AI tools and increasingly autonomous AI agents can hallucinate references, distort scientific evidence and generate claims that are not supported by their cited sources.
RefCheck detects these errors before they are published in scientific journals. It verifies citation integrity by checking that scientific claims are faithful to their sources and that references and bibliographic metadata are correct.
AI adoption is growing rapidly across research. The challenge for publishers is preventing false or distorted AI generated information from being published in scientific journals.
of researchers surveyed reported using AI tools in their work in 2025, compared with 57% in 2024. Wiley Newsroom
Reviewers assess scientific and methodological quality. They are not normally expected to open every cited paper and compare every statement with the primary source.
This creates a growing research integrity gap: scientific manuscripts can look credible and cite real literature while still containing information that the cited sources do not support.
The cited paper is real, but the manuscript distorts what the source actually reports.
The cited work is hallucinated, incorrectly identified or contains incorrect metadata.
Does the cited source support what the manuscript claims?
RefCheck examines the manuscript claim by claim, retrieves the cited source and compares the statement with the source evidence.
Does the cited work exist and do the reference details match the official record?
Authors, year, journal, volume, pages, DOI and other metadata are checked against authoritative scholarly databases.
“High consumption of ultra processed foods is associated with a 28% increase in the incidence of type 2 diabetes.”
“Participants in the highest consumption group showed a 36% higher risk of hypertension.”
The editorial team uploads the manuscript received from the researcher.
RefCheck runs citation fidelity and bibliographic integrity checks across the manuscript.
Editors receive the problematic claims, source evidence and bibliographic issues.
The editor can continue the process, request corrections or take another editorial action.
RefCheck can operate as a standalone platform or integrate into existing editorial systems through an API.
Screen manuscripts during initial checks, peer review or before the final editorial decision.
Check manuscripts before submission and provide researchers with evidence linked findings.
Embed citation verification into existing research and editorial platforms through an API.
We combine AI truthfulness with advanced methods from mechanistic interpretability and explainability to make AI systems safer and more trustworthy.
These methods help us understand how AI models work internally, how they “think”, and where and why they fail.
We are testing sparse autoencoders (SAEs) to identify internal features that may help distinguish faithful from unfaithful scientific claims.
Feature attribution, ablation and stability analysis help identify which internal signals contribute to reliable decisions and which may simply be noise.
Applied to research integrity, these methods offer a novel approach to a growing problem: increasingly autonomous AI systems that can generate false, misleading or incorrect scientific information. RefCheck uses these tools to understand and reduce those errors while strengthening the fight against paper mills and unreliable synthetic scientific content.
Publishers are facing growing volumes of synthetic scientific content, paper mills, hallucinated references and claims that distort real evidence. Detecting these problems before publication can reduce corrections, retractions, editorial workload and further contamination of the scientific record.
Identify unsupported claims and incorrect references before they are published and repeated by researchers or future AI systems.
Detect problems before they result in corrections, retractions, reputational damage and additional editorial work.
Add an evidence verification layer to publisher efforts against manipulated, fabricated and industrially produced scientific manuscripts.
Detect unreliable scientific content produced by increasingly autonomous AI systems before it enters the publishing workflow.
Before an AI generated error becomes part of the scientific record, verify the claim, verify the source and show the evidence.
Contact us