AI Tool Claims to Identify the Top 1% of Research Preprints: Should Researchers Trust It?

Artificial-intelligence tools for reviewing scientific manuscripts before publication are becoming increasingly common. One system, developed by Tel Aviv-based start-up QED Science, is designed to assess whether life-sciences research is original, scientifically valid and supported by the evidence presented.

The QED Science platform analyzes the claims in a manuscript, checks whether the data support those claims and identifies potential gaps in the research. The tool is free for researchers and has reportedly been used by more than 10,000 laboratories at 1,500 institutions in more than 70 countries.

In November last year, openRxiv — the non-profit organization that operates the bioRxiv and medRxiv preprint servers — announced a pilot programme to test the QED Science system on bioRxiv.

In an analysis published in June, QED Science used its AI system to rank more than 57,000 preprints posted on bioRxiv between May 2025 and April 2026. The analysis identified the top 1% of submissions according to the tool’s assessment.

The ranking has prompted debate among researchers. Some scientists argue that it could create another marker of prestige and further encourage the use of metrics in academic research. Others question whether artificial intelligence can evaluate scientific quality consistently, transparently and reliably.

Nature spoke with Niv Mastboim, co-founder and chief executive of QED Science, about the company’s AI-powered approach to scientific manuscript assessment.

How does QED Science’s AI platform evaluate scientific claims?

Many tools are available for reviewing research papers, but our approach differs in several ways. We trained the AI platform using multiple data sources, including open peer reviews, user feedback and synthetic data.

A key part of the process is helping the system understand what negative or non-supportive experimental results would look like. These are findings that do not support the original hypothesis or the conclusion being tested.

Scientific literature tends to over-represent successful and positive results. To assess research claims accurately, an AI system must also recognize evidence that fails to support a hypothesis. This can be learned from published null results, contradictory findings, failed replication studies and other examples of non-supportive evidence.

We develop internal metrics that allow the system to improve continuously. The platform evaluates each claim in a paper and assigns scores across several categories, including originality and validity.

The AI platform operates autonomously, but researchers also provide feedback. We use that feedback to refine and improve the system over time.

What was the purpose of QED Science’s top-1% preprint list?

The aim was to evaluate scientific work based on its content rather than the reputation of the journal, institution or authors. The 574 preprints included in the top 1% were the highest-scoring submissions among 57,455 bioRxiv preprints assessed by QED Science. Selection was based solely on the platform’s evaluation of originality and validity, independent of author identity, institutional affiliation or publication venue.

We also wanted to identify high-quality research that may have been overlooked by the existing publishing system. In a separate validation analysis, we examined 2,879 bioRxiv preprints posted in April 2025 that were later published in peer-reviewed journals. We then compared the rankings produced by our tool with the rankings of the journals in which the papers were eventually published.

QED Science rated 12.9% of those papers more highly than their eventual journals’ rankings might indicate. We refer to these studies as “hidden gems”. An expert panel assessed the strongest disagreements in a blinded evaluation and preferred the QED Science-favoured paper in 75% of decisive comparisons.

Could AI become the ultimate peer reviewer?

We are not currently offering our products to journals or publishers. Our goal is to provide authors with free, private and secure AI tools that can help them strengthen their research before publication.