Think Tank Research With AI: Protecting Citation Integrity
Policy research organizations sell one thing above all else: the confidence that what they publish is true and traceable. AI can speed up literature reviews, data work, legislative summaries, and the stream of op-eds and explainers that follow every major report. It can also introduce sources that do not exist, quotes nobody said, statistics pinned to the wrong study, and citations that are real but do not support the sentence they sit under. This guide explains how those failures happen, why they are uniquely damaging for think tanks and state policy centers, and how to build a verification workflow, disclosure policy, and staff guidelines that let your institute use AI without putting its name at risk in front of legislators, journalists, and funders.
A think tank's footnotes are not decoration. They are the product. When a legislative staffer cites your brief in a committee memo, when a reporter quotes your estimate on deadline, or when a foundation program officer forwards your report to a board, each of them is borrowing your credibility. They rarely check every source themselves. They trust that you did. That trust is built over years of careful work and can be damaged in an afternoon when someone discovers that a cited study does not exist.
This is no longer a hypothetical. In May 2025, the federal government's first "Make America Healthy Again" report was found to contain citations to studies that do not exist, along with references that mischaracterized real research. PolitiFact summarized the findings: several cited titles could not be found at all, others did not say what the report claimed, and reporters noticed markers in some URLs associated with an AI tool. The citations were formatted correctly, named reputable journals, and in some cases included realistic-looking identifiers. That combination, polished format and absent substance, is the signature of an AI-generated reference, and it is exactly what a busy reviewer is least likely to catch.
For policy research organizations, the stakes are sharper than for most nonprofits. Your audience includes opponents who read your work specifically looking for errors. A fabricated citation does not just embarrass the author of one brief. It hands critics a reason to dismiss everything your institute has published, and it gives allies a reason to stop citing you. Because many think tanks operate as 501(c)(3) research institutes whose public education work depends on being seen as objective and fact-based, a reputation for sloppy sourcing also undermines the very characteristic that distinguishes your research from advocacy.
None of this means policy shops should avoid AI. The tools are genuinely useful for scanning large literatures, summarizing a 400-page bill, cleaning datasets, and repurposing a long report into the testimony, explainers, and op-eds that actually reach decision-makers. The question is how to capture those gains while making fabrication structurally difficult rather than merely discouraged. This article covers how citation failures happen, a source-first drafting method, a practical citation-checking routine, provenance logs, review tiers matched to product type, disclosure policy, and guidelines for staff and visiting fellows.
Five Ways AI Breaks a Citation
"Hallucination" is the word everyone uses, but it lumps together several different failures that require different defenses. A large language model generates text by predicting plausible sequences of words. A citation is a very predictable shape: author names, a year, a title that sounds like the field, a journal, volume and page numbers. The model can produce that shape fluently whether or not a real document sits behind it. Understanding the specific ways this goes wrong helps your team know what to look for. For a broader primer on why models do this and the general techniques for limiting it, see our guide to reducing AI hallucinations.
The research on this is sobering. A study published in Scientific Reports in 2023 asked ChatGPT to write short literature reviews on 42 topics and checked the 636 resulting citations. About 55% of the citations from GPT-3.5 and 18% from GPT-4 were fabricated outright, and a substantial share of the real ones contained errors. Newer models are better, but later work keeps finding the same pattern. A 2025 study in JMIR Mental Health found that roughly one in five citations GPT-4o produced for literature reviews were fabricated, with rates varying widely by how well studied the topic was. That last detail matters for policy shops: the less mainstream your subject, the more likely the model is to fill gaps with invention.
1. The phantom source
A reference to a document that does not exist
Real authors, a real journal, a plausible title, and no such paper. These are the easiest to catch with a lookup and the most damaging when they are not, because they suggest nobody opened the source.
2. The fabricated quote
Words put in the mouth of a real person or document
A model asked to "support this point with a quote from the committee chair" may invent one. Quotes from legislators, agency officials, and court opinions are especially risky because they are checked by people who were in the room.
3. The misattributed statistic
A real number pinned to the wrong source, year, or population
The figure may be roughly right but come from a different survey, a different state, or an earlier year. In policy work, a 2019 baseline presented as current, or a national rate presented as a state rate, can change the conclusion.
