Attention is not the scarce thing anymore

I spent two weeks trying to name the economy we are in and could not get it down to one word. Creator economy, attention economy, experience economy, gig economy. Every few years another one arrives, gets a book, gets a conference track, then quietly stops explaining anything. I wanted to know whether the naming was doing any work at all, or whether it was consultants relabelling the same decade.

It does do work. Just not the work I assumed. Each of those names turns out to be a name for whatever happened to be scarce at the time.

The rule I ended up with

Herbert Simon wrote the cleanest version of this in 1971, in an essay about designing organisations for an information rich world. His line was that a wealth of information creates a poverty of attention.¹ Information got cheap, and that is exactly what made attention expensive.

So an economy gets named after its bottleneck rather than after what it produces. That turns the question inside out. Instead of asking which economy we are in, I started asking what just became free, and where the scarcity moved to.

There is a problem with leaning on Simon this hard. Jelle Bruineberg published a paper in Philosophical Psychology in 2025 arguing the whole attention economy literature rests on two assumptions that contemporary attention science rejects: that attention gets handed out by some central allocator, and that it is one uniform resource you spend on any task.² Tier A If he is right, and I think he mostly is, the naming rule is a filing system for evidence rather than a claim about how minds work. I am using it as a filing system.

The ledger

Three scarcities, stacked

When I ran the test on 2026 I could not get a single answer out of it. I got three, and they sit on top of each other, each one load bearing for the one above.

Bedrock: compute and energy

At the bottom the bottleneck stopped being capital or talent and became physics. Amazon, Microsoft, Alphabet and Meta have guided to somewhere between 720 and 745 billion dollars of capital expenditure for 2026. Amazon at 220 billion after raising it from 200 in July, Microsoft near 190, Alphabet at 180 to 190, Meta at 130 to 145 including finance lease payments.⁵ Tier B The year before was around 410 billion.

The International Energy Agency has data centre electricity roughly doubling, from about 415 terawatt hours in 2024 to around 945 by 2030, which is close to 3 percent of global demand.⁶ Tier B In the United States data centres account for nearly half of all electricity demand growth across that period. The binding constraint is not generation. It is interconnection queues, ramp rates and local capacity.⁷ The price pass through to households now gets modelled properly rather than argued anecdotally.⁸ Tier A

Extrapolating from any of this has failed before. Eric Masanet and colleagues showed in Science in 2020 that data centre workloads went up roughly sixfold across the 2010s while energy use stayed close to flat, because efficiency absorbed the growth.⁹ Tier A I had a draft of this essay that treated the buildout as a certainty. That paper is why I rewrote it.

The long range numbers are enormously wide too. One 2026 study puts global data centre demand between 1,800 and 5,000 terawatt hours by 2050, and notes that the uncertainty coming from efficiency dynamics inside the sector is wider than the uncertainty across completely different socioeconomic futures.¹⁰ Tier A That is a real admission. The sector's own engineering choices matter more than which world you assume it happens in.

This layer does not behave like software. It behaves like railway building in 1880. Enormous fixed capital goes in ahead of proven demand, and the argument about who overbuilt happens afterwards.

Middle: agents

Inference got cheap at a fixed capability threshold, and that qualifier is doing a lot of work. Stanford's AI Index tracks the cost of querying a model scoring 64.8 on MMLU, roughly GPT-3.5 level. It went from 20 dollars per million tokens in November 2022 to 7 cents in October 2024, using Gemini 1.5 Flash 8B.¹¹ Tier B More than 280 times cheaper in about eighteen months.

Epoch AI looked across six benchmarks and found the decline runs anywhere from 9 times to 900 times per year depending which capability you hold constant, median around 50.¹² Cheaper tokens do not translate into uniformly cheaper work.

The price curve is not the part I keep thinking about. The part I keep thinking about is who sits in the middle. Platforms in the attention economy sold your eyeballs to advertisers. This layer sells your intent to suppliers: an agent that already knows what you want, negotiating on your behalf, or getting paid to prefer somebody. Doc Searls described this in 2012 and called it the intention economy.¹³ It was waiting on an agent to exist. The architectures being proposed for it borrow directly from real time ad auctions.¹⁴

This is also where I have the least to stand on. Most of the serious work is framework building rather than measurement. A Google DeepMind team published a taxonomy in 2025 sorting agent economies by whether they emerge or get designed, and by how permeable they are to the human economy.¹⁵ Gillian Hadfield and colleagues surveyed the open institutional questions.¹⁶ Tier A Both are careful. Both are about what might happen.

Where agent markets have actually been tested they behave badly. In simulated markets with information asymmetry, LLM agents mostly failed to establish cooperation in one shot settings, expert fraud stayed entrenched across repeated rounds, and the consumer agents fixated on price while missing the incentives sitting inside the markups.¹⁷ Tier A Institutions built for humans do not port over cleanly.

Surface: trust

Generation costs nothing now, so output stopped being evidence of anything. The experimental literature here is unusually good, and it says something more specific than the industry line about authenticity.

