Researchers exploring the use of generative artificial intelligence for academic writing have found that while AI tools can assist in brainstorming search terms, they cannot be trusted to produce complete or accurate query strings without rigorous human oversight. The risks of inaccuracy, low recall, and fabricated citations make AI a supplementary tool rather than a replacement for traditional literature search methods.
What can AI actually do when generating search terms for academic writing?
Generative AI models like ChatGPT show promise in rapidly producing lists of keywords, synonyms, and related concepts that researchers might overlook during initial brainstorming. However, their capabilities come with significant limitations that affect the quality and completeness of literature searches.
Overview of AI performance in search term generation
| Aspect | Capability | Limitation |
|---|---|---|
| Speed | Generates dozens of terms in seconds | Terms may lack relevance to specific research questions |
| Synonyms & acronyms | Expands vocabulary quickly | May include incorrect or obscure terms |
| Query formatting | Structures searches for various databases | Formatting may not align with database syntax |
| Precision | Can retrieve relevant studies when well-prompted | Often misses relevant studies (low recall) |
| Source accuracy | Can cite references when asked | Frequently hallucinates citations and references |
| Critical thinking | Not present | Produces non-original, surface-level output |
Key insights on AI‑generated search terms
- AI tools excel at brainstorming and expansion of terms, synonyms, and acronyms that researchers might otherwise miss — sources: Clemson University and Chicago School.
- AI can format queries for different databases faster than a human researcher in some contexts — BCU Library.
- There is a trade‑off between precision and recall: AI may improve the former but significantly reduces the latter, making it unsuitable for systematic reviews that require exhaustive retrieval — same source.
- AI‑generated search strings are not yet reliable for high‑quality literature reviews; they require detailed human scrutiny and refinement — BCU Library.
- Inaccuracy and hallucinations are common: tools frequently generate false references and invented citations — ERIC, OUCI, SSRN.
- General AI models are not trained on specialized databases like PubMed, causing them to miss crucial studies in medical or scientific fields — BCU Library.
- Current LLMs lack critical thinking and produce non‑original content, a serious drawback for complex search logic development — OUCI.
Real‑world performance of AI in search term generation
The table below summarizes the main constraints and known failures of AI when used for this task.
| Factor | Finding | Source |
|---|---|---|
| Recall shortfall | AI‑generated searches miss many relevant papers | BCU Library |
| Citation hallucinations | Tools invent non‑existent references | ERIC |
| Database ignorance | ChatGPT not trained on PubMed or similar | BCU Library |
| Lack of originality | Output is derivative, not novel | OUCI |
| No accountability | AI cannot be listed as author | AMS |
| Copyright issues | AI‑generated text lacks originality for copyright | Publisso |
| Disclosure required | Use of AI must be reported in methods | AMS |
What are the critical constraints and risks when using AI for search term generation?
Human oversight remains essential
AI‑generated search strings are not yet reliable enough for high‑quality literature reviews. Every output must be rigorously reviewed and refined by a human researcher to ensure alignment with the research question — BCU Library.
Inaccuracy and citation hallucinations
AI tools frequently generate inaccurate references and “hallucinate” citations that do not exist. Researchers must verify every source and term produced — ERIC, OUCI, SSRN.
Multiple studies confirm that AI tools routinely generate citations that look plausible but are entirely invented. Researchers must cross‑check every generated reference against real databases and the original literature. Relying on AI‑generated citations without verification can lead to serious academic misconduct.
Database limitations
General AI models like ChatGPT are not trained on specialized scientific databases such as PubMed. This means they routinely miss relevant studies in medicine, biology, and other fields that rely on controlled vocabularies — BCU Library.
Lack of critical thinking
Current large language models lack the ability to think critically about search strategy. They produce non‑original, surface‑level outputs that do not incorporate nuanced understanding of the research domain — OUCI.
What ethical and publishing constraints apply when using AI?
No AI authorship allowed
AI tools cannot be listed as authors on research papers because they cannot take accountability for the work. Researchers must disclose any use of AI in the methods section — AMS and Publisso.
Disclosure is mandatory
Authors who use AI for search term generation must fully disclose this in their manuscript. Attribution of authorship carries accountability that AI cannot fulfill — AMS.
Copyright originality is lacking
AI‑generated text, including search term lists, does not meet the threshold of originality required for copyright protection unless substantially revised by a human author — Publisso.
Prohibited uses
AI should not be used to fill logical gaps in an argument, generate citations to support a predetermined statement, or copy generated passages verbatim without review — Publisso.
Always document exactly how you used AI, keep original prompts and outputs, and treat AI‑generated content as a draft that requires substantial human revision. This ensures transparency and protects against allegations of misconduct.
