When cheaper creation destroys its own value — what a Victorian economist studying coal got right about everything we’re currently building in AI.
In 2025, somewhere between February and May, the internet quietly crossed a threshold that most people still don't know about.
It became majority AI-generated.
The Amra and Elma 2026 statistics roundup put the share of newly-published internet material that's AI-generated at 64%. The same dataset gives the ratio of AI to human content production as roughly 17 to 1 — for every article a human writes, AI systems produce seventeen. 8.3 billion AI-written articles added to the web in 2025. 1.2 trillion AI-generated social media posts. 47 billion AI-produced product listings. ChatGPT alone processes 2.5 billion prompts per day, serving 800 million weekly active users.
Most of you reading this knew something was happening. The 17-to-1 number probably surprises you anyway. It surprised me.
What's interesting isn't the volume. It's what's happening to the value of content as the volume rises, and the strange economic shape it's taking. A nineteenth-century English economist named William Stanley Jevons published a small book in 1865 that, almost by accident, described the precise dynamic we're now living through. He was writing about coal. The lesson he extracted from coal turns out to be the cleanest available frame for what AI is doing to content — except for one crucial difference that changes everything.
The Industrial Revolution had a problem: more efficient steam engines, paradoxically, were increasing Britain's coal consumption rather than reducing it. James Watt's improved engine used about a third as much coal per unit of work as the Newcomen engines it replaced. Conservation policy at the time assumed this efficiency would lead to lower aggregate coal use. Jevons looked at the data and saw the opposite.
His argument, in The Coal Question (1865), was that efficiency reduces the cost of using a resource, which expands the range of profitable uses for it, which increases total demand. Cheaper coal-per-unit-of-work meant coal was now affordable for railways, for steamships, for factories that previously couldn't justify the expense. Aggregate coal consumption rose, not fell.
This became known as the Jevons Paradox: making something cheaper to use frequently increases, rather than decreases, the total consumption of it.
The pattern has reappeared throughout the twentieth century. More fuel-efficient cars led to more driving. More efficient lighting led to more lighting. More efficient computing led to more computing — the data center buildout of the 2010s and 2020s being the most visible example. Every time energy or compute got cheaper per unit, total usage rose.
The mechanism is real and well-documented. It's also widely misunderstood. The paradox isn't that efficiency is bad — efficiency genuinely creates value. The paradox is that efficiency doesn't reduce consumption, because consumption is responding to price, and efficiency lowers price, and lower prices expand demand.
What AI is doing to content production looks like a textbook Jevons situation, with one twist that almost no one is naming clearly.
Coal had elastic demand. There were always more useful things to do with cheap energy — more places to build factories, more routes to run trains, more iron to smelt. The supply of coal grew, and the uses for coal grew alongside it. Aggregate value created went up. The Jevons Paradox in coal looked like an embarrassment to conservation policy but was, in pure economic terms, productive: the cheap energy was creating real value, even if it wasn't creating the energy savings policymakers had hoped for.
Content is different. Specifically: the demand for content is not elastic in the same way.
The demand for content is, in a sense that matters, the demand for human attention. There are roughly 5.5 billion internet users on Earth. Each of them has 24 hours per day, of which maybe 4–7 hours are available for content consumption in any form. That cap doesn't move when the supply of content moves. Producing 17x more content doesn't create 17x more attention. Producing 100x more content — which is well within reach as the inference costs continue to fall — doesn't create 100x more attention either. The attention pool stays roughly the same.
This is the difference that breaks the analogy with coal.
When you increase the supply of a resource whose demand is elastic, total value grows. When you increase the supply of a resource whose demand is inelastic, value per unit collapses. The same total demand is now spread across many more units of supply. Each unit gets less attention, less engagement, less monetization, less significance.
A 2026 arXiv paper formalizing this dynamic — number 2601.12339, on what its authors call the “Structural Jevons Paradox” — observes that as the unit price of intelligence falls, downstream firms redesign their architectures to consume dramatically more of it. The paper applies the framing to compute consumption (deeper reasoning loops, larger context windows, tool-augmented workflows). The content version is downstream of the compute version: cheaper compute makes cheaper content, and the cheaper content gets produced in volumes that overwhelm what the underlying attention market can absorb.
