust in the publication process, a human must ultimately stand behind the shared result.
Given these developments, we must prepare for the fact that the economics of knowledge production have fundamentally changed. The focus of human effort is increasingly shifting from producing analysis to verifying it and taking responsibility for it. A 2026 paper by Christian Catalini, Xiang Hui and Jane Wu describes the broader dynamic: as generative AI sharply reduces the cost of producing outputs, the capacity to verify them emerges as the binding constraint.
It is important to note that certain verifiable fields can at least partly escape the emerging constraint. An AI-generated mathematical proof, for example, can be formally checked if the problem and rules are precisely specified. Software can often be tested in similar ways. But even mathematics faces a verification bottleneck. Mathematician Terence Tao has warned of potential “proof indigestion,” as AI-generated proofs accumulate faster than mathematicians can verify, explain and absorb them.
But most policy research is different. There is no mathematical test that can establish whether advice and considerations presented to a government are sound, whether evidence has been interpreted properly, or whether a recommendation properly weighs competing risks. Of course, AI can also assist with verification, but it should not eliminate the need for final human judgment and accountability.
Therein lies the challenge at the heart of the emerging AI economy of think tanks. The more analysis AI allows us to generate, the more important it becomes to establish what can actually be trusted by the users of our research, and who exactly stands behind it. There is also another point. Generative AI not only reverses the economics of producing analysis, but also the direction of responsibility (see Figure 1). If AI is used, as it will and should be, the AI disclaimer effectively becomes, “AI produced this for you, but you are responsible for checking its errors.”
This is why CIGI has developed the Human Analysis Standard (HAS): as a mechanism for transparency and to ensure that responsibility is explicitly given to a human.
Figure 1: The Changing Allocation of Human Effort in Policy Research
Note: Percentages are illustrative ranges of changing allocation of human effort rather than actual averages. See Barrick, Summers and Wood (2026); Li, Lin and Luo (2026).
The HAS does not seek to preserve an older model of research nor to limit the productive use of AI. It sets where human responsibility remains as production processes changes: human in the lead. The use of AI becomes mainstreamed, and transparent. Evidence remains grounded in credible sources and is reviewed, and real humans remain accountable for the result.
Established think tanks and other knowledge-producing organizations have built their legitimacy and credibility on a commitment to rigour, clarity and truth. As the knowledge production process evolves, those commitments must remain intact if we are to continue to impact policy and to be trusted by decision makers — and for the ultimate public good. CIGI’s AI standard does not call for organizations to ignore what is evolving, but rather calls for a baseline that ensures humans remain in the lead in the midst of it, so that these values and the legitimacy of our organizations endure, no matter the tools employed to help our mandates.