The digital record of science is surprisingly fragile. While a printed journal from the 18th century remains readable today, a research paper published just five years ago often points to a digital graveyard of "404 Not Found" errors. This phenomenon, known as link rot, occurs when the URLs used to host code, datasets, and model weights become inaccessible due to server decommissioning, personnel changes, or the cessation of funding.
To ensure long-term reproducibility, the research community is shifting away from decentralized, ephemeral hosting toward centralized, archival-grade infrastructure.
The Anatomy of Digital Decay
The primary cause of link rot is the reliance on "transient" hosting solutions. Researchers often host artifacts on personal faculty pages, departmental servers, or temporary cloud storage links (like Dropbox or Google Drive). As Joseph Halpern and Joseph Y. Halpern noted in their founding philosophy for the Computing Research Repository (CoRR), decentralized servers are highly susceptible to "link rot" when hosting environments change.
The scale of this decay is significant across disciplines:
- Artificial Intelligence: An automated analysis by Jialiang Lin and colleagues found that 8% of source code repositories linked in AI papers become inaccessible over time.
- Computational Biology: Sebastian J. Schultheiß discovered that while many bioinformatics web services remain reachable, only 45% are fully functional just a few years after publication.
- Machine Learning Datasets: A systematic review of NeurIPS papers by Yiwei Wu and Hanlin Li highlighted that reliance on personal lab websites leads to frequent accessibility failures.
The Openness Gap in Modern AI
Reproducibility requires more than just code; it requires the specific artifacts used in the experiment. Rui Li identifies an "openness gap" in biomedical AI: while many researchers share code, only about one-third provide the trained model parameters (weights). When these weights are hosted on non-archival platforms, they are often lost, forcing future researchers to incur massive computational costs to re-train models from scratch—if they can reproduce the environment at all.
Why Persistent Identifiers Work
The solution to link rot is the decoupling of an artifact's identity from its location.
In a traditional URL, the identity and location are fused: university.edu/~professor/data.zip. If the professor moves or the university renames its servers, the link breaks. Persistent infrastructure, such as arXiv or repositories that issue Digital Object Identifiers (DOIs), uses a redirection layer. The DOI acts as a permanent name that the repository resolves to the current physical location of the file.
By moving artifacts to centralized hubs, the maintenance burden shifts from the individual researcher to an institution dedicated to long-term preservation. As Hancheng Cao and Lucy Lu Wang demonstrate, this shift toward centralized hosting platforms significantly mitigates the loss of research artifacts.
The Incentive for Persistence
Beyond the altruistic goal of scientific integrity, there is a clear professional incentive for using persistent repositories. Hancheng Cao found a massive "citation bonus"—papers with active, persistent links to their methods receive up to 60% more citations than those without.
When research is easy to find and verify, it is more likely to be built upon. Conversely, as Fabian Ruperti warns, even a single digit error in a persistent identifier or a broken link in a database can effectively "orphan" a dataset, making it invisible to the global research community. The transition to robust, centralized digital libraries like DIRECT or RRD-Bio represents the field's commitment to turning transient "results" into a permanent, verifiable scientific record.