Can Generative AI Detect Duplicate References in a Bibliography?

Can AI Find Duplicate References That Are Difficult to Recognize?

You combine references from Scopus, Web of Science, PubMed, Google Scholar, and several articles recommended by generative AI. The resulting bibliography contains hundreds of entries, and some publications may appear more than once.

Identifying exact duplicates seems straightforward. But what if one record uses an abbreviated journal title while another uses the full name? What if the same article appears with different publication years, missing DOIs, or slightly different titles?

You might ask ChatGPT or another generative AI tool to identify and remove duplicate references. The question is whether it can reliably distinguish repeated records from genuinely different publications, especially when the bibliographic information is incomplete or inconsistent.

AI Can Suggest Duplicate References, but Its Decisions Require Verification

How Duplicate Reference Detection Works and Where AI Can Make Mistakes

What Counts as a Duplicate Academic Reference?

A duplicate reference occurs when two or more bibliographic records represent the same scholarly work or publication record, depending on the deduplication rules being applied.

Duplicates frequently arise when researchers combine search results from multiple databases. The same journal article may appear in Scopus, Web of Science, PubMed, and other services, with each database exporting slightly different metadata.

However, two references that share an author, title keywords, or research topic are not necessarily duplicates.

Consider three situations:

The third category requires particular care. A preprint and its published journal version may describe substantially the same research, but they are distinct scholarly objects with potentially different content and identifiers.

Whether such records should be linked, consolidated for screening, or retained separately depends on the purpose of the bibliography and the applicable review methodology.

How Does AI Identify Potential Duplicate References?

When given a reference list, a generative AI model can compare textual similarities among entries. It may recognize repeated titles, similar author combinations, matching DOIs, or differences that appear to result from citation formatting.

For example, one reference might use the full journal name while another uses its accepted abbreviation. AI may infer that the records describe the same article if the titles, authors, and publication details also correspond.

However, a language model is not inherently a deterministic bibliographic matching engine. Its response may depend on how the references are presented, the length of the list, the prompt, and the model's ability to compare entries consistently.

Without an external matching procedure or validation step, the model may produce a plausible list of duplicates without systematically evaluating every possible record pair.

It may also fail to preserve record numbers or introduce errors while rewriting the bibliography.

What Makes Duplicate Detection More Difficult Than Finding Repeated Titles?

Bibliographic records often contain inconsistencies that complicate matching.

These examples are illustrative. The underlying challenge is deciding whether differences reflect alternative descriptions of the same publication or genuinely distinct research outputs.

Can AI Detect Duplicates When the DOIs Are Identical?

Matching DOIs provide strong evidence that two records refer to the same registered object, provided the identifiers have been accurately recorded.

For example, if two journal references contain the same verified DOI but differ in capitalization or punctuation, they are strong candidates for deduplication.

However, the DOI should be normalized before comparison. One record may contain the identifier as a URL, while another contains only the DOI string.

For instance, these forms represent the same identifier:

• 10.1038/s41598-023-41032-5
• https://doi.org/10.1038/s41598-023-41032-5

These examples identify the genuine Walters and Wilder (2023) article in Scientific Reports.

DOI matching is not sufficient for every deduplication task. Some publications lack DOIs, and incorrect DOI metadata can create false matches or conceal genuine duplicates.

The distinction between a valid DOI and the publication it actually identifies remains relevant here.

Can AI Incorrectly Merge Two Different Papers?

Yes. This is a false-positive duplicate decision, in which two distinct records are incorrectly classified as representing the same publication.

Imagine that a research team publishes two articles about AI-assisted feedback. Both papers have similar titles and overlapping authors, but one investigates student perceptions while the other evaluates writing performance.

An AI system relying heavily on title similarity might identify them as duplicates even though they report different investigations.

The consequences can be substantial. Removing one record could eliminate relevant evidence from a literature review or create an inaccurate bibliography.

This is why deduplication should consider several bibliographic fields together, including title, authors, publication venue, year, DOI, and source type.

When uncertainty remains, retain the records until their relationship can be established.

Can AI Miss Duplicates That Look Different?

Yes. A false-negative decision occurs when two records represent the same work but are not identified as duplicates.

For example, one database may export a complete article title and DOI, while another supplies an abbreviated title, different publication date, and no identifier.

The records may appear sufficiently different that a text-based comparison fails to connect them.

Missing duplicates can inflate the number of apparently unique records in a literature search and create unnecessary screening work.

In a manuscript bibliography, undetected duplicates may produce repeated entries or inconsistent in-text citations for the same publication.

Both false-positive and false-negative decisions matter, although their consequences differ according to the task.

Why Is Duplicate Detection Particularly Important in Systematic Reviews?

Systematic reviews commonly search multiple databases to improve coverage of the relevant literature. As a result, the same publication may appear repeatedly across exported search results.

Deduplication is usually performed before title-and-abstract screening, although the exact workflow depends on the review protocol and software.

The PRISMA 2020 statement includes reporting of duplicate records removed before screening. Its flow diagram distinguishes records identified, records removed before screening, and subsequent screening and inclusion decisions.

Deduplication therefore affects more than the appearance of a reference list. It contributes to the reproducibility and transparency of the study-selection process.

If an AI tool removes records without preserving the original identifiers, matching decisions, and counts, researchers may struggle to reconstruct how many records were identified and excluded.