4. The misgrounded citation
A real source that does not support the claim
The hardest failure to catch. The paper exists and is on topic, but it does not say what your sentence says, or it says something more qualified. Simple existence checks pass it every time.
5. The drifted summary
Accurate at first, distorted through repeated reuse
A report's careful finding ("associated with a modest increase in some counties") becomes, after AI-assisted summarization into an explainer, then an op-ed, then a social post, "proven to increase." Each step seemed fine. The cumulative effect is a claim the original research never made, now circulating under your institute's name. Derivative content is where many think tanks are most exposed, because it gets the least review.
It is tempting to assume that tools which search the web or a document library, often called retrieval-augmented tools, solve this. They help, but they do not eliminate the problem. When Stanford researchers tested commercial legal research tools built on retrieval, they found the tools still produced incorrect or misgrounded answers in roughly one out of six queries or more. The researchers specifically flagged "misgrounded" responses, where the tool describes something plausibly but cites a source that does not back it up. Retrieval shifts the failure from phantom sources toward misgrounded ones, which are harder to spot. Your verification process has to be built for that.
Why the Reputational Math Is Harsher for Policy Shops
Most nonprofits that publish a flawed statistic face a correction and some mild embarrassment. Policy research organizations face something closer to a credibility test, because the people who read their work most carefully are often the people who disagree with it. A state policy center testifying against a tax proposal can expect the bill's sponsors, opposing think tanks, and industry associations to read every footnote. Finding a phantom source gives them a talking point that has nothing to do with the merits: "their own citations are made up."
Courts have shown how unforgiving this can be. In Kohls v. Ellison, a federal court in Minnesota excluded an expert declaration filed in defense of the state's election deepfake law after the expert, a respected researcher, acknowledged that an AI tool had introduced citations to nonexistent articles. The court accepted that the errors were unintentional and still declined to rely on the declaration, reasoning that the fabrications undermined its credibility as a whole. The same logic applies informally in every committee hearing room. A legislator who learns that one citation in your testimony was invented has little reason to trust the rest. The broader legal system is tracking a steady stream of these incidents; a public database maintained by researcher Damien Charlotin catalogs court decisions addressing AI-hallucinated content, with the United States accounting for the majority of entries.
Journalists are another pressure point. Reporters increasingly treat fabricated citations as a story in themselves, and they know how to look for them. When a syndicated summer reading list that ran in the Chicago Sun-Times turned out to recommend books that did not exist, NPR and others covered it widely, and the freelance writer responsible publicly acknowledged that he had not checked the AI's output. For a think tank, the equivalent headline is far worse because accuracy is the core of your value proposition, not a side feature.
Funders feel this too. Foundations that support policy research are buying influence on public debate through credible evidence. A program officer who has to explain to their board why a grantee's flagship report was corrected for fabricated sources is unlikely to forget it at renewal time. The reputational cost lands on the institution, not just on the individual analyst, which is why citation integrity needs to be an organizational system and not left to each author's diligence.
A brief note on 501(c)(3) research and nonpartisanship
Many research institutes rely on the IRS treatment of "nonpartisan analysis, study, or research," which allows a 501(c)(3) to publish work on legislation without it counting as lobbying, provided the work offers a sufficiently full and fair exposition of the pertinent facts that readers can form their own conclusions. A publication can take a position, but unsupported opinion does not qualify. The governing definition sits in Treasury Regulation 56.4911-2, and practitioner guides such as the Bolder Advocacy fact sheet on nonpartisan analysis explain the details.
This is not primarily a tax article, and your counsel should guide any lobbying questions. But the connection is real: fabricated or misgrounded facts weaken the "full and fair exposition" that defines your research as research. And AI-generated derivative content, such as a summary that adds a direct call to contact legislators, can shift a piece from education into lobbying, sometimes in ways that matter for a 501(c)(4) affiliate's reporting too. Our guides to lobbying disclosure and AI and election-year nonpartisanship in AI content cover those boundaries in more depth.
