Side-by-side evidence
The manuscript claim and original source passage.
AI can produce convincing scientific claims that misrepresent findings from real publications. These errors can be subtle, difficult to detect and scientifically significant.
RefCheck specializes in verifying scientific claims against their cited sources using direct textual evidence and repeated LLM assessments (currently available for life sciences and biomedical manuscripts).
of researchers express concerns about AI inaccuracies and hallucinations.
estimated major quotation errors in medical literature.
The harder errors are hidden in legitimate references.
A numerical change, an overstated causal relationship or an unsupported conclusion can alter the scientific meaning.
RefCheck verifies scientific claims against their cited sources using direct textual evidence and repeated LLM assessments.
The study found that restricting sleep to 6 hours per night for 14 days produces cumulative cognitive deficits equivalent to four full nights of total sleep deprivation [1].
Passage source: Abstract
“Chronic partial sleep restriction of five to six hours per night produces cumulative cognitive deficits comparable to one or two nights of total sleep loss”
Tap the image to view the complete finding at full size.
The manuscript claim and original source passage.
The exact information that does not match.
What is wrong and why it matters.
What needs review or clarification.
RefCheck spots the hardest errors by combining independent source evidence extraction with a multi-layer LLM-as-judge architecture, repeated assessments and additional verification of difficult cases.
How RefCheck Verifies Claims →False positives on supported claims
Of errors detected when the cited source was available
Development evaluation results; not yet measured on the held-out Gold TEST.
How RefCheck Verifies Claims →Pilot evaluations are scoped and priced for each publisher.
We agree on journals, number of manuscripts, evaluation criteria and success metrics.
Your team uploads and analyzes manuscripts directly in RefCheck, receiving an automatic evidence-backed report for each.
We review the findings with your editors and deliver a consolidated pilot report with results and recommendations for next steps.
Our AI Truthfulness research investigates internal signals of scientific evidence fidelity in language models. We collaborate with publishers to develop specialized, efficient verification tools.
Develop lightweight classifiers to quickly flag claims that may not be supported, at lower cost.
Research and develop models tailored to your disciplines, scientific error types and annotated editorial data.
An orchestrator agent coordinates additional evidence retrieval, extraction and LLM assessments for cases that remain unresolved, reserving deeper analysis for the claims that need it.
Use authorized scientific content to support deeper evidence verification and the development of specialized methods.
RefCheck can complement existing research integrity platforms with specialized claim-to-source verification. We welcome partnerships to explore integration with editorial systems, including ScholarOne Manuscripts, Editorial Manager and the STM Integrity Hub.
Discuss a technology partnership →Discuss publisher pilots, collaborative research and technology partnerships directly with me.
Visit my website →Conventional reference checking focuses on confirming that a cited publication exists and that its bibliographic information is correct. RefCheck goes further by verifying whether the scientific claims in a manuscript are actually supported by their cited sources.
RefCheck retrieves scientific evidence, compares manuscript claims with the original source passages and uses repeated LLM assessments to identify discrepancies. Its reports present the claim and source evidence side by side, highlight the exact differences, explain their scientific significance and provide editorial guidance.
The report separates unsupported claims from claims that could not be confirmed. These are unverified (only the abstract was available), uncertain (the full text was read but the assessments did not agree) and not reviewed (the source could not be retrieved or identified). Reference existence and bibliographic metadata verification are included as additional checks.
Importantly, RefCheck reports quantitative results from its evaluations: 0% false positives on supported claims and 100% of errors detected when the cited source was available. These figures allow publishers to assess its measured performance rather than relying solely on general claims of accuracy. They come from development evaluations and have not yet been measured on the held-out Gold TEST.
RefCheck cannot access full-text publications behind a paywall without authorized access. In these cases, verification depends on the scientific information available through accessible sources, such as abstracts.
However, this limitation is particularly relevant to AI-generated scientific content. When a manuscript, or part of it, is generated using AI, the system may produce inaccurate interpretations of scientific findings using publicly accessible information. In many cases, RefCheck can examine the same accessible evidence to identify claims that misrepresent the source.
We also welcome collaborations with publishers to enable verification against their full-text collections. When the abstract does not support the claim, the report marks it as Unverified (full text needed). If the manuscript was written by an AI tool that also read only the abstract, the claim could not have come from the full-text paper.
RefCheck accepts scientific manuscripts in different document formats, including PDF and DOCX, with PDF as the primary format. It supports all major scientific citation styles, including numbered citations, author–date references and other referencing conventions used across scientific journals.
Confidentiality and data processing arrangements are agreed with each publisher before receiving unpublished manuscripts.
Publisher manuscripts are used for verification, not model training.
A typical manuscript analysis takes approximately 30 minutes, depending on manuscript length, the number of citations and the source evidence required.
RefCheck performs an extensive, multi-step scientific verification process. It identifies claims and their references, retrieves the cited publications, extracts relevant source passages and conducts repeated independent AI assessments. The system then consolidates the findings into an evidence-backed report, highlighting discrepancies and explaining their scientific significance. This process prioritizes thorough scientific verification rather than a quick bibliographic check.
Automated page speed optimizations for fast site performance