What Does Research Say About Automated Reference Deduplication?

Automated deduplication has been studied extensively in evidence synthesis, although findings from conventional matching algorithms should not be treated as direct evidence of generative AI performance.

Bramer and colleagues (2016), writing in the Journal of the Medical Library Association, described a method for removing duplicate records from database search results using EndNote. Their work illustrates how bibliographic fields and sequential matching procedures can support deduplication.

Rathbone and colleagues (2015), in the Systematic Reviews journal, compared approaches to identifying and removing duplicate records and discussed the importance of accurate deduplication for systematic-review workflows.

These studies concern bibliographic deduplication methods rather than the performance of current generative language models.

Consequently, it would be inappropriate to claim that ChatGPT achieves a particular duplicate-detection accuracy based solely on those investigations. A generative AI system would need to be evaluated directly against a validated reference dataset to establish its precision and recall.

How Do Precision and Recall Apply to Duplicate Detection?

Duplicate detection can be evaluated using two familiar information-retrieval measures.

Precision describes the proportion of records or record pairs flagged as duplicates that are genuinely duplicates.

Recall describes the proportion of genuine duplicates that the system successfully identifies.

In this example, 90% precision means that most flagged pairs are correct. A recall of 75% means that one-quarter of the genuine duplicate pairs were missed.

These percentages are illustrative and do not represent the measured performance of any particular AI tool.

Neither metric alone establishes that a deduplication system is suitable for every research task. In systematic reviews, the consequences of incorrectly removing distinct studies may justify a cautious approach to uncertain matches.

How Do Reference Managers Detect Duplicate Records?

Reference managers typically use structured metadata and matching rules to identify potential duplicates.

Zotero provides a Duplicate Items collection that helps users identify records likely to represent the same item. Its documentation explains that duplicate detection considers bibliographic information such as titles, DOIs, and ISBNs, with additional matching considerations for some item types.

Users can review suggested duplicates and merge them, retaining a single library item while preserving relevant relationships and selected metadata.

This differs from asking a language model to rewrite an entire bibliography without preserving structured record identities.

However, reference-manager matching is not infallible. Similar publications may be flagged incorrectly, while records with substantial metadata differences may remain undetected.

Researchers should review proposed merges, particularly when the records represent preprints, conference papers, journal articles, or other related outputs.

Should You Use AI to Remove Duplicates or Only to Flag Them?

For most scholarly workflows, using AI to flag potentially duplicate references is safer than allowing it to delete or merge records automatically.

A useful AI-assisted output might identify the record numbers, explain which fields match, identify discrepancies, and classify the pair as a likely duplicate or a case requiring manual review.

For example, a model might report that references 14 and 37 have identical DOIs but different journal abbreviations. That provides a specific, testable reason to investigate the pair.

By contrast, an instruction such as "remove all duplicate references and return the corrected bibliography" gives the model greater opportunity to omit entries or alter metadata without explaining its decisions.

The researcher should retain the original records and make final merge decisions using verified publication information.

Can Duplicate References Be Confused With Different Versions of the Same Research?

Yes. This is one of the more consequential challenges in deduplication.

A preprint may later appear as a peer-reviewed journal article. A conference paper may be expanded into a journal publication. A research team may also publish separate analyses using the same dataset.

These records can share titles, authors, methods, or findings while representing distinct scholarly outputs.

Whether to consolidate them depends on the purpose of the review and the relationship between the publications.

For a systematic review, multiple reports of the same underlying study may need to be linked rather than simply discarded. PRISMA 2020 distinguishes records, reports, and studies because these are not interchangeable units.

This distinction prevents researchers from assuming that every repeated study description represents a redundant publication.

Can AI Detect Duplicate References Created by Its Own Earlier Outputs?

Potentially, but it should not be relied upon to identify every repeated or inconsistent citation.

An AI-generated bibliography may contain the same paper under different author initials, publication years, or title variations. It may also contain a genuine reference alongside an inaccurate version of that reference.

In such cases, deduplication should be combined with bibliographic verification. Two entries may appear to describe the same publication, but the researcher must establish which details are correct.

The risk of AI combining details from different publications further complicates matching because a composite citation may resemble several genuine records without accurately identifying any one of them.

Finding Duplicate References With Different Bibliographic Details

This is a relatively straightforward duplicate because the DOI and publication identity agree.

More difficult cases involve missing identifiers, conflicting publication dates, or closely related articles. In those situations, an AI-generated duplicate suggestion should trigger investigation rather than automatic deletion.

Common Misconceptions About AI Bibliography Deduplication

When Should You Use AI for Duplicate Reference Detection?

AI can be useful for reviewing a small bibliography, identifying suspicious similarities, and explaining why particular records may represent the same publication.

For large bibliographies and systematic reviews, a structured deduplication workflow is generally preferable because it can preserve record identities, apply consistent matching rules, and support an auditable decision process.

Where possible, perform duplicate detection on structured reference records rather than relying exclusively on formatted bibliography text.

This approach fits within a broader workflow for using AI with academic references that separates source identification, metadata verification, and bibliography management.

Before Removing Duplicate Academic References

Frequently Asked Questions About AI Duplicate Reference Detection

AI Can Flag Possible Duplicates, but Researchers Must Confirm the Matches

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.