The Half-Life of Knowledge: What the Evidence Actually Shows About CS and AI

TheThe half-life of a computer science degree is five years." "AI skills expire every two years." "Medical knowledge doubles every 73 days." These numbers circulate constantly, always with the same implication: things are worse now, and worst of all in technology. Some of that is true. Most of the specific numbers behind it are not what they're presented as — and the field that actually has the best-documented evidence, oddly, isn't AI at all, and it contradicts the popular story.

Where the metaphor comes from

TheThe phrase itself is old: economist Fritz Machlup coined "half-life of knowledge" in his 1962 book Knowledge Production and Distribution in the United States, borrowing the term from radioactive decay to describe how long it takes for half the knowledge in a field to be superseded. It stayed a mostly academic idea until Samuel Arbesman's 2012 popular book The Half-Life of Facts put a number-friendly, shareable frame around it — the one now recycled across a thousand blog posts and corporate learning-and-development decks.

The metaphor has a real crack in it. Radioactive half-life is a fixed physical constant, set by the isotope alone. Knowledge half-life shifts with the field, the decade, and who's measuring it. A fact can be superseded, but a foundational result more often gets absorbed rather than discarded — Newtonian mechanics got re-scoped to the regime where it still holds when relativity arrived, rather than being erased by it. Applying a decay curve built for isotopes to something that can also accumulate, get re-contextualized, or survive intact for a century is already an approximation, and how good an approximation depends enormously on which kind of "knowledge" you mean.

Three kinds of knowledge, one word

MostMost of the apparent contradiction in this topic dissolves once "knowledge" is split into the three things it's actually being used to mean:

Foundational. Architectural principles, proofs, core results. This tier gets extended and re-scoped far more often than it decays, and can sit undisturbed for decades or, in mathematics, close to a century.

Applied / methodological. General practice, professional skill, "how to do the job." This is where most of the citation-half-life literature (psychology, economics, medicine) actually sits, typically five to ten years.

Ephemeral / tooling. The specific API, framework, model version, or leaderboard standing in front of you right now. This is the layer that can turn over in months, and it's also the layer almost every viral "half-life" statistic is secretly describing while using language borrowed from the other two tiers.

The diagram below places a representative set of domains and subdomains on a single (logarithmic) time axis, grouped by field. The pattern worth looking for is not any single dot — it's that computer science and AI is the only field here with dots in all three tiers at once, from a five-month release cadence up to quarter-century-old theory, while the other fields shown cluster much more narrowly.

View the chart data as a table

Domain	Field	Tier	Approx. half-life
Top mathematics journals	Mathematics & Physics	Foundational	~90 yr
Theoretical physics	Mathematics & Physics	Foundational	~30 yr
Theoretical CS / core algorithms	Computer Science & AI	Foundational	~25 yr
Transformer architecture (2017–)	Computer Science & AI	Foundational (illustrative)	9 yr old, not yet decayed
Engineering degree (1960s cohort)	Computer Science & AI	Applied	~10 yr
General software-engineering practice	Computer Science & AI	Applied	~5 yr
Software-engineering skills, today	Computer Science & AI	Ephemeral	~4 yr
Agentic-framework "generation"	Computer Science & AI	Ephemeral	~1 yr
Frontier LLM release cadence	Computer Science & AI	Ephemeral	~4–6 months
Top medical-journal citation age (1996–2020)	Medicine & Biology	Applied	~7.2 yr, stable
Economics	Social Science & Economics	Applied	~9 yr
Psychology (2016 Delphi poll avg)	Social Science & Economics	Applied	~7 yr
Finance	Social Science & Economics	Applied	~5.5 yr

The rigorous evidence says the opposite of the popular story

CitationCitation age is the one part of this that's actually been measured properly, and the best recent study is a 2023 BMJ Open paper by Chow et al., which tracked citation ages across 726 articles in eight top medical and science journals (BMJ, JAMA, NEJM, The Lancet, Annals of Internal Medicine, Nature, Science, PNAS) from 1996 to 2020. If the "everything's accelerating" story were right, you'd expect citations getting steadily younger over that period. They found the opposite: citation lag was essentially stable, with a small increase in the use of older references over the last decade. More than 70% of references in medical journals were still within 10 years old throughout the study period, and that share held steady across it.

