03 · What You Need to Know
“Access to a Paper” Can Mean Several Different Things
There Is No Single Universal Collection of All Research Papers
Published scholarship is distributed across many systems. Some articles are openly available from publishers or repositories. Others require subscriptions. Some disciplines rely heavily on conference proceedings, preprints, books, or specialized databases. Older publications may be digitized unevenly, while regional journals may have limited representation in major discovery systems.
Even large scholarly infrastructures do different jobs. Crossref, for example, provides extensive bibliographic metadata deposited by publishers and other members, including titles, dates, identifiers, funding information, licenses, references, and some abstracts. Crossref explicitly notes that it collects metadata rather than the full text itself, although its records can contain links to full-text locations.
A system that searches Crossref may therefore discover that an article exists without possessing the article itself.
Metadata, Abstract Access, and Full-Text Access Are Different
When researchers say an AI “found the paper,” several very different things may have happened.
Bibliographic access
The system can retrieve information such as the title, authors, journal, publication date, DOI, or other metadata.
Content access
The system can retrieve enough of the paper itself, such as the abstract or full text, to analyze what the study actually reports.
There are additional levels between these two. A system may see only a search snippet. It may obtain an abstract but not the methods or results. It may retrieve an open manuscript version rather than the final version of record.
These differences matter because the amount of accessible content limits what the AI can defensibly say about the paper.
Training on Scholarly Text Is Not the Same as Having a Research Database
A language model may have encountered scholarly material during training. That does not turn its parameters into a searchable bibliographic archive.
The model generally cannot provide a trustworthy inventory of every document represented in its training data. Nor should researchers assume that it can faithfully reconstruct a particular paper merely because some version of the paper, its abstract, or discussions of it may have appeared in training material.
This distinction helps explain why non-retrieval models can produce plausible but fabricated citations. In the 2026 OpenScholar evaluation, general-purpose models without retrieval showed poor literature coverage and high rates of fabricated paper titles in demanding scientific-literature tasks. Retrieval substantially improved performance.
Generative fluency is therefore not evidence of bibliographic coverage.
Even Very Large Scientific Retrieval Systems Are Collections With Boundaries
Specialized AI systems can connect language models to enormous scholarly corpora. OpenScholar, for example, was developed using a retrieval datastore containing 45 million open-access scientific papers.
Forty-five million papers is an extraordinary collection. It is still a defined collection, not every scholarly publication ever produced. Its authors describe the datastore as being built from open-access academic papers.
This illustrates a useful principle: whenever an AI literature tool claims broad scholarly access, ask what corpus or databases it actually searches. Coverage should be described, not imagined.
Open Access Improves Availability but Does Not Make the Literature Complete
Open-access papers are particularly useful for AI retrieval because their content can often be obtained without subscription barriers. Repositories and open scholarly infrastructures can therefore support very large research corpora.
Yet not every publication is open access, and open collections may differ by discipline, geography, publication type, language, and historical period. A retrieval system built primarily from openly available literature can consequently have systematic coverage gaps.
Those gaps matter if the missing literature differs from the accessible literature in ways relevant to your research question.
Paywalls Create a Separate Access Problem
A paper may be discoverable through metadata while its full text remains available only through a publisher subscription, institutional license, individual purchase, or other authorized access route.
Crossref's documentation makes this distinction explicit: metadata can include a full-text URL, but the presence of that URL does not guarantee access. The destination may still require a subscription, login, or license.
Whether a particular AI system can access and read paywalled research papers therefore depends on the system's authorized access and should not be inferred merely because it knows the paper's title or DOI.
Indexing Is Also Not the Same as Full-Text Access
A scholarly database can index a publication without providing unrestricted access to its full text. Likewise, a search engine may return a paper's title and abstract while directing the researcher to a publisher page for the complete article.
This distinction is particularly important when researchers ask AI to conduct a literature review. Discovering candidate papers, screening abstracts, extracting full-text methods, and synthesizing results are different stages with different access requirements.
An AI tool capable of the first is not automatically capable of the others.
Retrieval Quality Determines What the AI Actually Sees
Even when an AI system has access to a large scholarly corpus, it cannot place millions of papers into a single prompt. Retrieval systems identify a manageable subset of documents or passages judged relevant to the query.
That creates another potential source of omission. The relevant paper may exist in the corpus but fail to be retrieved because of terminology, indexing, ranking, query formulation, or retrieval limitations.
In other words, corpus coverage and retrieval recall are separate issues. Having a paper somewhere in the database does not guarantee that the system will find it for your question.
Literature-Retrieval Studies Show Why Completeness Cannot Be Assumed
Empirical evaluations reinforce this caution. A 2024 study compared general-purpose LLMs with the reference sets of human-conducted systematic reviews. Under the evaluated conditions, the models showed low recall and precision, generated fabricated references, and exhibited geographical and open-access biases.
Those results should not be generalized mechanically to every newer model or specialized retrieval system. AI literature tools have advanced substantially, and retrieval-grounded systems can perform much better. The broader lesson remains relevant: a fluent list of papers should not be treated as evidence that a search was comprehensive.
Recent biomedical retrieval systems continue to be benchmarked specifically against systematic-review reference sets because literature recall is an empirical property that must be measured, not assumed.
AI Can Know About a Paper Without Having Read the Paper
This distinction deserves particular emphasis. A model or retrieval system might know a paper's title, authors, DOI, publication year, and abstract while lacking the full methods, results, tables, supplementary files, or discussion.
Secondary sources can create an additional complication. The model may know what other authors have said about a paper without having access to the original paper itself.
Consequently, recognition is not evidence of access. The question of whether an AI can accurately describe a paper it has not accessed is especially important when the system begins supplying detailed claims about methods or findings.
No Single AI Search Should Be Treated as an Exhaustive Systematic Search
For exploratory literature discovery, AI-assisted search can be extremely useful. It can help formulate search concepts, discover terminology, identify candidate papers, follow citation relationships, and synthesize retrieved evidence.
A systematic review asks a harder question: did the search identify the eligible evidence according to a reproducible methodology?
That requires explicit databases, search strategies, eligibility criteria, dates, screening procedures, and other methodological decisions. Unless an AI tool provides and validates those capabilities for the particular workflow, researchers should not substitute “I asked the AI for relevant papers” for a systematic search.
Watch Out
If an AI says “the literature shows” or “all available studies indicate,” ask what literature it actually searched. Without defined coverage and a defensible search process, the system cannot establish that it has examined all relevant published research.