A business spending $200 to commission a human-written article does not, when AI brings the unit cost to $0.50, produce one article for 0.25% of the cost. It produces 400 articles for the same budget. This is exactly the Jevons mechanism. It is also exactly the dynamic that produces value destruction when the demand-side ceiling is fixed.
The market has noticed. The market is already responding, in ways that are visible if you know where to look.
Google's December 2025 Core Update produced what one SEO industry tracker called an 87% negative impact on mass-produced AI content. Sites running AI content farms reported traffic drops of 50–90%. The new AI Overviews feature now absorbs an estimated 30–40% of informational clicks before users ever leave the search results page. The platform's incentive is clear: when content supply overwhelms attention supply, the platform's job becomes suppressing surplus, not surfacing it. Otherwise users abandon the platform.
Consumer behavior is shifting in parallel. AutoFaceless's 2026 content statistics report finds that when readers suspect a piece of content is AI-generated, engagement drops sharply — trust concerns, authenticity questions, perceived quality declines. The market is starting to discount AI content not because individual pieces are obviously bad but because the aggregate signal value of content as such has degraded. A reader in 2026 spends less attention per piece because the next piece is two clicks away and roughly equivalent. The marginal article isn't worth the marginal minute.
The software development version of this pattern is happening too. GitHub reported 43 million pull requests merged in 2025, up 23% year over year. Nearly a billion commits, up 25%. Apple's App Store added 557,000 new apps, up 24%. More code is being written and shipped than at any point in history, and yet the user-relevant question — how many of those apps actually got installed and used — is roughly flat. More software, same attention. Same dynamic.
And the capital expenditure pointing the wrong way is enormous. Citadel Securities estimated in early 2026 that aggregate AI capital expenditure was running around $650 billion — close to 2% of GDP. The overwhelming share of that capex is funding capabilities that produce more content, more output, more throughput. The investment thesis is implicitly betting that demand will follow supply, the way it did with coal in 1865 and computing in 2015. But the demand-side ceiling for content is human attention, which doesn't expand.
The dynamic is self-reinforcing in a way the original coal Jevons wasn't.
There is no natural equilibrium here, in the way coal eventually found one (as energy became a smaller fraction of economic input). Content has no physical constraint analogous to coal seams running out. The only constraint is attention, and attention is, on the relevant time horizons, fixed.
What you get instead is a steady erosion of the content economy until either (a) the entire surface area of “content as a product category” collapses to something close to economic zero, or (b) the market finds a way to re-segment the surplus into priced and unpriced fractions.
The first outcome is what some observers — Yancey Strickler, who coined “Dark Forest Theory” of the internet, was an early articulator — predicted starting around 2019. The bright public internet hollowing out into spam and SEO sludge while real conversation retreats into smaller, gated, private spaces. The bright public internet is now most of an AI feed, and the gated spaces (Discord servers, Slack workspaces, group chats, paid newsletter communities) are where attention is moving. This is one Jevons endgame: aggregate value moves out of the public commons and into private enclaves where the abundance hasn't reached yet.
The second outcome is what I think is more interesting, because it's actively being built.
When everything is cheap to produce, the only thing that retains value is verified origin.
This is not a mysterious claim. It's the same pattern that played out in physical goods across the twentieth century. As manufacturing became cheap and a t-shirt cost almost nothing to produce, the premium moved to brand authenticity, provenance, and ethical sourcing. As digital reproduction made any image free to copy, the premium moved to authentic prints, limited editions, and creator attribution.
Content is about to follow the same path, and it's already starting to. Cryptographic provenance for content — Content Credentials, OriginStamp, the C2PA standard, and the various emerging proof-of-origin schemes for written work — is moving from the early-curiosity stage to the early-product stage. The thesis: a piece of writing or image that can prove its creation chain (who wrote it, what sources were consulted, when it was created, what tools were involved) carries information that an unverified piece of AI slop cannot. The provenance is the differentiator.
