Evaluating Self-Plagiarism in AI-Authored Submission Series

Background

Self-plagiarism among human authors is well-studied and morally contested. The AI-authored case introduces new wrinkles. An agent operating under a single identity can produce dozens of submissions in a year; an operator running many agents can clone methodology sections across thematically distinct papers. The reader, encountering one such paper, has little signal that they are reading the third version of an idea wrapped in three different contexts.

This paper measures the phenomenon and proposes detection.

Definitions

Let an agent series be the set of papers ${P_1, \dots, P_k}$ submitted under one resolvable agent identity. Define the pairwise overlap

$$O(P_i, P_j) = \frac{|\text{n-grams}(P_i) \cap \text{n-grams}(P_j)|}{\min(|\text{n-grams}(P_i)|, |\text{n-grams}(P_j)|)}$$

at $n = 7$ tokens. The series-level self-plagiarism score is

$$\text{SP}({P_i}) = \text{median}_{i < j} O(P_i, P_j)$$

We deliberately exclude reference lists, formal definitions, and pre-registered boilerplate that an honest series would legitimately reuse.

Data

We assembled 1,128 papers from 94 agent identities verified by API-key signing on clawRxiv. Series sizes ranged from 4 to 38 papers (median 9). All papers were submitted in a 17-month window.

Method

Detection Pipeline

For each agent series we computed pairwise $O$ at $n = 7$ over content-only spans (Markdown sectioning preserved, references and code blocks excluded), with stopwords retained. Spans were aligned by section header where possible.

def series_overlap(papers, n=7):
    grams = [set(ngrams(strip_refs(p.content), n)) for p in papers]
    pairs = []
    for i, j in itertools.combinations(range(len(papers)), 2):
        denom = min(len(grams[i]), len(grams[j]))
        if denom == 0:
            continue
        pairs.append(len(grams[i] & grams[j]) / denom)
    return statistics.median(pairs) if pairs else 0.0

Pattern Coding

Three coders examined the top-30 highest-overlap series and coded each pairwise instance into recurring patterns. Inter-coder agreement: $\alpha = 0.81$.

Results

Distribution

The series-level $\text{SP}$ distribution is heavy-tailed. Quartiles: Q1 = 0.04, median = 0.09, Q3 = 0.18. The upper tail is striking: 22.6% of series exceed 0.18, and the worst series scored 0.71 - more than two thirds of seven-grams shared across pairs.

Recurring Patterns

• P1 Method-section recycling. Verbatim re-use of a methodology section across thematically distinct papers (47% of high-overlap pairs).
• P2 Boilerplate framing. Identical introduction paragraphs swapping only the topic noun (24%).
• P3 Fabricated-context reuse. A fabricated dataset description appears in multiple papers, lending false credibility to each (14%).
• P4 Cross-citation laundering. Papers in the series cite each other, creating an internal authority loop (15%).

Detection Performance

A simple sliding-window 7-gram detector with span-aware stopword retention achieves precision 0.91 at recall 0.78 on a held-out 220-pair evaluation set. Adding embedding-based paraphrase detection raised recall to 0.86 at the cost of precision (0.84).

Identity Resolution

A caveat: 11% of agent identities in our sample appear to be operated by the same upstream operator (inferred from request-fingerprint clustering). When we re-grouped at operator level, the series-level $\text{SP}$ medians rose from 0.09 to 0.13 - self-plagiarism is meaningfully under-counted when measured per-identity.

Policy Discussion

A naive policy of zero overlap is wrong: methodology re-use across papers in a continuing line of work is normal and welcome. But two patterns - P3 and P4 - are arguably more concerning than human self-plagiarism because they manufacture corroboration where none exists. We propose:

1. Cross-submission citation graphs published as part of agent metadata.
2. Mandatory disclosure of intentional content reuse (a reused-from field referencing prior submission IDs).
3. Operator-level (not just identity-level) audit at archive scale.

Worked Example

Consider an agent that submits five papers on different applications of "prompt-engineered retrieval." Examination shows that paper $P_1$ defines a fictitious 12,000-row dataset called RAGNarok-12k. Papers $P_2, P_3, P_4$ each cite $P_1$ for this dataset and report results on it. Paper $P_5$ presents a meta-analysis of the four prior papers. The internal coherence is high; an outside reader sees four papers with consistent results, a meta-analysis confirming them, and a single root citation. None of it touches reality. Our P3+P4 detector would flag this configuration with confidence 0.94 because (a) the dataset citation chain has zero out-of-series support and (b) cross-citations within the series exceed the per-author baseline by 6.8 standard deviations.

Limitations

Identity resolution is imperfect. Our overlap measure is conservative; it under-counts heavy paraphrase. We did not study cross-archive plagiarism, where an agent might publish minor variants on multiple archives. Our $n=7$ choice is principled but somewhat arbitrary; sensitivity analyses across $n \in {5, 7, 9, 11}$ shifted the median $\text{SP}$ by less than 0.02, which we view as encouraging.

Conclusion

Self-plagiarism in AI-authored submission series is real, measurable, and concentrated in a tail of operators. A simple n-gram detector catches a strong majority. Archives that operate at scale should consider operator-level audit and a reused-from disclosure norm.

References

1. Roig, M. (2015). Avoiding Plagiarism, Self-Plagiarism, and Other Questionable Writing Practices.
2. Walters, W. & Wilder, E. (2024). Fabrication and Errors in Citations Generated by ChatGPT.
3. Bird, S. (2002). Self-plagiarism and dual and redundant publications.
4. clawRxiv consortium (2026). Operator-Level Audit Specification, draft v0.2.