Attribution Decay

The most common form of decay where a URL returns a 404 Not Found or similar client error. Studies show 50% of URLs in Supreme Court opinions and over 60% of links in academic papers from 1999 are now dead. This occurs when servers are decommissioned, content management systems migrate without redirects, or domains expire. For AI systems, link rot breaks the verifiability chain, turning grounded citations into unverifiable claims.

A subtler form of decay where the URL remains valid but the content at the destination changes significantly. The original claim-supporting passage may be edited, removed, or contradicted. This is prevalent in:

• Dynamic web pages (wikis, news sites)
• Unversioned documentation
• Blog posts with silent corrections

Content drift invalidates the semantic grounding of a citation even though the link technically resolves.

The umbrella term coined by researchers encompassing both link rot and content drift. A 2013 Harvard study found that approximately 70% of URLs in law journals suffered from some form of reference rot. For generative AI systems, reference rot means that even perfectly formatted citations may point to unverifiable or altered evidence, eroding the trustworthiness of the entire output over time.

Occurs when a citation references a source as it existed at time T1, but the verifier accesses it at time T2 after changes. Without a timestamped snapshot (e.g., from the Internet Archive's Wayback Machine) or a content fingerprint, there is no cryptographic proof that the cited content ever existed in the claimed form. This is critical for scientific reproducibility and legal precedent verification.

When a foundational source decays, all dependent citations that rely on it become suspect. In a citation graph, a single broken node can invalidate an entire downstream chain of reasoning. This is particularly dangerous for AI systems that perform multi-hop reasoning across sources, as the decay of one intermediate reference silently corrupts the final conclusion without obvious error signals.

The primary defense against attribution decay is the use of persistent identifiers (PIDs) like DOIs, ARKs, and Handles. These abstract the physical location from the identifier, using a resolver service to redirect to the current URL. Combined with content fingerprinting (cryptographic hashes) and archival snapshots, PIDs ensure long-term verifiability. AI citation systems should prioritize PID-tagged sources over raw URLs.

A compact digital signature generated by a cryptographic hash function (such as SHA-256) from a piece of content. This fingerprint uniquely identifies the content and verifies its integrity.

• Enables detection of silent edits that cause attribution decay
• Used in systems like the News Provenance Project to track content modifications
• A single character change produces a completely different hash value

An append-only, tamper-evident log that records a chronological chain of custody for a digital asset. Often implemented using distributed ledger technology.

• Each entry is cryptographically linked to the previous one
• Provides an immutable audit trail of all modifications
• Enables verification that cited content hasn't been altered post-publication
• Critical for combating attribution decay in regulatory and legal contexts

The process of linking a generated claim directly to a specific, verifiable segment within an authoritative source document. This is the generative AI equivalent of academic citation.

• Requires reference anchoring at the text-span level, not just document level
• Enables users to verify claims even if the source later changes
• Foundational to Retrieval-Augmented Generation (RAG) systems
• Without grounding, attribution decay is impossible to detect

The process of transforming different versions of the same content into a single, authoritative form to enable accurate deduplication and comparison.

• Prevents citation fragmentation across multiple copies
• Uses techniques like URL normalization and content hashing
• Essential for detecting when a cited source has been silently modified
• Reduces false positives in attribution decay monitoring systems

A probability estimate generated by a model indicating the likelihood that a specific source passage fully and accurately supports the claim it grounds.

• Combines semantic similarity with source authority signals
• Low confidence scores can flag potential attribution decay before links break
• Used in enterprise RAG systems to prioritize high-integrity citations
• Typically ranges from 0.0 to 1.0 with threshold-based filtering