Citation Receipts: Auditing Retrieval-Backed Citation Events in Digital Libraries
This paper introduces the Citation Receipt Protocol (CRP), a system that generates cryptographically signed, auditable records of authenticated source retrievals to enable digital libraries to verify specific citation events through sampled checks, while explicitly acknowledging its limitations in certifying global completeness or semantic citation validity.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the modern world of research, a scholar's work is often measured by how often other researchers cite it. For decades, this counting was straightforward: a human writer read a paper, decided it was relevant, and added a reference in their own work. That act created a permanent, verifiable record. Today, however, a new kind of writer has entered the room: artificial intelligence. These systems can read thousands of papers and write summaries, literature reviews, and research drafts in seconds. They often include citations, pointing readers to the original sources. But a quiet problem has emerged. When an AI cites a paper, it might have actually read that paper to find the answer, or it might have simply guessed the title based on patterns it learned years ago. Worse, even when the AI does read the paper, the publisher of that paper often has no way of knowing it happened. The AI company sees the interaction, and the user sees the answer, but the library that holds the paper sees nothing. This creates a gap in the record of scientific discovery, where a paper could influence thousands of AI answers without ever receiving a single official credit.
A researcher named Kumar Abhinav has proposed a way to close this gap, not by changing how AI thinks, but by creating a new kind of digital receipt. Imagine a library where every time a book is checked out, the system automatically prints a small, tamper-proof ticket that says exactly which book was taken, when, and by whom. Abhinav's work, called the Citation Receipt Protocol, does something similar for the digital age. It is designed to prove that an AI system actually retrieved a specific version of a scientific paper before it used that paper to write an answer. The system works by having the AI provider sign a digital receipt the moment it fetches a source and displays it in an answer. This receipt is then locked into a public, unchangeable log, much like a notary public stamping a document and adding it to a permanent registry. This allows anyone, including the paper's publisher or an independent auditor, to check the log and verify that the citation was real and backed by an actual retrieval of the source text.
To test if this idea works in the real world, Abhinav ran a series of controlled experiments using a large language model that was trained on data up to January 2026. The researchers created a set of 100 scientific papers that were published after that training date, meaning the AI should not have known about them from its memory alone. They then asked the AI questions about these new papers under three different conditions. In the first condition, the AI had no access to the new papers at all. In the second, the AI was allowed to search a local library of these new papers and retrieve the top five results. In the third, the researchers forced the AI to look at a specific target paper. The results were clear. When the AI had no access to the new papers, it never cited them. When it was allowed to search and retrieve the papers, it cited the target paper in 95 out of 100 attempts. Crucially, for every single one of those 95 visible and declared citations that matched an authenticated source, the system successfully generated and verified a digital receipt. This proved that the protocol could distinguish between a citation based on a fresh retrieval and a citation that was merely a guess or a hallucination.
The study also looked at whether this system could detect an AI provider attempting to omit citations to make their system look better or to avoid reporting usage. They set up a test where the provider deliberately omitted receipts for 10% of the citations. The audit system, which acts like a random spot-check, was able to detect these missing receipts with high reliability. However, the study also found that the system is not a magic bullet that solves every problem. It cannot prove that the AI actually understood the meaning of the paper or that the citation was the only reason the AI gave a specific answer. It also cannot tell if the AI was trained on the paper in the past; it only proves that the paper was fetched and cited at that specific moment. Furthermore, the system relies on the honesty of the client recording the answer; if the user's device is compromised, the audit could be fooled.
The researchers measured the cost of adding this system and found it to be very small. Each digital receipt is a tiny file, roughly the size of a short text message, and creating or checking one takes less than a second on a standard computer. This suggests that the system could be added to existing digital libraries without slowing them down. The work also highlighted a critical distinction in how we count citations. A simple count of how many times a paper appears in an AI answer is not enough. The researchers showed that without the receipt protocol, a provider could easily hide citations or invent fake ones. With the protocol, the count becomes a verified record of a specific event: a specific version of a paper was fetched, and a specific answer was generated.
Ultimately, this research offers a practical tool for transparency in the age of artificial intelligence. It does not claim to solve the deep philosophical questions of how AI learns or whether it truly "understands" science. Instead, it provides a mechanical way to ensure that when an AI points to a source, that pointer is backed by a verifiable record of the source being accessed. For digital libraries and publishers, this means they can finally see which of their papers are being used by AI systems in real-time, rather than relying on the AI companies to tell them the truth. For the scientific community, it offers a way to track the flow of information with a new level of precision, ensuring that the credit for discovery is given where it is due, even when the writer is a machine. The study concludes that while this protocol does not certify the quality of the AI's thinking, it successfully creates an auditable trail for the citations it produces, turning a hidden interaction into a public, verifiable fact. It is worth noting that the author, Kumar Abhinav, has disclosed a potential financial competing interest: he personally owns a filed patent application related to the Citation Receipt Protocol.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.