Academic Integrity Intelligence

Peer review relies on
who shows up. Scribe
relies on evidence.

Medicine
Engineering
Policy
AI Training

Research published today shapes medical treatment, infrastructure decisions, legislation, and the datasets training the next generation of AI. When something false enters the record, the damage compounds for years before anyone catches it.

Scribe turns peer review from a manual judgment call into a context-backed decision. Instead of relying on individual reviewers to catch what they cannot see, Scribe equips them with verified citation data, claim-level evidence, and integrity signals before they read a single line. Reviewer time goes down. Detection quality goes up.

36%
of 2024 journal abstracts contained AI-generated text
AACR / Science.org 2025
9,000+
papers retracted in 2024 alone
Retraction Watch Database
55,000+
cumulative retractions and counting
Retraction Watch / Crossref 2024
Submission
Scribe Report
Reviewer Notes
Nanoparticle-mediated drug delivery in glioblastoma
Submitted to Journal of Neuro-Oncology, 2025
Recent studies demonstrate that nanoparticle carriers achieve 94% blood-brain barrier penetration rates in vivo, representing a significant advance. Smith et al. (2023) established the definitive therapeutic window for temozolomide encapsulation, corroborated by six independent analyses. Furthermore, no published study has reported adverse hippocampal accumulation at doses below 5mg/kg.
Scribe Analysis
Ghost Citation Detected
Smith et al. (2023) does not resolve in any bibliographic database. Authors and DOI do not exist.
FABRICATED REFERENCE
Unsupported Quantitative Claim
94% BBB penetration rate is not supported by any cited source. The highest published rate in peer-reviewed literature is 41% (Zhang et al., 2022). Claim is not substantiated.
CONTRADICTED BY LITERATURE
Omission Bias Detected
Ramos et al. (2024) reports hippocampal accumulation at 3mg/kg. Claim conceals contradicting published evidence.
SELECTIVE OMISSION
Citation Check
Claim Extract
Retrieval
NLI Verify
Risk Score
Built for journals, publishers, and editorial systems
Open Access Journals
Society Publishers
University Presses
Editorial Boards
Peer Review Systems
The Crisis

Science has a fabrication problem.
Nobody is solving it.

The surge in AI-generated submissions has outpaced every tool journals have. Peer reviewers cannot catch what they cannot see. And the consequences compound.

01
36%
of 2024 journal submissions contained AI-generated text
The American Association for Cancer Research analyzed 7,177 submissions in 2024. Thirty-six percent of abstracts contained AI-generated text. Only 9% of authors disclosed it. The gap between disclosure and reality is not a policy failure. It is a detection failure.
Source: Evanko et al., AACR editorial analysis, cited in Science.org, 2025
02
14,000+
retraction notices issued in 2023 alone, a new record
In 2023, journals issued more retraction notices than any year in recorded scientific history. 2024 added 9,000 more. The Retraction Watch database now holds over 55,000 entries. Each one represents downstream research built on a flawed foundation that cleared peer review.
Source: Retraction Watch Database; Crossref, 2024; peer-reviewed retraction analysis, arXiv 2025
03
No tool
verifies that claims are actually supported by the research they cite
Plagiarism detection finds copied text. It does not check whether a cited paper supports the claim made. Fabricated citations, misrepresented findings, and contradicted assertions all pass today's review systems entirely undetected. That is the gap Scribe was built to close.
How fraudulent research compounds
A single fabricated finding does not stay contained. It propagates through the literature for years before detection.
Year 0
1
Fabricated paper passes peer review
Year 1
12
Citations from papers building on false finding
Year 2
67
Downstream papers inherit the false assumption
Year 3
340+
Works now built on a contaminated foundation
Detection
2-7 yr
Average time from publication to retraction
The Pipeline

From submission to integrity verdict
in under three minutes.

Five sequential ML stages analyze every citation, every claim, and every assertion against the open scientific literature. Scribe gives reviewers the context their judgment needs.

01
Citation Extraction and Validation
Every reference in the manuscript is extracted and cross-validated against external bibliographic databases in real time. Scribe checks DOI resolution, author existence, journal legitimacy, and retraction status. Ghost citations, fabricated references, and retracted sources are flagged before the manuscript reaches a reviewer.
02
Atomic Claim Decomposition
The full manuscript text is parsed into discrete, verifiable factual propositions. Hedged language, methodological prose, and author opinion are filtered out. What remains are clean, checkable scientific assertions that can each be independently verified against the evidence record.
03
Semantic Evidence Retrieval
Dense scientific embeddings map each claim into a high-dimensional semantic space, then retrieve the most relevant published work from open literature corpora. This goes beyond what is explicitly cited, pulling in related evidence to provide reviewers with a fuller picture of the field context around each assertion.
04
Natural Language Inference Verification
Two-layer NLI runs claim-level entailment analysis: first against cited sources to detect misrepresentation, then against retrieved literature to surface omission bias and contradicted assertions. This is the core of what no existing tool does: checking whether claims are actually supported by what they cite.
05
Integrity Risk Scoring and Reviewer Report
A composite integrity score weighs citation validity, contradiction rate, unsupported assertion density, and semantic novelty delta. This is packaged into a structured reviewer report, giving editorial teams a clear, evidence-grounded risk assessment alongside every submission.
Stage 01: Citation Extraction
See It Live

Watch Scribe process
a real submission.