Where AI Genuinely Helps in Research Production
Before building defenses, it is worth being clear about where AI earns its place, because a verification workflow that is too burdensome will simply be ignored. The common thread among good uses is that the AI works on material your team has already gathered and can check, rather than supplying facts from its own memory. When a model summarizes a bill you have uploaded, you can compare its summary to the text. When it "recalls" what a study found, you have nothing to compare it to until you go find the study yourself.
Literature scanning is a strong fit when used for discovery rather than citation. An analyst starting a project on, say, state child care subsidy cliffs can use AI research tools to map the landscape: which agencies have published on it, what methodological debates exist, which states have tried reforms. Our article on AI research agents for nonprofit strategy describes how these agents work. The output is a reading list and a set of questions, not a bibliography. Every item on the list must be retrieved and read before it can be cited.
Legislative and regulatory summarization is another high-value use. A model can turn a long bill, a fiscal note, or a proposed federal rule into a structured section-by-section summary in minutes. Paired with tools for tracking legislation, this lets a small state policy center follow far more bills than its staff could read in full. The safeguard is simple: every summarized provision should carry the section number it came from, so a reviewer can jump straight to the source text.
Good fits
AI works on sources you supply and can verify
- Section-by-section summaries of bills, fiscal notes, and rules you upload
- Discovery scans that produce a reading list, not citations
- Writing and debugging analysis code that a human reviews and reruns
- Restructuring, tightening, and plain-language editing of human drafts
- Checking a draft's claims against a supplied source packet
High-risk uses
AI supplies facts from memory
- Asking a chatbot to "add citations" to a finished draft
- Requesting quotes from officials, testimony, or court opinions
- Generating statistics, cost estimates, or state comparisons without data files
- Summarizing a study the AI has not been given in full
- Producing derivative content from a summary rather than the original report
Data analysis deserves its own caution. AI coding assistants are excellent at writing the code to merge Census tables, reshape state budget files, or run a regression. The risk is not usually that the code fails, but that it runs and produces a confident number based on a wrong join, a filtered-out category, or a misread variable definition. Treat AI-written analysis code the way you would treat code from a new research assistant: it must be read by someone who understands the data, rerun from raw files, and spot-checked against known totals before any number leaves the building.
Source-First Drafting: Make Fabrication Structurally Hard
The most effective single change a research shop can make is to reverse the order in which sources and prose come together. In a "prose-first" workflow, someone drafts claims (with or without AI) and then goes looking for citations to support them. That is exactly the workflow that invites phantom and misgrounded citations, because the pressure is to find something that fits a sentence already written. In a "source-first" workflow, the analyst assembles and reads the evidence base first, and the draft is built only from that packet.
Source-first drafting works with AI rather than against it. Once the packet exists, you can give the model the actual documents and instruct it to draft only from those materials, to tag every factual sentence with the document and page it relied on, and to say plainly when the packet does not support a point. The model's tendency to fill gaps does not disappear, but now every claim has a pointer that a reviewer can check in seconds, and an untagged claim is an immediate red flag.
The source-first workflow, step by step
Adapt to your product types; the order is what matters
- 1. Scope the question. Write down the research question and the claims you expect to need evidence for. This list becomes your checklist later.
- 2. Discover, then retrieve. Use AI search tools to find candidate sources, then download each one from its original publisher, agency, or database. A source you cannot retrieve does not enter the packet.
- 3. Read and annotate. A human reads at least the relevant sections of every source and notes the exact passages, tables, or figures that matter, with page numbers.
- 4. Draft from the packet. If AI assists with drafting, it works only from uploaded packet documents, with instructions to cite document and page for every factual claim and to flag unsupported points rather than fill them.
- 5. Close gaps by research, not generation. When the draft needs a fact the packet lacks, the analyst goes and finds a real source, adds it to the packet, and then writes the sentence.
- 6. Hand off the packet with the draft. Reviewers receive the sources alongside the manuscript, so checking is a matter of opening a file rather than searching the internet.
