What is knowledge decay in AI?

Knowledge decay is the growing gap between when a source was last verified and when your AI system keeps citing it. The retrieval is still 'correct' — it returns the document it was supposed to — but the document itself is out of date, so the answer is confidently wrong.

What is a knowledge half-life?

A knowledge half-life is the rate at which a given source loses accuracy over time. Different source types decay at very different rates: a pricing page may go stale in weeks, a product doc in months, a regulation in a year or more. Treating them all as equally fresh is the core mistake.

How is knowledge decay different from model hallucination?

A hallucination invents a source that never existed. Knowledge decay cites a real source that has since changed. Decay is harder to catch because the citation looks legitimate — the document is real, it was retrieved correctly, and only the freshness is wrong.

What Is Knowledge Decay? Why Your AI's Sources Go Stale (and What It Costs You)

A support assistant tells a customer your refund window is 30 days. It cites an internal policy doc — the right doc, retrieved correctly, with a tidy citation. The only problem: that policy changed to 14 days three months ago, and nobody re-indexed the source. The model didn’t hallucinate. It quoted a document that quietly went stale.

That gap — between when a source was last verified and when your AI keeps citing it — is knowledge decay. It’s the most under-instrumented failure mode in production AI, because every dashboard you already have says the system is healthy.

What knowledge decay actually is

Retrieval-augmented generation (RAG) was supposed to fix hallucination by grounding answers in your own documents. It does — right up until those documents change and the index doesn’t. After that, RAG faithfully grounds your AI in information that is no longer true.

Knowledge decay is not a model problem. It’s a source problem. The embeddings are fine. The vector search is fine. The retrieval scores are high. The source underneath simply moved, and your pipeline has no idea.

Knowledge has a half-life

The trap is treating every source as equally fresh. It isn’t. Different knowledge decays at wildly different rates:

Pricing and offers — days to weeks.
Product documentation — weeks to months, faster after a release.
Internal policies — unpredictable; they change the day legal says so.
Regulations and standards — slower, but the stakes when they shift are highest.

A source’s half-life is how quickly it loses accuracy. A knowledge base mixes sources with half-lives that differ by orders of magnitude, and a system that refreshes everything on the same schedule is either wasting effort on the stable docs or shipping stale answers from the volatile ones.

The three ways decayed knowledge fails

Silent staleness. The cited source is real and retrieved correctly — it’s just out of date. The answer looks fully sourced, which is exactly why no one questions it.
Cross-source contradiction. Two documents now disagree (an old policy and its replacement both sit in the index). Top-k retrieval picks one silently, and you never learn which. We dug into this failure mode separately.
Over-confident citation. The model attaches a clean citation to a decayed doc, lending false authority to a wrong answer — the worst case for a regulated or customer-facing deployment.

Why retrieval accuracy isn’t freshness

Most RAG evaluation measures whether the system retrieved the relevant document. None of it measures whether that document is still true. You can hit 95% retrieval precision and still ship yesterday’s answer to today’s customer, because precision and freshness are different axes — and almost nobody instruments the second one.

This is the same blind spot FinOps solved for cost. AI Vyuh FinOps watches what your AI costs; ProvenanceOps watches what your AI cites. Both are invisible until the invoice — or the incident — arrives.

Measuring decay: the Knowledge Decay Index

You can’t manage what you can’t see. A Knowledge Decay Index scores each source by how long it’s been since verification, weighted by its half-life and how heavily your AI relies on it. Continuous source-attribution monitoring then watches every connected source — Confluence, Notion, Glean, SharePoint, custom RAG pipelines, the open web — and flags the moment a document changes, so every dependent AI feature is marked the instant its ground truth shifts.

The point is to move from reactive (a customer catches the stale answer) to proactive (you refresh the source before the model ever cites the old version).

Where to start

You don’t need to re-architect your stack to find out how bad the decay is. Start by measuring one knowledge base: join the waitlist for a free Provenance Audit and get a Knowledge Decay Index score plus the top stale-document candidates for a single source. It’s the fastest way to see how much of your AI’s “grounded” output is quietly running on expired information.

ProvenanceOps is part of the AI Vyuh AI-operations stack — the layer that proves what your agents know is still true.