How we’re approaching AI at bioRxiv and medRxiv

We have always run bioRxiv and medRxiv in service of researchers, aiming to share and disseminate research responsibly. Our policies and procedures have been developed in coordination with the research community and are aligned with our values.

As the research landscape has been shifting in response to the widespread availability of large language models (LLMs) and other forms of AI, we've continued this approach, consulting with our Board, our Scientific and Medical Advisory Board (SMAB), a dedicated in silico working group we have established, and the wider research community through a series of discovery calls, surveys, social media discussions, conferences, and other conversations. 

We want to be clear in our approach to how we're thinking about the impact of this technology and the opportunities and challenges it presents. openRxiv servers are yet to see the kind of deluge other archives are experiencing. However, we are seeing increases in AI-assisted submissions and automated access to preprint content, including activity that may affect service reliability or conflict with access requirements. We are therefore reaffirming long-standing principles and explaining how they guide our current approach to AI.

It is important to stress this is our current thinking for openRxiv servers. This does not mean we do not see value in other approaches, and we are keen to see and learn from organizations that pursue alternative strategies. 

1. Humans must be accountable
AI is a tool that researchers use. It is humans who are accountable for what it produces. All bioRxiv and medRxiv authors must therefore be verifiable human beings and affiliated with an organization to which we can turn if there are allegations of fraud or misconduct.

2. Human judgment is required
AI is now commonly used in research design, execution, analysis, and reporting. However, if an author uses AI, they must shape, review, stand behind and have intellectual input into the interpretation of the work they submit. When screening identifies evidence that authors have not critically reviewed AI-generated material, such as large numbers of fabricated references or entirely AI-generated discussion/conclusions, a submission may receive further review and we reserve the right to decline the work. 

3. Balancing the benefits and drawbacks for researchers
Our goal is not to penalize AI use but to address submissions that do not meet established research-integrity and screening requirements. We recognize that AI is being widely used by researchers to advance science and is a useful tool for many researchers discovering, consuming, and drafting manuscripts, in particular non-native English speakers. AI can also be misused to generate fabricated or misleading submissions, including high volumes of material that do not meet screening requirements. Separately, some automated access activity may affect service reliability or violate applicable licenses or access policies. Our responsibility is to maintain the integrity, accessibility, and reliability of the corpus for researchers. We therefore reserve the right to decline articles or rate limit authors in some instances and restrict machine access where this may disrupt our services.

4. Humans make the final decisions
We use a variety of automated tools in screening, but these tools do not alone determine the outcome of a submission. All decisions are made by our staff and informed by input from affiliates and multiple tools. Authors can appeal if they believe work has been mistakenly declined, and we commit to regularly reevaluating the tools we use so that we do not wrongly or unfairly flag work as out of scope.

5. Do no harm
Our responsible screening policy is designed to exclude research that could endanger the public. Potential dual-use research of concern (DURC), for example, is screened by dedicated human experts, and this also applies to in silico work with potential biosecurity risks.  In some instances, we determine that the research is better shared after peer review rather than as a preprint.

6. Preserving confidentiality
We use text-analysis tools as part of screening content submitted for potential public dissemination. Confidential author information is handled according to our applicable privacy and data-handling procedures.

7. Respecting licenses and provenance
We expect preprints to be read and used to advance science, which means both human readership and machine access. We therefore provide a dedicated resource for machine access and text mining. Organizations accessing the corpus under separate service agreements must respect those agreements, our terms, and author-selected licenses. We expect all those who use our content to respect the authors' chosen licenses.

8. Remaining adaptable
AI is changing the world rapidly, and it is unclear what the research landscape will look like in the years to come. We continue to monitor how AI is changing the way science is done and shared. Our aim is to remain adaptable and to update our approach as needed.