Consider how this plays out in practice. A health policy analyst preparing a brief on Medicaid unwinding in her state assembles state agency enrollment reports, a few federal datasets, and three peer-reviewed studies. She uploads them to an approved AI workspace and asks for a first-draft background section citing document and page. The model produces a clean draft with one sentence about national disenrollment trends that has no tag. That missing tag tells her the model reached beyond the packet. She either finds a federal source for the national figure or cuts the sentence. Either way, nothing unsupported reaches the reviewer.
Source-first drafting also improves the research itself. Analysts who read the evidence before writing are less likely to overstate findings, more likely to notice qualifications and limitations, and better prepared to answer hostile questions at a hearing. AI then accelerates the parts of the job that are genuinely mechanical, such as organizing, summarizing, and drafting, without being trusted for the part that requires judgment.
Citation Checking and Provenance Logs
Even with source-first drafting, every product needs a citation check before publication. The key is to check at the right depth. Verifying that a source exists catches phantom citations but misses misgrounded ones. Verifying that the source supports the specific claim catches both but takes longer. A good process does the fast checks on everything and the deep checks where the stakes are highest.
Level 1: Existence
Does the source exist exactly as cited? Resolve every DOI and URL, confirm title, authors, year, and publisher, and confirm government documents on the agency's own site. Fast, and it should cover every citation in every product.
Level 2: Support
Does the cited passage actually support the sentence? Open the source to the cited page, read the relevant passage, and confirm the claim matches in scope, direction, population, and time period.
Level 3: Fidelity
Does the claim reflect the source fairly, including its caveats? Check quotes word for word, recompute key numbers from underlying data, and confirm you have not dropped a qualification the authors considered important.
AI can assist the checking itself, with care. A useful technique is to give a model the draft and the source packet and ask it to list every factual claim alongside the passage it believes supports it, marking any claim it cannot match. This works well as a triage tool that points human checkers at the weakest spots. It should never be the final word, because the same tendencies that produce misgrounded citations can produce misgrounded confirmations. The human checker confirms; the model only helps decide where to look first.
A provenance log is what turns a careful process into a defensible one. It is a simple record, usually a spreadsheet or a structured table in your project management tool, that shows where each significant claim came from and who verified it. When a reporter or a legislative staffer asks "where did this number come from," the log lets anyone on your team answer in minutes. When an error does slip through, the log shows where the process broke, which is far more useful than blaming an individual.
What a provenance log should capture
One row per key claim, statistic, or quote
- The claim as it appears in the draft, with its location (section or paragraph)
- The source, with a stable link or file path to the retrieved copy, and the exact page, table, or timestamp
- Whether AI was involved in finding, summarizing, or drafting the claim, and which tool
- For calculated figures, the dataset, version or vintage, and the script or workbook used
- Verification level reached (existence, support, fidelity), the checker's name, and the date
Saving local copies of sources is part of provenance, not an afterthought. Government web pages are reorganized, datasets are revised, and links rot. If a state agency updates an enrollment report after your brief is published, you want to be able to show exactly what the figure said on the day you cited it. Many institutes already archive sources for their own reference; making it a required step in AI-assisted projects costs little and pays off the first time a number is challenged.
Human Review Tiers Matched to Product Type
Not every product needs the same scrutiny, and pretending otherwise leads to either gridlock or quiet shortcuts. A sensible approach sets review tiers by two factors: how much weight the product will carry with decision-makers, and how hard it would be to correct once it is out. Flagship reports and formal testimony sit at the top. Internal memos sit at the bottom. Derivative content, which is where AI is most heavily used and least reviewed in many organizations, often belongs higher than people assume.
Tier 1: Highest scrutiny
Flagship reports, legislative testimony, formal comment letters, expert declarations, fiscal or economic estimates
Every citation checked at the support level, all quotes and key statistics at the fidelity level, and all calculated figures independently reproduced. A full provenance log. Review by a second analyst plus the research director, and external peer review where your institute uses it. Testimony and anything submitted under oath or to a court gets an extra read specifically for citations, since that is where fabricated sources have caused the most visible harm.
Tier 2: Standard research review
Policy briefs, issue explainers, bill analyses, data notes
Every citation checked for existence, key claims and every statistic checked for support, and quotes checked word for word. Provenance log for statistics and quotes. One reviewer other than the author. Bill analyses should be checked against the current version of the bill, since AI tools may summarize an introduced version after amendments.
