Dr. R, a computational materials chemist at a European university, asked Scopus AI to summarise recent work on two-dimensional covalent organic frameworks. The tool returned a confident paragraph and six references. Five matched real Scopus records. The sixth paired a plausible author name with a journal she knew published virology, and the DOI pointed to an unrelated 2019 article. She spotted the error quickly because the field is small enough for her to recognise its journals. Most first-year doctoral students would not have noticed.
Scopus AI is Elsevier's generative-AI layer built on top of the Scopus citation and abstract database. A user asks a question in natural language; the system retrieves relevant Scopus records and uses them to ground a written answer. The answer arrives with citations attached, which makes it read like a labmate's literature briefing. The problem is that the citations come from a generated summary, and the summary is only as reliable as the retrieval step that fed it.
What Scopus AI Is and How the Expansion Works
Scopus AI began as a limited beta in 2023 and became commercially available in January 2024. Elsevier describes the tool as drawing on peer-reviewed literature indexed in Scopus, including articles, books, conference papers, and reviews. Generated answers are accompanied by source abstracts so readers can move from summary to original text. Since the launch, the company has broadened the feature set, adding follow-up questions and ways to explore related topics within a search session.
For subscribing universities, the practical change is that Scopus AI now sits inside a database many researchers already use. That lowers the barrier to adoption. A researcher who would not open a separate AI tool will try a search box that is already on the library's discovery page. The expansion is therefore less about new technology and more about placement: the tool moves from an experiment to a default option.
Independent Tests and the Citation Verification Problem
Early independent scrutiny arrived before the commercial rollout reached most campuses. A January 2024 review in The Scholarly Kitchen found the tool useful for broad topic discovery but cautioned that its answers depend on Scopus coverage and require subject knowledge to evaluate. A 2024 news feature in Nature surveyed AI academic search tools and reported that different systems miss known papers, surface weak references, and struggle with interdisciplinary work. Elsevier's product page shows the source-citation display and notes that users should consult the original abstracts.
A hallucinated citation is a reference that looks complete but does not match a real, retrievable document. The DOI may not resolve, or it may resolve to a different paper. The failure can be small: one wrong author, a swapped volume number, a title merged from two abstracts, or a DOI copied from another record. Because the generated prose is fluent, the error passes standard proofreading.
The main risk is concentrated in fields where Scopus coverage is thinner or where terminology crosses disciplines. A tool can retrieve an abstract from an adjacent field and attach a plausible citation that a specialist would never have chosen. In high-volume fields, the odds of a generated summary being useful are higher simply because there are more indexed papers to retrieve.
This creates a base-rate problem. Most outputs are good, but a researcher who uses hundreds of generated references without checking will eventually import an error. One fabricated citation in a grant proposal or submitted manuscript can trigger an integrity query, and the rest of the work is then read under suspicion. The exception is the error caught early by a subject expert who knows the expected journals, authors, institutions, and methods. That exception is not a safeguard; it is an argument for treating every output as unverified.
What This Means for Your Lab
A lab does not need to ban Scopus AI to make it safe. It needs a verification routine that happens before any reference enters a manuscript.
- Paste each generated reference into Crossref or the journal's DOI resolver before you store it in your reference manager.
- Open the abstract Scopus AI cites, not just the generated summary. The summary can flatten a study's limitations.
- Have early-career researchers bring one AI-generated answer to group meeting and check it against the literature out loud, so the checking habit becomes public.
- Keep a short note in your lab wiki about which AI tools are allowed for which tasks and who reviews the output before submission.
Publishers are dealing with the same problem from the author side. Manuscripts have been found to include references that do not correspond to real papers, and STM has pushed for disclosure rules around AI use. The pattern is now visible in both directions: tools that generate citations, and authors who import those citations without checking. The overlap has turned citation verification into a due-diligence step rather than a niche concern.
Institutions Are Teaching Verification, Not Prohibition
University libraries are increasingly adding Scopus AI to instruction sessions, partly because students are already using it. A research support librarian at a large UK university told me her default slide now reads: Scopus AI is a starting point, not a source. Similar guidance appears in library subject guides across institutions, often with a checklist for verifying generated references.
That pragmatic approach matches where journal policies on generative AI have landed. Rather than banning generative AI, editors are asking authors to disclose how they used it and to confirm that they checked citations. The working assumption is that disclosure plus verification is stronger than outright prohibition, because prohibition drives the practice underground.
Scopus AI is likely to become a standard part of discovery workflows before its accuracy problem is fully solved. That does not make the tool unusable. It makes the user's job longer by one step. Next time you run a Scopus AI query, copy the first citation that excites you into Crossref or the publisher's own page and confirm it exists. If it does not, tell your librarian. That single habit, repeated by enough labs, is what will turn the debate about citation hallucination into a narrower one about edge cases.
Photo by Brett Jordan on Unsplash