That result makes more sense once you separate the tiers above: what top journals cite is overwhelmingly applied and foundational work — the layer that was never expected to churn in a couple of years even under the popular narrative. The study says nothing about how fast a specific drug protocol or diagnostic tool becomes outdated in practice (ephemeral-tier questions it wasn't designed to answer), and it directly undercuts the most commonly repeated version of the medical half-life claim: that medical knowledge's "half-life" fell from 50 years (1950) to 7 years (1980) to 3.5 years (2010), and was "on track" to hit 73 days by 2020. That figure traces back to a single 2011 essay by physician Peter Densen. It's a doubling-time estimate for the total volume of published medical literature — a different metric from citation decay or fact-falsification, dressed in similar language — and it's the version that gets quoted everywhere because 73 days makes a far better headline than a flat 25-year citation-age trend.

Where the fast-decay story genuinely holds up: applied CS, not "AI" as a topic

ComputerComputer science is not one thing here, and the diagram above is the actual finding worth taking seriously: it's the only field shown with a foot in all three tiers. Split it into layers and the picture gets much more specific:

Theoretical CS and mathematics decay slowly, like theoretical physics. Citation half-lives in the hard theoretical sciences commonly run 20–40+ years, with some top mathematics journals exceeding a century — consistent with the "accumulates rather than decays" pattern above. A proof doesn't go stale.

Applied/tooling knowledge in software genuinely does churn fast, and there's a real historical trend line for it. A widely cited 1966 IEEE Spectrum piece on "technical obsolescence" put the half-life of an engineering degree at roughly 35 years for a 1920s graduate, falling to about a decade by 1960 (worth flagging: that 1966 figure is itself usually cited second-hand through later retrospectives rather than the original piece, so treat the exact numbers as an oft-repeated estimate, not a hard data point). By the 2000s, software-specific estimates from practitioner literature (e.g. Kruchten's 2008 IEEE Software piece putting the half-life of software engineering ideas at roughly five years) had that number down under a decade, and current industry estimates cluster around 2.5–5 years for technical skills generally.

The AI-specific "two-year half-life" figures are the least rigorous of the bunch. They come almost entirely from industry blog posts and vendor content, Salesforce among them, rather than bibliometric studies of the kind Chow et al. did for medicine. The practitioner-level observation behind them — that a framework, a model API, or a specific prompting technique can be obsolete within a couple of years — is plausible and matches everyday experience. It's a different kind of claim, though, than "the field's citation half-life is two years," and treating anecdote-grade numbers as if they carried BMJ-Open-grade rigor is exactly the pattern this whole topic keeps falling into.

Language models and agentic tooling: the fastest layer yet, sitting on the slowest

LargeLarge language models are the sharpest illustration of the three-tier split, because both extremes are visible in the same eight-year window. The architecture underneath almost every frontier model today — attention in place of recurrence, from the 2017 paper "Attention Is All You Need" — has spent eight years being scaled, tuned, and extended, which is exactly the foundational-tier pattern described above. Meanwhile the layer built on top of it is the fastest-churning thing in this entire piece: multiple labs now ship major frontier model updates every few months (Anthropic alone has released around two dozen Claude versions since launch), and METR's task-completion-horizon benchmark — a methodologically grounded, continuously updated measure of how long a task an AI system can reliably complete — found that horizon doubling roughly every seven months as of its March 2025 analysis, with some more recent readings suggesting the post-2023 doubling time has compressed to around four months. Benchmarks follow the same pattern from the other direction: GPQA was considered a serious challenge for frontier models in early 2024 and was largely saturated within about a year.

Agentic tooling shows the same churn one layer up the stack: practitioners commonly describe three distinct "generations" of agent framework emerging in three years — retrieval-augmented generation, then structured agentic workflows, then autonomous tool-calling loops — and LangChain, still one of the most-starred projects in the space, is now openly and publicly argued to have been overtaken by the problem it was built to solve moving on without it. None of this is measured with anything like Chow et al.'s rigor — it's practitioner consensus, which is exactly why it belongs in the ephemeral tier rather than being quoted as if it says something about "AI knowledge" as a whole.