Tier 3: Derivative content review
Op-eds, blog posts, newsletters, social posts, talking points drawn from published research
Every factual claim traced back to the original published report, not to an intermediate summary, and checked for drift in strength or scope. Any new fact not in the parent report is treated as Tier 2. A designated reviewer, often communications staff working with the original author. For guidance on drafting leadership commentary, see our article on AI-assisted op-eds for executive directors.
Tier 4: Internal working products
Research memos, meeting prep, literature maps, internal summaries
Author responsibility, with a clear label that the document contains unverified AI-assisted material. The rule that matters most here is that internal products cannot be cited or quoted externally until their claims go through the appropriate tier. Many fabricated citations enter published work by way of an internal memo someone trusted.
Time pressure is the main enemy of any tier system. When a bill moves unexpectedly and a hearing is scheduled for tomorrow, the temptation is to let AI draft testimony quickly and skip the checks. Build for this in advance. Keep a vetted, regularly updated evidence library on your core issues, with sources already retrieved and annotated, so rapid-response work draws on material that has already been verified. Speed then comes from preparation, not from cutting verification.
Writing an AI Disclosure Policy for Published Research
Disclosure is where many policy shops hesitate. Some worry that saying anything about AI use will invite critics to dismiss their work as machine-generated. Others disclose so vaguely ("AI tools may have been used") that the statement conveys nothing. The more durable position is a clear, specific policy that describes how AI is used, what it is not used for, and how human verification works. Done well, disclosure is a statement of method, and method is something research institutions are already comfortable explaining.
Academic publishers have converged on some useful norms that translate well to policy research. Under the ICMJE recommendations followed by many medical journals, AI tools are not listed as authors, because authorship implies accountability. Humans remain responsible for everything in the work. Substantive use of AI in research methods, such as coding qualitative data or generating analysis code, is described in the methods section, while routine editing assistance typically does not require disclosure. Your policy can adopt similar distinctions. Our broader guide to AI content ethics and disclosure covers the general principles.
What a good policy includes
- A public statement of how the institute uses AI and how claims are verified
- Methods-section disclosure when AI shaped data processing, coding, or analysis
- A clear rule that named human authors are accountable for every claim
- A corrections policy that explains how errors are fixed and noted publicly
- Any disclosure rules that apply to testimony or court filings in your jurisdictions
Common disclosure mistakes
- Boilerplate so vague it implies nothing was checked
- Disclosing only after an error is discovered, which looks like concealment
- Characterizing fabricated citations as "formatting issues" when corrected
- Inconsistent practice across the (c)(3), the (c)(4), and individual fellows
The corrections policy deserves special attention because it is what people will judge you on if something goes wrong. When the MAHA report's fabricated citations came to light, the initial public response describing them as formatting issues drew additional criticism. A research institute that discovers a fabricated or misgrounded citation should correct it promptly, add a visible correction note explaining what changed and whether the finding is affected, and review whether the same workflow touched other products. Candor about the error, and evidence that your process caught or fixed it, is what preserves trust. It also helps to remember that courts and legislatures are adopting their own expectations for AI use in filings and testimony, so check the rules that apply where your staff submit work.
Guidelines for Staff, Fellows, and Affiliated Scholars
Think tanks have an unusual workforce for a nonprofit. Alongside full-time analysts, many rely on visiting fellows, nonresident scholars, university faculty who publish under the institute's name, research assistants, and interns. These contributors bring their own habits and tools, and they may not know your verification expectations. A fellow who would never invent a source in a journal article may still paste a chatbot's suggested references into a quick blog post without checking them. Your guidelines need to reach everyone whose work carries your name.
Keep the guidelines short enough to be read and specific enough to be followed. A two-page document that sits alongside your general AI acceptable use policy works better than a long manual. Include it in fellowship agreements and contributor terms so expectations are set before work begins, not after a problem.
Core rules for anyone publishing under your name
- Never cite what you have not opened. Every source in a published product has been retrieved and read, at least in its relevant section, by a named person.