How has the guidance on AI use in academic search evolved?
The guidance from publishers, libraries, and scholarly societies has developed rapidly since 2023.
| Time period | Development | Source |
|---|---|---|
| Early 2023 | Initial warnings about AI hallucinations in references | ERIC |
| Mid 2023 | Libraries publish guidelines stressing human oversight | BCU Library |
| Late 2023 | AMS and Publisso issue authorship and copyright rules | AMS, Publisso |
| 2024 | Studies confirm low recall and critical thinking deficits | OUCI, SSRN |
| Ongoing | AI remains recommended only for rapid reviews, not systematic ones | BCU Library |
What is established and what remains uncertain about AI‑generated search terms?
| Established information | Information that remains unclear |
|---|---|
| AI can generate large lists of keywords and synonyms quickly | Whether AI can ever achieve recall comparable to a human expert |
| AI frequently hallucinates citations and references | The exact rate of hallucination across different domains and tools |
| Human oversight is mandatory for all AI‑generated search strings | The optimal balance between AI speed and human refinement effort |
| AI cannot be listed as an author | How disclosure requirements will be enforced by journals |
| AI output lacks copyright originality | Whether future models may overcome the critical thinking limitation |
| AI is unsuitable for systematic reviews | The precise conditions under which AI might be reliable for rapid reviews |
How do these findings impact researchers who rely on AI for search generation?
The consensus from library science, publishing ethics, and empirical studies is clear: AI is a valuable brainstorming tool but cannot replace the expertise of a human researcher in constructing comprehensive, accurate search strings. Researchers who use AI must treat its output as a starting point, not a final product. They must verify every term against controlled vocabularies and thesauri, cross‑check citations against real databases, and disclose their use of AI in the methods section.
What do authoritative sources say about using AI for academic search terms?
“AI‑generated search strings are not yet reliable for high‑quality literature reviews; they require rigorous scrutiny and refinement by humans to ensure alignment with research questions.”
— BCU Library guide on generative AI for searching
“AI tools cannot be listed as authors on research papers because they cannot take accountability for the work; researchers must disclose AI usage in the methods section.”
— American Mathematical Society (AMS) Notices
What does this mean for the future of AI in academic literature searches?
Given the persistent limitations in recall, accuracy, and critical thinking, AI is best suited for rapid reviews where time is limited, rather than for comprehensive systematic reviews that demand exhaustive retrieval of all relevant studies. Researchers planning to use AI should break down their research topic into functional components, iteratively refine generated keywords, validate output against existing literature and controlled vocabularies, and use specific prompting that explicitly frames the context as academic.
1. Break down your research question into core concepts. 2. Ask AI to brainstorm synonyms, related terms, and acronyms for each concept. 3. Review the output and add missing terms from your own expertise. 4. Cross‑reference all terms with database thesauri and subject headings. 5. Use the refined list to build your final search string manually. This approach leverages AI’s speed while maintaining human quality control.
Frequently asked questions
Can I use AI to generate search terms for my dissertation?
Yes, but only as a brainstorming tool. You must verify every term and document your use of AI in your methods section. Never copy AI‑generated strings without review.
Why does AI miss relevant studies in systematic reviews?
AI models are not trained on specialized databases like PubMed, and they prioritize precision over recall. This causes them to retrieve a narrower set of results than a human‑built search.
How can I reduce the risk of citation hallucinations?
Always cross‑check every AI‑generated citation against the original source in a trusted database. Do not rely on the AI’s confidence level.
Is it allowed to list ChatGPT as a co‑author?
No. AI tools cannot be authors because they cannot take accountability. Major publishers and scholarly societies explicitly prohibit this.
Do I need to disclose AI use in my paper?
Yes. Most journals now require disclosure of AI‑generated content in the methods section. Check your target journal’s specific guidelines.
Can AI‑generated search terms be copyrighted?
Generally no. AI‑generated text does not meet the threshold of originality for copyright protection unless substantially revised by a human author.
What is the biggest risk of using AI for search term generation?
The combination of low recall and hallucinated citations can lead to incomplete literature reviews and potential academic misconduct if citations are used without verification.
Is AI better than a human librarian for literature searches?
No. Human librarians and information specialists have domain knowledge and critical thinking that AI lacks. AI can supplement but not replace their expertise.
Should I use AI for a rapid review?
Yes, if time is limited. AI can quickly generate a broad set of terms, but you must still verify results against known literature and accept that some relevant studies may be missed.
Can I use AI to format my search query for PubMed?
AI can generate a query, but it may not be accurate. Always test the query in PubMed or use the database’s own advanced search builder.
Related stories
Sophie Marlowe is Culture & Lifestyle Editor at BritainNow.uk, covering arts, culture, lifestyle and society.