So what's actually changed, historically

TwoTwo real, separable trends sit underneath the folklore. First: applied and ephemeral technical knowledge specifically — the tooling, framework, model-version, and API layer — really has compressed, from a multi-decade half-life for an engineering degree in the mid-20th century down to a few years for general software skills and, at the sharpest edge, a matter of months for frontier-model and agent-framework specifics. The 1966-to-now trend line for engineering degrees is a genuine, if second-hand-sourced, historical signal, and METR's benchmark is a genuine, well-documented modern one. Second: foundational and theoretical knowledge — in mathematics, in the core of computer science (the transformer paper being the freshest example so far), and apparently in general medical/scientific citation practice too, per the one study that actually measured it — doesn't show that pattern at all. The applied and ephemeral surface is what's speeding up. Conflating that surface with the field underneath it is how a physician's 2011 back-of-envelope doubling-time table ends up cited as evidence that a computer science degree is worthless in five years.

What this means for how we teach

TheThe tiers above point at a fairly direct implication for curriculum design, and it follows from timing more than pedagogy. A degree programme is written years before it's taught and taught over several more years after that: a syllabus fixed today reaches a graduating student half a decade from now. That lag is survivable for foundational material, which is durable enough to still be true on the far side of it — a quarter-century, in the theoretical-CS figure above, or close to a century for the top mathematics journals. The same lag is fatal for ephemeral material. A framework, a model version, or a specific agentic-tooling pattern taught today can be gone before the student who learned it graduates, which is the same mechanism behind the engineering-degree half-life falling from 35 years to a decade across the 20th century, just compressed further.

That gives formal teaching and a student's own ongoing learning genuinely different jobs. Depth — the kind a curriculum can actually deliver, built and checked over months, examined, built on in later courses — belongs on the foundational tier: algorithms, complexity, the mathematics under a model, the architectural principles that were still standing a quarter-century after publication. The ephemeral tier asks for something else entirely: awareness of what exists, the judgement to tell a genuine advance from a repackaged one, and the habit of picking up a new tool quickly when a task calls for it. A programme that tries to teach the ephemeral tier with the same depth it gives the foundational one is spending its scarcest resource — instructional time that took years to design — on material with a half-life measured in months. A programme that teaches students how to evaluate and acquire that tier for themselves is training a skill with the same durability as the foundational material itself.

Related Topics

• Running Agentic LLM Workflows Locally — the practical side of the agent-framework churn discussed above: what it's actually like to build on a layer that's still finding its shape.

References

1. Fritz Machlup, The Production and Distribution of Knowledge in the United States, Princeton University Press, 1962; concept summarized at "Half-life of knowledge", Wikipedia (including its own note on the unproven exponential-decay assumption).
2. A. Chow et al., "Does knowledge have a half-life? An observational study analyzing the use of older citations in medical and scientific publications," BMJ Open, 2023. pmc.ncbi.nlm.nih.gov/articles/PMC10231019
3. P. Densen, "Challenges and Opportunities Facing Medical Education," Trans Am Clin Climatol Assoc, 122:48–58, 2011. Source of the 50yr/7yr/3.5yr/73-day medical doubling-time figures. researchgate.net/publication/51231641
4. R. N. Charette, "An Engineering Career: Only a Young Person's Game?", IEEE Spectrum. Source of the 35-year/10-year engineering-degree figures (itself citing a 1966 IEEE Spectrum piece on "technical obsolescence"). spectrum.ieee.org/an-engineering-career-only-a-young-persons-game
5. "Expected Citation Rates, Half-Life and Impact Ratio," Clarivate/Web of Science Group. How cited half-life is defined and measured as a bibliometric indicator. clarivate.com/webofsciencegroup/essays/expected-citation-rates-half-life-and-impact-ratio
6. "Task-Completion Time Horizons of Frontier AI Models," METR. The doubling-every-few-months capability benchmark. https://metr.org/time-horizons/
7. A. Vaswani et al., "Attention Is All You Need," NeurIPS 2017; summarized at "Attention Is All You Need", Wikipedia.
8. "On Agent Frameworks and Agent Observability," LangChain. On the pace of agent-framework churn, from inside one of the frameworks itself. langchain.com/blog/on-agent-frameworks-and-agent-observability