A PhD student types into an AI literature search tool: “How does online learning affect first-year students’ motivation?” Within seconds she has ten closely matching papers, each with a one-line summary. All ten go into her review. At a seminar, a lecturer asks why there is nothing before 2015, no dissertations, and nothing using “e-learning” rather than “online learning.” She cannot answer, because she does not know where or how the tool searched.
AI-powered search tools are genuinely useful, especially for exploration. The problem is not using them; it is treating them as if they were systematic searching.
How semantic search differs from keyword search
| Aspect | Keyword search (traditional databases) | AI semantic search |
|---|---|---|
| Input | Keywords, AND/OR/NOT operators, filters | Natural-language questions |
| Matching | Words in titles, abstracts and indexing terms | Meaning, even without shared words |
| Reproducibility | The same string gives the same results on the same date | Results can shift with model versions and phrasing |
| Coverage | You know which database you searched | Often built on one open index; scope not always clear |
| Ranking | By relevance, date or citations, as you choose | By model-computed relevance, less transparent |
| What it finds best | Papers using terms you thought of | Papers using terms you did not think of |
That last row is the case for using both: semantic search widens your view, keyword search gives you controlled coverage.
Common blind spots
- Index coverage: many tools search aggregated open indexes. Regional journals, books, dissertations and policy reports are often thin or missing.
- Language: papers in Vietnamese, Spanish, Chinese and other languages tend to rank low or not appear at all.
- Older work: foundational studies from before widespread digitisation have sparse abstracts and match poorly by meaning.
- Null and contrary findings: a question that carries an assumption (“the positive effect of…”) pulls in agreeing papers.
- One-line summaries: machine-generated summaries can be wrong, especially about study design and sample size.
A combined workflow for a literature review
- Explore (AI): ask three or four differently phrased questions, including the opposite framing (“no effect”, “negative impact”). Note every new term that appears in the results.
- Build a keyword string: use those terms to write an AND/OR search with synonyms (online learning OR e-learning OR distance education…).
- Search systematically: run that string in two or three databases suited to your field, plus regional sources and dissertation repositories.
- Chase citations: for your five or six key papers, check their reference lists (backwards) and the papers citing them (forwards). Many AI tools offer citation maps that help here.
- Cross-check: are the AI-found papers in your systematic results? For any that are not, work out why your string missed them and fix it.
Reading results without being steered by them
AI tools rank by relevance to your question, so a biased question returns a biased list. Habits that help:
- Always ask one question that runs against your own hypothesis.
- Do not read the one-line summary instead of the actual abstract. Open the paper and check the methods for design and sample size.
- Watch the mix of study types: if everything is cross-sectional, look deliberately for experiments and systematic reviews.
- Check publication years: if nearly everything is from the last five years, go looking for foundational work.
Logging your search
Reviewers increasingly ask how you searched. For a systematic review, the answer must follow PRISMA, and AI searching is usually accepted only as a supplementary source, not a replacement for databases. For a narrative review a short paragraph is enough, but it should exist. A simple search log:
| Date | Tool or database | String or question | Filters | Results | Kept |
|---|---|---|---|---|---|
| 12 Mar | AI tool (name, version) | Exact question | 2010 onwards | First 50 | 9 |
| 13 Mar | Subject database | Full AND/OR string | English, Vietnamese | 412 | 31 |
When an AI tool is enough
You do not always need the full workflow. Finding a few papers for a talk’s opening slide, checking whether an idea has already been tried, or locating one specific methods paper are all good uses for an AI tool. The line comes when you intend to write “studies show” or “no research has examined…” Those sentences require systematic coverage, and no single tool guarantees it.
Beware of sources that do not exist
Search tools built on real bibliographic indexes rarely invent papers. General-purpose conversational AI assistants, however, when asked to “give me references on…”, can produce convincing citations that do not exist. These are different kinds of tools; do not use a chat assistant as a literature database. Any source from any tool goes into your paper only after you have opened the real thing.
A useful exercise for the review you are writing: take five papers you found with an AI tool and try to find them with your keyword string in a database. Any that do not appear show you which terms your string is missing, and fixing that is the most valuable edit you can make this week.
Câu hỏi thường gặp
Can AI literature search tools replace traditional databases?
Not yet, particularly for systematic reviews. They are excellent for exploration and surfacing new terms, but coverage, reproducibility and transparency of ranking are limited, so use them alongside databases.
What is semantic search in academic research?
It matches the meaning of your question rather than exact words, so it can return papers using different terminology. Its weaknesses are reproducibility and uncertainty about which sources were searched.
Can I use AI search tools in a PRISMA systematic review?
Usually only as a supplementary search, reported with the tool, date and exact query. The main search still needs reproducible keyword strings run in bibliographic databases.
How do I document a literature search done with AI?
Record the date, tool name and version, the exact question, filters, how many results you screened and how many you kept. This log helps you write the methods and answer reviewers.
Do AI literature tools make up references?
Tools that search real indexes rarely invent papers, though their summaries can be wrong. General chat assistants can generate non-existent citations, so always open the actual paper before citing it.
