The Half-Life of Facts: Why Everything You Know Has an Expiration Date

In 1923, the cytologist Theophilus Painter peered through a microscope at human testicular cells and counted forty-eight chromosomes. He was wrong by two. But for the next thirty-three years, every biology textbook, every medical school lecture, every popular science article gave the same answer: forty-eight. The number survived the Second World War, more than a generation of geneticists, and the Watson–Crick paper on DNA. It was finally corrected in 1956, when Joe Hin Tjio and Albert Levan, using a better staining technique, counted forty-six and published their result in the journal Hereditas. Three decades of confident, peer-reviewed, settled fact disappeared in a single paper.

What kept the wrong number alive for so long wasn’t conspiracy or fraud. It was confirmation bias on an institutional scale. Each researcher who repeated Painter’s count expected to see forty-eight, so they saw forty-eight. An entire generation of geneticists learned the wrong number, taught it, and passed it on.

This is the central problem of human knowledge. Not that we don’t know things, but that what we know has an expiration date most of us never check.

Facts decay like radioactive isotopes

Samuel Arbesman wrote a book about this in 2012 called The Half-Life of Facts, and the analogy in the title is more than poetic license. You cannot predict which uranium-238 atom in a sample will decay next, but you can predict with extraordinary precision that half of them will have decayed within roughly 4.5 billion years. Facts behave the same way at the population level. You cannot predict which specific claim in a hepatology textbook will be overturned next month, but you can predict, within a decade, when half of its claims will be wrong.

That precision exists because someone bothered to measure it. In 2002, Thierry Poynard and colleagues published “Truth Survival in Clinical Research: An Evidence-Based Requiem?” in the Annals of Internal Medicine. They examined 474 articles on cirrhosis and hepatitis published between 1945 and 1999, then asked expert panels to rate each conclusion as still true, obsolete, or disproved. The results: 60 percent still true, 19 percent obsolete, 21 percent flatly false. The half-life of a hepatology paper’s central conclusion: forty-five years. Cite a hepatology meta-analysis from 2002 today, and there is roughly a coin-flip chance its conclusion still holds in 2047.

That number sits in the middle of a hierarchy that surprises nobody who has thought about it and surprises everybody else.

The rate at which a domain’s facts decay tracks how quickly the underlying subject changes and how much new research is being produced. Mathematics doesn’t change because what’s true today was true yesterday in a more rigorous sense than empirical fields can offer. The Pythagorean theorem doesn’t have an expiration date. But hepatology does. Psychology does. Software engineering does most aggressively of all — the 31.5 percent annual depreciation rate that NBER economists derive from productivity data implies that half of what a software team knows today will be obsolete within roughly two years. The skill of “configuring a Webpack bundler” had a half-life of about two release cycles.

The gold standard decays fastest

The Poynard study buried a result so counterintuitive it deserves its own paragraph. Conclusions from meta-analyses had lower twenty-year survival, 57 percent, than conclusions from individual non-randomized studies (87 percent) or randomized trials (85 percent). The supposed gold standard of evidence — the meta-analysis — decayed faster than the merely good evidence it was built on.

Why? Because a meta-analysis is a group photograph of a field’s knowledge at one moment. It captures everyone in frame when the shutter clicks; every researcher who later joins, leaves, or changes their mind makes the snapshot less faithful. An individual study is more like a portrait — it’s either a faithful rendering of one person’s findings or it isn’t, and you can revise it without restaging everyone else.

The same inversion shows up in artificial intelligence. MIT Sloan researchers reported that large language models use roughly 34 percent more confident language when they’re hallucinating than when they’re stating known facts. (The figure is widely cited but the methodology has not been fully disclosed; treat it as directional rather than precise.) The model is most certain precisely when it is most wrong. In both cases, the markers we instinctively use to judge reliability point away from accuracy rather than toward it. Heuristics that work in low-volume, slow-moving information environments invert when the volume and velocity rise.

The web has its own decay constant

The internet was supposed to fix all this. Information would be cheap to publish, easy to update, and perpetually current. What actually happened is that the medium turned out to have its own decay curve, and a faster one than the printed page.

In May 2024, the Pew Research Center published “When Online Content Disappears,” which measured link rot across millions of pages. As of October 2023:

• Twenty-five percent of all webpages that existed between 2013 and 2023 were no longer accessible.
• Thirty-eight percent of webpages from 2013 were gone by 2023.
• Eight percent of pages published in 2023 were already inaccessible within the same year.
• Fifty-three percent of English Wikipedia articles contained at least one broken reference link.
• Eleven percent of all Wikipedia reference links pointed to dead URLs.
• Eighteen percent of tweets disappeared within three months of posting.

