The Short Answer
Cite the source you read, not the tool that found it. You would not cite Google Scholar or your library catalogue for leading you to a paper, and the same goes for Elicit, Consensus, Semantic Scholar, ResearchRabbit, ChatGPT or Wonders.
| What you did | What you cite | Do you mention the tool? |
|---|---|---|
| Used an AI tool to find a paper, then read the paper | The paper | Not in the reference list. In methods or acknowledgments if your venue asks for AI disclosure |
| Read only the AI summary of a paper | Nothing yet. Read the paper first | n/a |
| Reported the tool's own output (a synthesis, a consensus figure, an extracted table) | The tool's output, using your style's format for generative AI | Yes, that is the citation |
| Ran a systematic or scoping review search with the tool | The papers | Yes, in the methods: tool, date, query |
| Got a reference from a chatbot that you cannot find anywhere | Nothing. It may not exist | n/a |
For the format of a citation to the AI tool itself, see how to cite AI tools.
If you do need to name the tool, in a methods section or a disclosure, see how to cite Consensus, Elicit and other AI research tools. For answer engines that list web sources, see how to cite Perplexity.
When You Might Need to Mention the Tool
There are a few exceptions where acknowledging your discovery method makes sense:
In your methodology section if you're conducting a systematic review or describing your search strategy. Reviewers and readers care about how you found your sources. This isn't a formal citation—it's methodological transparency.
If the AI tool provided the summary or synthesis you're using. This is different from finding a source. If Consensus gave you a synthesis such as "most of the studies found X" and you're reporting that synthesis, you're not citing the original papers—you're citing Consensus's analysis. That requires attribution.
If your institution or journal requires disclosure of AI use. Some venues want to know about any AI involvement in your research process, including literature discovery.
What Actually Needs a Citation
The paper, article, or source itself—always. Using AI to find it doesn't change your obligation to cite the original work.
If you read the abstract through an AI tool but then accessed and read the full paper, cite the full paper normally.
If you only read the AI-generated summary and never accessed the original, that's a problem. You shouldn't cite sources you haven't actually read. More on that below.
The "I Only Read the AI Summary" Problem
This is where things get ethically sticky. AI tools often provide summaries, key findings, or extracted claims from papers. It's tempting to cite the original paper based on that summary alone.
Don't do this.
If you cite a paper, you're representing that you've engaged with and understood that source. Citing based solely on an AI summary means you can't verify the summary is accurate, you might miss context that changes the meaning, and you're essentially citing something you haven't read.
Better options: access the actual paper, use the AI summary to decide if the paper is worth reading, or if you truly can't access it, see our guide on citing sources you can't fully access.
How to Describe AI-Assisted Search in Methods
If you're writing up a formal methodology—for a thesis, systematic review, or publication—here's how to handle it:
"Initial literature discovery was conducted using [Wonders/Elicit/Semantic Scholar], supplemented by manual searches in [databases]. AI-generated summaries were used to screen relevance; all cited sources were read in full."
This is transparent, honest, and increasingly normal in academic writing.
Common Scenarios
You asked a chatbot for paper recommendations. Chatbots can make up citations that don't exist. Verify every recommendation actually exists before citing. If the paper is real and you read it, cite the paper—not the chatbot.
You used Elicit to find papers and extract claims. Cite the papers you actually read. If you used data that Elicit extracted from the papers in your analysis, mention this in your methods and check the extracted values against the papers.
You used Consensus to see what "the research says" on a topic. If you're citing Consensus's synthesis (for example, the distribution shown by its Consensus Meter), you need to attribute that to Consensus. If you then read those studies yourself, cite them directly.
You used Semantic Scholar or ResearchRabbit to map a citation network. No citation needed for the tool. Cite the papers you read.
What a Citation on an AI Answer Actually Means
Two quite different things get presented the same way. One is a record the system retrieved and used: it searched an index, got back real entries, and showed you what it consulted. The other is a reference the model produced from its training, which may describe a paper that does not exist. On screen both arrive as a tidy author-year line under a confident paragraph, and nothing in the formatting distinguishes them.
Unless the tool says outright that a given citation came from a document it retrieved and opened, you cannot tell from the answer which of the two you are looking at. So the practical reading of any citation attached to an AI answer is: this is a lead, and it has not yet been checked. A citation is a pointer to a source, not evidence that the source exists or that it says what the answer claims.
That holds for the two mistakes people make in opposite directions. A reference that resolves to a real paper still does not mean the paper supports the sentence it was attached to. And a reference that you cannot find is not proof of fabrication either — it may be mis-transcribed, paywalled, or indexed under a different title. Either way, the next step is the same: go and look.
The Verification Step You Can't Skip
AI tools can surface papers that are retracted, misrepresent findings, or, in the case of general-purpose chatbots, simply don't exist. Before citing any source you discovered through AI:
- Resolve the identifier. Paste the DOI into doi.org or search the title in Crossref, your library discovery service or the journal's own site. A DOI that does not resolve, or that resolves to a different paper, is the fastest signal something is wrong.
- Access and read the actual source. The record existing is not the same as the record saying what you were told.
- Check the claim against the text. Find the passage the answer is standing on, and confirm the paper states it and states it for the population, period or outcome you are writing about.
- Check for a retraction or correction. The publisher's page and the Crossref record both carry retraction and correction notices.
This takes time. It's also non-negotiable for credible academic work.
The exact formatting for your citations depends on your required citation style. Check the specific guide for your discipline below.
Sources
- American Psychological Association, Citing generative AI in APA Style: Part 2, AI as a search engine (9 September 2025)
- MLA Style Center, How do I cite generative AI in MLA style? (Updated and Revised) (13 August 2025)
- Elicit Help Center, Citing Elicit, search methodology, and using Elicit's content in your own work
- Consensus, Consensus LibGuide for academic research
- Science, Change to policy on the use of generative AI and large language models
- ICMJE, Recommendations: Defining the Role of Authors and Contributors, section on AI-assisted technology
Accessed 21 September 2026.



