AI evidence extraction

The Ghost Couple: Correlated LLM Name Priors and Their Haunting of the Web and Academic Publishing

Michał Brzozowski, Neo Christopher Chung · ArXiv.org · 2026

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

1/4

Quality (LMQS)

Evidence

Citations

0.00

FWCI

Methodology & findings

Study design

Multi-method forensic approach combining: (1) systematic API probing of nine Claude versions, ten GPT versions, and Gemini-2.5-flash with controlled prompts (30 single-name, 30 pair, 30 trio prompts at temperature 1.0); (2) web corpus collection via Serper.dev Google Search API across multiple archetypes; (3) academic infrastructure analysis via ResearchGate and Zenodo API queries; (4) manual verification of 436 ResearchGate publication records and 1,655 Zenodo records; (5) temporal analysis using citation_publication_date meta tags and DataCite server-side timestamps; (6) control corpus design using demographically matched non-ghost surnames..

Primary method

Predict-then-confirm methodology: controlled API probing establishes model-specific name priors, which are then used as search signatures to recover AI-generated content in the wild, converting controlled experiments into detection tools. Comparative cross-model analysis (Claude, GPT, Gemini).

Main result

The study demonstrates that "large language models do not merely default to high-probability individual names when generating fictional experts: they produce correlated character ensembles: pairs and trios whose co-occurrence rates far exceed chance and are consistent across independent generations." Additionally, "on Zenodo, a CERN-operated repository that mints real DataCite DOIs, we identify 1,655 ghost-authored records claiming nonexistent journals with fabricated publication dates: server-side DataCite timestamps prove deliberate backdating, and 991 records were registered in a single month." The authors found that Elena Vasquez appears at "67% in claude-sonnet-4-20250514, decaying monotonically" across model versions, and "the pair is fully suppressed in claude-sonnet-4-6."

Research paradigm

Empirical forensic analysis with qualitative and quantitative components

Author conclusions

The authors conclude: "We have shown that LLMs generate correlated character ensembles, not merely high-probability individual names, that are model-family-specific, version-specific, and actively suppressed at release boundaries; the suppression is itself evidence that the priors were strong enough to be noticed. These ghost names propagate from model outputs into AI-generated web content and from there into academic publishing infrastructure. On Zenodo alone, 1,655 ghost-authored records with real DataCite DOIs were registered in a 60-day automated burst, claiming nonexistent journals with backdated publication dates; the infrastructure for large-scale scholarly record contamination is already in place. The academic record is being quietly haunted."

Risk of bias

Selection bias: API probing limited to publicly accessible checkpoints only; Measurement bias: Google Search API subject to recency bias in indexing; Classification bias: Potential misidentification of real researchers whose names coincidentally match ghost names; Sampling bias: Prompt set size (n=30) may miss lower-frequency name priors; Temporal bias: Web page publication dates unreliable for slop sites; only ResearchGate dates trustworthy; API-only probing misses internal model behavior and fine-tuned variants; Prompt set size (30 per condition) may miss lower-frequency name priors; Google Search recency bias in web corpus collection; Manual verification introduces human annotation bias; Possible false positives: real researchers whose names coincidentally match ghost priors; Temporal analysis relies on user-controllable metadata (Zenodo publication_date field); Selection bias: ghost names that are more distinctive may be more discoverable in web searches; Recency bias in Google Search indexing affecting web corpus collection; Unreliability of page-level publication dates on AI-generated content sites; Incomplete coverage of internal/fine-tuned models not accessible via public APIs; Potential false positives from legitimate researchers sharing names with ghost priors; Selection bias in manual review of ResearchGate and Zenodo records; Temporal lag between model release and appearance of ghost content in wild

Limitations

The authors state that "Our probing study covers only publicly accessible API checkpoints
internal or fine-tuned models are not covered
Prompt set size (30 prompts per condition) is sufficient to establish dominant priors but may miss lower-frequency names
Web corpus collection via Google Search (Serper) is subject to recency bias in the age field
page-level publication dates from slop sites are unreliable
ResearchGate paper dates are more trustworthy but require systematic collection at scale, which is ongoing

Open questions raised

The authors identify the need for: (1) systematic collection of ResearchGate data at scale; (2) investigation of internal or fine-tuned models beyond publicly accessible APIs; (3) larger prompt set sizes to detect lower-frequency name priors; (4) real-time monitoring infrastructure as fake-paper generators continue deployment.
The authors identify: (1) Internal or fine-tuned models not covered by API probing; (2) Need for systematic ResearchGate collection at scale; (3) Downstream impact on scholarly aggregators (Semantic Scholar, Google Scholar) not yet fully indexed but infrastructure in place; (4) Understanding of why ensemble crystallization occurs (trio > pair > solo structure may reflect narrative fine-tuning data volume/structure); (5) Need to date LLM-generated content using publication date proxies across platforms.
The paper identifies the need for systematic collection of ResearchGate data at scale and proposes publication dates on fake academic records as a temporal proxy for model deployment windows. It notes that ghost-authored papers cite real work but real papers' 'cited by' lists do not yet include them, suggesting future research into citation network contamination.

Data: Zenodo API queries: 1,655 ghost-authored records with DOIs (10.5281/zenodo.*); ResearchGate corpus: 436 publication records across ghost names; Web corpus via Serper.dev: 515 URLs (Claude), 714 URLs (Gemini), 816 URLs (GPT); Web corpus via Serper.dev Google Search API (not directly downloadable; reproducible via API queries documented in Appendices B and C); Zenodo API queries (1,655 records with DOIs in range 10.5281/zenodo.*); ResearchGate data via Serper API (436 publication records); DataCite API for timestamp verification. No downloadable dataset provided.; Zenodo ghost-authored corpus (1,655 records with 10.5281/zenodo.* DOIs); ResearchGate publication records (436 records identified); Web corpus from Google Search API (1,515 URLs across three model families); Control corpus with demographically matched non-ghost surnamesExtracted from: pdfAgreement 66%

Explore related topics

Hallucination Control Research Integrity Content Analysis