These are not just dead links. Every broken reference link in Wikipedia is a fact that can no longer be verified, and the 53 percent figure represents an enormous epistemic vulnerability sitting under what is, by some measures, the most actively maintained knowledge base in human history. The maintenance burden scales with the stock of knowledge, not with the rate at which new knowledge is added.

That is the structural problem. The new problem stacked on top is retractions. According to Lendvai and Sasvári (2025) in the Journal of Information Science, scientific retractions grew from 140 in 2000 to over 11,000 in 2022 — a compound annual growth rate of roughly 22 percent, faster than any growth rate in the publication corpus itself. A 2024 paper documented that English Wikipedia articles citing retracted papers had received over 250 million cumulative page views, and that retracted health science papers were more than three times as likely to be cited on Wikipedia as non-retracted ones. Wrong facts spread faster through the citation graph than right ones.

The language model is downstream of all of it

Now layer the language model on top. An LLM trained on a 2024 snapshot of the web has, embedded in its weights, every fact in that snapshot — the ones that were already wrong, the ones about to become wrong, and citations to papers since retracted. It does not know which is which. It cannot, structurally. Its training corpus is a fossil, and the world the corpus was sampled from has moved on.

Industry research (directional) suggests data only six months old can produce a 19 percent rise in hallucination rates for AI market forecasts. Retrieval-augmented generation reduces outdated hallucinations by roughly 30 percent in production — real progress, not a panacea. The math underneath is unforgiving. A knowledge base that decays at, say, 22 percent per year (a typical rate for B2B contact data) feeding a model fine-tuned eighteen months ago means a substantial fraction of what the model “knows” is already false. The model uses confident language to assert it. The downstream system trusts the confidence. The half-life problem becomes a runtime problem.

This turns knowledge decay from academic curiosity into something tech leaders should care about. Every team deploying RAG over an internal documentation set is shipping a product on top of a corpus with a measurable decay rate. Stack Overflow’s documentation engineers reported in 2024 that internal docs become “suspect” at six months and “actively misleading — worse than no documentation” at twelve. Roughly half of documentation site traffic now comes from AI agents. Stale docs don’t just degrade humans anymore — they degrade the models humans rely on to read.

The math is short and unforgiving

Here is the math. For any knowledge system, let:

• P = production rate (new facts added per unit time)
• λ = decay rate (fraction of existing facts becoming wrong per unit time)
• S = stock of facts in the system

The maintenance burden is λ × S, and it grows with S. The system becomes maintenance-dominated at the moment λ × S > P.

Once you write it down, the inevitability is obvious. Production rates are bounded by editorial capacity, headcount, or LLM token budgets. The stock of facts is unbounded — it grows monotonically with every documentation page and every API spec update. The decay rate is set by the world, not by the system. Every knowledge base, given enough time, crosses the threshold where keeping the existing stock current consumes more effort than adding to it. Wikipedia is past it. Most enterprise knowledge bases are deeply past it. The chromosome textbooks of 1923 to 1956 were past it from the moment Painter published.

Derek de Solla Price, the father of scientometrics, demonstrated in 1963 that scientific literature doubles roughly every fifteen years. The Poynard hepatology half-life is forty-five years. Compound those: the stock of medical literature triples in the time it takes half of the existing stock to become wrong. The corrections cannot keep up with the additions, and the additions are accelerating.

Knowledge as undeclared depreciation

One analogy from the wider economy pulls this together. A factory depreciates its machinery on a schedule because everyone understands machines wear out — a truck bought today will be worth less next year, and the IRS lets the company expense the difference.

Knowledge does not appear on any balance sheet, but it should. A company’s knowledge base of ten thousand internal documentation pages, in a domain with a three-year fact half-life, is carrying an enormous unbooked write-down. Most companies treat the corpus as a pure asset and budget zero for maintenance. The depreciation happens anyway — quietly, distributed across every customer support ticket where the documented answer is wrong, every onboarding deck that misrepresents the system as it stood eighteen months ago, every analyst report that cites a benchmark from a model version that was deprecated last quarter.

