Safeguards against GenAI hallucination in literature reviews

For many academics, generative artificial intelligence (GenAI) has already become a routine companion in the literature review process. Researchers increasingly rely on these tools to scan literature, map unfamiliar fields, summarise dense papers and write up reviews. 

But GenAI output can be unreliable. Hallucination happens when a system produces fluent but entirely fabricated references, quotations or claims. In a literature review, where credibility rests on traceable sources, hallucinated results risk undermining academic integrity. 

Prompt optimisation 

The first safeguard is to design prompts that indicate accuracy is more important than completeness. Instead of asking GenAI to “review the literature”, researchers should require the model to cite only papers it can confidently verify as real, and to say explicitly when it cannot. 

Three principles matter most. First, instruct the model to cite only articles it can confidently verify as existing in peer-reviewed journals, and to state explicitly when it cannot do so. Second, require highly specific bibliographic information, particularly DOIs and verbatim quotations with location. Third, restrict outputs to journals indexed in major databases. Authors should then independently verify all suggested references using the DOI to ensure the citations are authentic and accurate.

Example prompt:

You are assisting with a scholarly literature review.

Strict rules you must follow:

• Only cite articles that you are confident exist in peer-reviewed journals.
• If you are not certain a reference is real, you must explicitly state “I cannot verify this reference” and must not fabricate details.
• For each cited article, provide: full author list, year of publication, exact article title, journal name and DOI (if known).
• If you cannot provide a DOI with high confidence, say so explicitly.
• Quote one exact sentence from the article and indicate where it appears (page number or section). If you cannot quote verbatim, do not cite the paper.
• Restrict citations to journals indexed in Web of Science or Scopus.
• If fewer than x verified papers exist, return fewer papers rather than inventing any.

Task: Review the peer-reviewed literature on [your topic]. Focus on conceptual framing and key empirical findings.

Output format: For each paper, provide: full reference, quoted sentence with location, one-paragraph analytical summary written in your own words

Important: Accuracy is more important than completeness. Silence is preferable to fabrication.

Keyword expansion 

GenAI can help with keyword generation to prevent researchers overlooking important strands of the literature. Different disciplines often use different terms to describe similar ideas. For example, work on “greenwashing” may also be framed as “symbolic environmental disclosure”, “ESG decoupling”, or “decoupling between talk and action”. GenAI can help identify alternative keywords, related concepts and variations in terminology. Once a set of keywords has been developed, searches should be conducted directly in Google Scholar, Scopus or Web of Science. 

Specialised academic AI tools 

Not all AI tools hallucinate in the same way. A growing number of academic platforms integrate GenAI with curated scholarly databases. Tools such as Elicit, Consensus or similar systems interpret natural-language questions, search existing databases and return links to real papers, using GenAI only to summarise or compare results. 

These platforms can be particularly useful when entering a new research area. They help identify relevant studies, common methods and frequently cited papers while maintaining a direct connection to original sources.

• Using literature reviews to strengthen research: tips for PhDs and supervisors
• Streamline the literature review process with these tips
• A practical guide to writing a literature review

Literature visualisation 

Literature visualisation provides another powerful safeguard against hallucination. Tools such as Connected Papers or ResearchRabbit represent the literature in visual networks, showing how studies are connected through shared citations. Papers appear as nodes, with size often indicating citation volume and connections representing shared references or influence. 

Because the visualisations are built from existing citation data, every paper in the network is traceable. Visualisation also supports better judgement. Instead of defaulting to the most recent or most easily accessible papers, researchers can see which works anchor a field and how ideas have evolved. This approach is particularly helpful for understanding how a field is structured and for identifying foundational studies. 

A staged workflow to prevent GenAI hallucination

The most effective use of GenAI in research occurs when these tools are combined in a staged workflow. For example, a researcher might begin by using GenAI to expand keywords, then search for papers in established databases, explore relationships using visualisation tools and finally, return to it to help summarise selected articles. Each stage serves a distinct purpose, and crucially, the workflow itself becomes the safeguard.

GenAI can speed up literature reviews but risks introducing false references. The solution is not to avoid GenAI, but to use it strategically. Researchers should treat it not as an independent scholar, but as a powerful assistant for preliminary tasks. A disciplined approach ensures the literature review’s credibility remains rooted in traceable evidence.

Jie Zhang is an associate professor of accounting, while Lili Jiu and Yi Luo are assistant professors of accounting. All work at the International Business School Suzhou (IBSS) at Xi’an Jiaotong-Liverpool University in China.

If you would like advice and insight from academics and university staff delivered direct to your inbox each week, sign up for the Campus newsletter.