- No AI-generated quotes. Quotes come from transcripts, recordings, documents, or interviews, and are checked word for word against the original.
- Numbers trace to data. Every statistic links to a specific table, dataset, or calculation, not to an AI summary.
- Use approved tools for unpublished work. Embargoed reports, confidential data, and draft testimony go only into tools the institute has vetted for data handling.
- Log substantive AI use. Note in the provenance log when AI helped find, summarize, analyze, or draft material.
- Report errors fast. Anyone who suspects a citation problem in published work flags it immediately, with no penalty for raising it.
Training matters as much as rules. A one-hour session where analysts watch a model confidently produce a plausible but nonexistent study on a topic they know well is more persuasive than any policy document. Let people see the failure modes on their own subject matter. Then show them the source-first workflow and the provenance log so they leave with a method, not just a warning.
Culture is the final layer. Research directors set the tone by asking "where did this come from" as a normal part of review, not as an accusation. Junior staff should feel comfortable telling a senior fellow that a citation could not be verified. Institutes where checking is seen as routine professional practice catch problems early. Institutes where it is seen as a sign of distrust tend to find out about problems from their critics.
Turning Verification Into a Credibility Advantage
Most institutes will approach this work defensively, as risk management. That is reasonable, but it misses an opportunity. As AI-generated policy content multiplies, legislators, journalists, and funders will find it harder to tell careful research from plausible-sounding text. Organizations that can demonstrate rigorous sourcing will stand out. Verification becomes part of your value proposition rather than a cost center.
With legislators and their staff, the advantage is reliability under pressure. Staffers remember which organizations give them numbers that hold up in floor debate. A state policy center that can answer a challenge to its fiscal estimate within an hour, by pointing to the dataset, the calculation, and the reviewer, builds the kind of reputation that gets it called before the next hearing. With journalists, the advantage is speed and confidence. Reporters on deadline favor sources who can provide the underlying document immediately. A provenance log makes that routine.
With funders, the advantage is demonstrable stewardship. Program officers are increasingly asking grantees how they use AI. A clear answer that describes your verification tiers, disclosure policy, and corrections record reassures them that their investment in your credibility is protected. It is also a natural topic for grant reports; our guide on building public trust in your AI use offers ways to frame that conversation.
A 90-day starting plan
Sequence the changes so they stick
- Weeks 1-2: Audit current practice. Ask staff and fellows, without blame, how they currently use AI in research and writing. Spot-check citations in a sample of recent derivative content.
- Weeks 3-4: Set the tiers and rules. Define your product tiers, the core rules for contributors, and the minimum provenance log fields. Choose approved tools for unpublished material.
- Weeks 5-8: Pilot on real projects. Run two or three projects through the source-first workflow and adjust based on what slows people down unnecessarily.
- Weeks 9-10: Train everyone. Hold hands-on sessions for staff, fellows, and communications teams, and update fellowship and contributor agreements.
- Weeks 11-12: Publish your policy. Post a public methods and AI use statement with your corrections policy, and brief your board and key funders.
Conclusion
Think tanks and policy research centers have always been in the business of trust. AI does not change that; it raises the stakes. The same tools that let a small state policy center analyze more bills, scan larger literatures, and reach more audiences can also slip phantom sources, invented quotes, and misgrounded claims into work that legislators and journalists rely on. Because these errors look polished, they are easy to miss, and because your critics read closely, they are likely to be found.
The answer is not to ban AI but to build a research process where fabrication is structurally difficult. Gather and read sources before drafting. Let AI work on material you supply, with every claim tagged to its source. Check citations at a depth matched to the product's stakes, and keep a provenance log that lets anyone trace any claim. Disclose your methods clearly, correct errors candidly, and make sure every fellow and contributor knows the rules before they publish under your name.
Institutes that do this well will find that verification is not just protection. In an information environment filling up with plausible but unverified text, demonstrable rigor is a reason for decision-makers to keep calling you. That is a credibility advantage worth building now, before the first fabricated citation forces the issue.
Build a Research Workflow You Can Defend
We help policy research organizations design AI verification workflows, provenance logs, disclosure policies, and staff training that protect the credibility of everything you publish.