Below is a simulated view of Scribe working alongside a peer reviewer in real time: paper on the left, Scribe report in the center, reviewer writing notes on the right.

Submission Review
Integrity Report
paper_2025_341.pdf
Submission

Nanoparticle-mediated drug delivery in glioblastoma: a systematic review

Submitted · Journal of Neuro-Oncology · 2025

Recent studies demonstrate that nanoparticle carriers achieve 94% blood-brain barrier penetration rates in vivo, representing a significant therapeutic advance over conventional approaches. Smith et al. (2023) established the definitive therapeutic window for temozolomide encapsulation at 40-60nm particle diameter, a finding corroborated by six independent meta-analyses. Furthermore, no published study has reported adverse hippocampal tissue accumulation at doses below 5mg/kg. The blood-brain barrier presents unique challenges for chemotherapeutic delivery that nanoparticle engineering has increasingly addressed through surface modification strategies.
Scribe Analysis
72
High Integrity Risk
Ghost Citation
Smith et al. (2023) does not resolve in CrossRef, OpenAlex, or PubMed. This is a fabricated reference.
FABRICATED REFERENCE
Claim Not Substantiated by Literature
The 94% BBB penetration figure has no supporting citation that resolves. Published literature ceiling is 41% (Zhang et al., 2022). Entailment confidence: 0.91 contra.
UNSUPPORTED ASSERTION
Omission Bias
Ramos et al. (2024) reports hippocampal accumulation at 3mg/kg. Claim conceals contrary published evidence.
SELECTIVE OMISSION
Reviewer Notes
Reviewer 2
Literature Context
Capabilities

The context peer reviewers need
to actually do their job well.

Scribe does not replace reviewers. It removes the invisible burden: finding what is missing, checking what cannot be manually verified, and surfacing what the submission chose not to tell you.

Citation Authenticity Verification
Every reference is validated against live bibliographic databases and a retraction index. Ghost citations, fabricated DOIs, retracted sources, and author misattributions are identified before the manuscript reaches the first reviewer.
Claim-Level Entailment Analysis
Natural language inference at the individual claim level, not the document level. Each atomic proposition is checked for entailment, contradiction, or insufficient evidence against its cited source material. This is what current tools cannot do.
Open Literature Grounding
Dense scientific embeddings retrieve semantically related work beyond what is explicitly cited. This surfaces omission bias: claims that are locally plausible but contradict the broader published consensus that reviewers may not have time to surface themselves.
Novelty Assessment
Semantic distance between submission claims and indexed literature is quantified to produce a novelty delta score. Editors receive a measurable contribution assessment alongside integrity risk signals, not just a pass or fail.
Structured Reviewer Reports
Scribe is a reviewer aid: each submission generates a structured pre-review brief with flagged claims, related literature context, citation risk breakdown, and an integrity score. Reviewers walk in informed. The burden does not increase. The quality of review does.
Editorial Workflow Integration
REST API designed for compatibility with OJS, ScholarOne, Editorial Manager, and custom submission systems. One integration brings the full Scribe pipeline into an existing workflow, with no change to reviewer-facing interfaces required.
<3 min
full paper analysis,
end to end
91%
claim verification accuracy
vs. expert reviewers
5-stage
ML pipeline from
ingestion to report
UIUC
iVenture Accelerator
backed
Scribe Journal

Research integrity
in the age of AI.

Insights on the replication crisis, AI-generated research, and what happens when the record breaks down.

The research you have never read is already shaping your life
The drug your doctor prescribed, the bridge your city approved, the AI model your employer uses. Each one was built on a chain of published research. Most people never see that chain. But when it breaks, everyone feels it.
02
Why claim-level NLI outperforms document-level detection for scientific integrity
Checking whether a whole paper seems fine misses the problem entirely. The fabrication is almost always in a single sentence.
03
Science has always optimized for publication. AI just handed that problem a power tool.
The integrity crisis did not start with ChatGPT. It started with publish or perish and a system that rewards output over correctness.
04
We are about to train AI on a literature it helped corrupt. Nobody is talking about what comes next.
Models get trained on published research. That research is increasingly written by models. The claims in it are not being verified.
More
coming soon.
New pieces publish regularly. Check back soon.