A reference can include authors, a title, a journal and a year without corresponding to a real paper. That is a fundamental problem for AI literature search: an answer may appear sourced while offering nothing that can be located. Walters and Wilder (2023) documented both fabricated citations and errors in real references in their evaluation of outputs from GPT-3.5 and GPT-4 at that time.
If those citations enter a review, proposal or paper, the cost goes beyond wasted search time. Claims may lack evidence, readers cannot verify them, and later work may proceed from a false premise. Presenting fabricated citations as genuine support may also constitute academic misconduct. Correct-looking reference formatting does not establish that a paper exists.
AI can search for literature when it can access real sources and answer from the returned material. The model helps interpret the question, organise queries and summarise information; retrieval tools supply records, links or accessible text. This gives the answer something to check, and raises a practical question: which connections suit which materials?
Three common routes for AI literature search
Begin with the location of the material and the task. Finding new papers, screening candidates and revisiting earlier reading are related activities, but need not use the same entry point.
| Route | Useful for | Examples |
|---|---|---|
| Connect online scholarly sources | Discovering papers and checking bibliographic or publication details | Wiley, CNKI-MCP, scholarly APIs |
| Use a dedicated AI search service | Moving from a research question to candidates and further screening | Elicit |
| Connect a local literature library | Revisiting evidence in collected papers, notes and annotations | Zotero-MCP |
MCP and APIs implement connections. MCP, the Model Context Protocol, provides a common way for AI applications to call tools; an API gives programs access to data. A local library can also connect through MCP, so the protocol is neither limited to online search nor the name of a particular database.
These routes can complement one another. Online sources support discovery, a local library helps revisit accumulated reading, and dedicated AI services provide an interface for searching and organising results around a question.
Literature search solutions to consider
Wiley: obtaining evidence from publisher content
Wiley’s Scholar Gateway connects scholarly content with AI applications. Its official description includes source metadata and DOI links, while full-text access may depend on institutional or personal subscriptions. References can therefore be checked against their publisher source.
Wiley’s Nexus MCP documentation describes natural-language search, date filters and full-text retrieval. It illustrates how a publisher supplies tools that obtain relevant publications or passages and retrieve further content within access permissions. The connection changes how material is reached, while licensing conditions continue to apply.
CNKI-MCP: making Chinese-language queries available to AI
CNKI-MCP is an independent, unofficial project for CNKI search and article-information retrieval, available through MCP, CLI and Python API access. Its API documentation covers combinations of title, author, journal and year, detail retrieval for abstracts and keywords, and journal-issue lookup.
A request to find a topic within a particular journal can become a query followed by organisation of actual results. This supports locating Chinese papers, checking their records and supplementing a reading list. The project does not download full texts, and relevant login or institutional authentication remains necessary.
It also includes Subagents and Agent Skills. The former supply specialised search roles; the latter guide query design, tool use and result checking. Alongside callable interfaces, the project provides methods for organising an agent’s retrieval task.
Elicit: organising search and screening around a question
Elicit provides researcher-facing workflows. Its keyword-search introduction describes generating queries from natural-language research questions; its API introduction describes returning structured titles, authors, abstracts and DOIs.
Such services suit researchers who want to inspect candidates and continue screening and organising them in one interface. Results can help narrow an initially broad question. Whether a study belongs in the project still depends on its population, method and relevance.
OpenAlex, Semantic Scholar and Crossref: different roles for scholarly APIs
Some solutions offer data interfaces for applications. They can supply AI workflows without themselves being configured conversational assistants. What each service returns determines the role it can usefully play.
OpenAlex search queries scholarly works through titles, abstracts and indexed full text, combined with filters. It supports broader discovery; searchable content does not mean every record has downloadable full text.
The Semantic Scholar Academic Graph API supplies paper searches, author information, references and citation relationships. A relevant paper can become a starting point for further exploration. Citation counts supply context but cannot alone establish relevance to a question.
The Crossref REST API supplies member-deposited scholarly metadata for DOI lookup and checking titles, authors and publication details. It helps identify a specific record from an incomplete citation. Establishing whether the paper supports a claim requires its relevant content.
Zotero-MCP: reusing accumulated reading
If relevant papers are already in a personal library, another online search may be unnecessary. Zotero-MCP connects local or web Zotero libraries, supports searching by title, author, tag and collection, and reads metadata, available full text, notes and annotations.
It addresses questions such as which collected papers discuss a mechanism or which earlier notes contain relevant evidence. The scope is accumulated material that can inform the current project. Full-text reading still depends on available attachments and text; a local library does not automatically imply a locally running AI model.
Choose according to what the tool can return
For minimum-wage and employment research, records identify candidate papers and abstracts support initial screening. Assessing employment measurement, comparison groups or assumptions requires the corresponding text. Retrieving a record and reading enough evidence for a judgement are separate achievements.
A task therefore guides the choice. Discovery requires coverage and filters; a known title or DOI calls for record checking; comparing earlier reading depends on accessible full texts, notes and annotations. A tool’s name cannot establish these conditions.
Keep source links and version information with results. They let researchers return to the material and identify which paper or version supports a discussion. A real paper does not make every generated summary correct; traceability provides a way to notice and correct mistakes.
Common questions about AI literature search
Does finding a paper mean AI can download its full text?
No. Record search, content reading and full-text downloading are different capabilities, subject to access rights and subscriptions. CNKI-MCP, for example, searches and retrieves article information without providing full-text downloading. Check what a service actually returns.
If I already have a DOI, do I need a keyword search?
Start by looking up the DOI and confirming the authors, title and publication details, then locate the relevant content. Crossref supports this type of record lookup. A DOI identifies a work but does not prove it supports a proposed claim.
Can a local library replace online searching?
A local library revisits collected material; online search discovers work outside it. Their coverage differs and they can be combined as needed. Failure to find a paper in a personal library does not establish that the research does not exist.
Continuing the research in Marx Max
Literature tasks can also develop around a project in Marx Max. Researchers describe a question, and AI uses available tools to find material and organise records and sources. Accessible texts can then be discussed alongside project notes. Current routes include CNKI searches and NBER working papers; this does not imply that every general solution described above is built in.
The next step depends on the research task. Methodological literature can inform analysis and comparison in empirical research. Selected papers can also become group-meeting presentation material. Retrieved sources, material actually read and the researcher’s own judgement can continue to inform that work.