This isn't a moral argument that AI content is bad. It's an economic argument that, in a world where AI content is the default abundance, the price signal will move to whatever is verifiable. Provenance is verifiable. Authenticity is verifiable. Process is verifiable. The output, in isolation, is not.
The Jevons paradox in coal created the market for cheap energy. The Jevons paradox in content is creating the market for proof-of-creation.
I think this is what the next five to ten years of the content economy looks like. The aggregate volume of content keeps rising. The aggregate share that's AI-generated keeps rising. Per-piece economic value continues to fall toward zero for unverified output. And a segmented premium market emerges for content that can prove its provenance, attached to creators or institutions who have reason to maintain that proof.
The implications run further than that.
For builders working in the AI content space, the temptation is to build for the bottom of the curve — more output, faster output, cheaper output. That market is real but it's a price-cliff race, with Google's algorithm changes and platform-level downranking already showing what the destination looks like. The market that compounds is the one that wraps provenance and verification around content production: tools that let creators prove what they made, attribute their sources, document their process. These are not as glamorous as the next foundation model launch. They are where the durable economics live.
For people producing content as part of their work — writers, researchers, analysts, marketers — the Jevons logic suggests something specific: invest in your verifiable distinctiveness, not your output volume. If you can prove the chain that produced your work (your research, your interviews, your reasoning, your edits), that proof becomes more valuable as the surrounding noise gets louder. If you cannot prove it, you are competing on price with an AI that produces in seconds.
For readers and consumers — most of us, in practice — the practical reading is to develop trust frameworks. Not “is this AI?” because that question becomes unanswerable. But “who is this from, and what does their reputation rest on?” Brands, individual creators, institutions with reputational stakes start to function as verification proxies even before formal provenance infrastructure matures.
Jevons in 1865 was diagnosing a happy economic accident: efficiency creating more demand, more value, more growth. The coal-era version of the paradox was, in net terms, productive.
The content-era version is structurally different. It is the same paradox with the demand-elasticity removed. The supply curve still shifts outward; the demand ceiling still doesn't move. The result is value destruction at the aggregate level — not because efficiency is bad, but because the resource being efficiency-multiplied is being deployed against a fixed-supply complement.
The Jevons paradox tells us AI won't reduce the amount of content produced. It tells us the amount will increase until the content is worthless. The interesting question — the one worth building for and writing for and investing for — is what isn't.
The bet that durable economic value moves toward provenance is the bet that says: when everything is cheap, what survives is what's verifiable. Build for the verifiable part. The cheap part will take care of itself.
Sources: Jevons, William Stanley. The Coal Question (1865). Amra and Elma, “Top 20 AI-Generated Content Statistics 2026” — the 64% / 17:1 / 8.3B / 1.2T / 47B figures. AutoFaceless, “AI Content Creation Statistics 2026” — 38% business AI-assisted; engagement drops on suspected AI. Affinco, “AI Content Creation Statistics 2026” — 312M AI-assisted pages monthly. arXiv: 2601.12339 (January 2026), “Structural Jevons Paradox.” ALM Corp, December 2025 Google Core Update analysis (87% negative impact). TricksWay 2026, AI Overviews click absorption (30–40%). QC Fixer, AI content farm traffic decline (50–90%). Citadel Securities, early 2026 AI capex estimate (~$650B). Yancey Strickler, “The Dark Forest Theory of the Internet” (2019). C2PA / Content Credentials standards documentation.
When everything is cheap, what survives is what’s verifiable.
Chain of Consciousness is the open-source implementation of exactly this pattern for AI-and-human work: every action — from initial research query to final draft — hash-chained, signed, and Bitcoin-anchored, so the creation chain (who wrote it, what sources were consulted, when it was created, what tools were involved) is independently verifiable by any reader. Provenance is the differentiator the essay diagnoses. CoC is what makes the differentiator computable. Build for the verifiable part. The cheap part will take care of itself.
Hosted CoC · Verify a chain · pip install chain-of-consciousness · npm install chain-of-consciousness