R&D economists have measured this. The standard assumption for R&D capital depreciation is 15 percent per year. Industry-specific NBER estimates put software knowledge depreciation around 31.5 percent and pharmaceuticals around 41 percent. A pharmaceutical company is, in econometric terms, losing about 41 percent of its R&D knowledge value annually. The accountants don’t book the loss because there is no GAAP entry for it, but the productivity data picks it up. The knowledge depreciates whether anyone records it or not.

Where the analogy breaks

The radioactive-decay framing isn’t perfect, and it is worth saying where it bends.

First, knowledge does not decay symmetrically. Some “facts” are refined toward greater accuracy rather than overturned — Newtonian mechanics did not become wrong, it became a useful approximation embedded in a more general theory. A reader who treats every old fact as expired throws out a lot of partial truth alongside the actual errors. The half-life model captures the worst case, not the average case.

Second, the half-life numbers are imprecise. Poynard used expert panels. Neimeyer used Delphi polls of psychologists. Citation half-lives depend on citation behavior rather than on truth content. The right reading of “forty-five years for hepatology” is “decades, not years or centuries,” not a precision constant for capacity planning.

Third, the AI tools that amplify the decay loop also offer the most plausible path to managing it. RAG, automated staleness detection, and projects like the University of Illinois’s WINELL system (Wikipedia Never-Ending Updating with LLM Agents, 2024) — which monitors online sources for facts that update existing Wikipedia articles — don’t solve the problem, but they redirect human attention to the cases where it matters most.

The point of the half-life framing is not that all knowledge is equally untrustworthy. It is that the rate at which a given claim degrades is measurable, varies by domain, and is something you can plan around if you bother to estimate it.

What to actually do about it

For anyone building software systems in 2026, the practical takeaway is that knowledge is not a static asset, and pretending it is creates compounding liabilities. Four moves are concrete enough to recommend.

Triage by domain half-life. Tag every document, RAG corpus, and wiki entry with its expected shelf life. Half-life under one year (security advisories, model benchmarks, prices, contact data): set up automated monitoring. Half-life one to ten years (architecture decisions, regulatory frameworks, API contracts): schedule a re-verification cycle calibrated to the domain. Half-life ten years or more (mathematical proofs, foundational theory, physical constants): invest in initial accuracy and stop worrying about it.

Write decay-aware metadata. Every factual claim should carry a date of writing, a source-reliability tier, a domain half-life estimate, and a last-verified date. The metadata is cheap. The lack of it is what makes knowledge decay invisible until something breaks loudly.

Budget maintenance like infrastructure. A team that ships a hundred documentation pages per quarter and budgets zero hours for maintenance is, mathematically, building a knowledge liability with a known doubling time. Allocate explicit maintenance hours as a fraction of the production rate. The right ratio depends on the domain half-life, but it is never zero.

Treat AI confidence as evidence of nothing. When a model gives you a confident answer, that confidence is not a signal of correctness. It may be the opposite. Build verification into the workflow at the points where being wrong is expensive — citations, claims about external systems, anything load-bearing. The confidence inversion is not a bug to be patched; it is a property of the architecture.

The chromosome textbooks of 1923 to 1956 were not badly written. They were, by the standards of their day, careful, peer-reviewed, and authoritative. They were just wrong, and nobody noticed for thirty-three years because nobody had a process for asking whether the count should be re-verified. We have better processes now. We just have to remember to use them, on the documents we wrote eighteen months ago, before the world has moved on without us.

Sources: Arbesman, The Half-Life of Facts, Current, 2012; Poynard et al., “Truth Survival in Clinical Research: An Evidence-Based Requiem?”, Annals of Internal Medicine 136(12):888–895, 2002; Neimeyer et al., “The Diminishing Durability of Knowledge in Clinical Psychology,” Review of General Psychology 18(4), 2014; Charette, “An Engineering Career: Only a Young Person’s Game?”, IEEE Spectrum, 2013; NBER R&D depreciation estimates (Li & Hall, 2020); Pew Research Center, “When Online Content Disappears,” 17 May 2024; Lendvai & Sasvári, Journal of Information Science, 2025; de Solla Price, Little Science, Big Science, Columbia University Press, 1963; University of Illinois WINELL system overview, 2024; Tjio & Levan, Hereditas 42:1–6, 1956. Several industry figures (Stack Overflow internal-doc decay, the 19% six-month hallucination uptick, the 30% RAG mitigation, the 34% confidence-language gap) are reported as directional from secondary sources where the primary methodology was not fully disclosed.