Labelling something as AI generated lowers how credible people find it. Sacha Altay and colleagues ran two preregistered experiments with 4,976 people across the US and UK. Labelling a headline as AI generated lowered perceived accuracy and willingness to share it, whether the headline was true or false, and whether a human or a machine had written it.¹⁸ Tier A The effect was about a third the size of labelling something false. The aversion traces back to an assumption that AI generated means fully automated with nobody supervising.

Wording tracks that. Content described as AI influenced reads as more credible than AI assisted, which reads as more credible than AI generated.¹⁹

Then the finding that changed what I do. Benjamin Toff and colleagues ran a survey experiment using real AI generated journalism, and the trust penalty was largely cancelled out when the article disclosed the list of sources it had been generated from.²⁰ Tier A A separate preregistered study found the same shape. Thin labels depress credibility. Rich disclosure that names the sources, the editorial checks and the people involved recovers most of the loss.²¹

So what is scarce at the top is warrant, meaning disclosed process somebody can go and check. I had written authenticity in an earlier draft, which was lazy of me. Authenticity is a feeling, and feelings are cheap to simulate now. The practical version of this is uncomfortable if you were planning to add a badge and move on, because a thin disclosure is worse than none. It buys you the penalty without the remedy.

One deflating note before anyone treats disclosure as armour. Isabel Gallegos and colleagues ran an experiment with 1,601 Americans. The messages shifted policy views by about 9.7 percentage points on average, and the authorship labels had no significant effect on attitude change, on accuracy judgments, or on intention to share.²² Tier A Disclosure is an honesty practice. It does not stop the persuasion.

The fault

None of the three layers is evenly distributed, and this is where I am least willing to sound confident, because the evidence genuinely conflicts.

Pointing toward displacement

Analysis of 285 million Lightcast job postings between 2018 and 2025 found postings for occupations above the median AI substitution score fell about 12 percent relative to those below it. 6 percent in the first year, 18 percent by the third. Entry level roles requiring no advanced degree fell 18 percent, roles requiring no experience 20 percent, administrative support 40 percent.²³ Tier C On freelance platforms, workers in heavily affected categories lost both employment and earnings, and the top performers were hit disproportionately rather than protected.²⁴ Tier A A PRISMA review covering 94 studies found reductions of 14 to 41 percent in entry and mid level software and content postings between 2022 and 2024, though the authors are explicit that these span non overlapping designs and are not pooled estimates.²⁵ Tier A

Pointing against

Then Norway. Researchers used population wide administrative registers from 2015 to March 2025 and found no robust employment displacement among young workers in highly exposed occupations. Coefficients were negative but small and not significant. The detail that stopped me: running the same test at fake earlier dates produced larger effects than the real one, which points at trends already underway before ChatGPT shipped.²⁶ Tier A A broader review found task level productivity gains of 15 to 50 percent sitting alongside limited aggregate disruption.²⁷ Tier C And there is an older warning worth keeping in view. David Card and John DiNardo showed in 2002 that wage inequality flattened through the 1990s while computing kept advancing, which the skill biased technical change story could not account for.²⁸ Tier A

My best guess at reconciling these is that they measure different things. Postings measure hiring intention, which moves fast. Administrative registers measure employment, which moves slowly and counts everyone already holding a job. Both can be true at the same time. Task based models expect the asymmetry anyway: automation lifts the skill premium and can push real wages down for displaced workers even while total output rises, and the gains are not Pareto improving when retraining is expensive and hard to reverse.²⁹ Tier A

So I will say it at the strength I can defend. The fault shows up clearly in hiring. It is unproven in aggregate employment. Theory expects it to widen. That is weaker than the headlines and stronger than the sceptics.

What I am doing about it

Five things, roughly ordered by how confident I am in each.

Publish the receipts

Best evidenced move on the list. Disclosing sources and method measurably recovers the credibility penalty that AI involvement otherwise costs you.²⁰²¹ Working files, rejected directions, the reasoning, named people, so somebody can walk the trail backwards.

Never ship a thin disclosure

A bare AI generated label lowers trust and gives nothing back, and the aversion is specifically about unsupervised automation.¹⁸ If you disclose, and you should, disclose thickly enough to name the human judgment inside it.

Be readable by agents

Discovery is moving into that middle layer, and the architectures being proposed select on structured, attributable capability.¹⁴¹⁵ This one is speculative and I am weighting it accordingly. Structuring the work costs very little, so I am doing it anyway.

Do not build things that only exist inside somebody else's model

Everything at the bedrock layer is capital looking for rent on a spending increase of roughly three quarters year on year.⁵ Files a person can open, edit and hand to somebody else without a subscription are the hedge against that rent getting repriced.

Treat the fault as a forecast

Hiring data shows the gap. Population data has not confirmed it.²³²⁶ Position for it, because the theory expects it. Do not make irreversible bets on a result that has not replicated.

How this could be wrong

Three ways, one per layer, and I would rather name them than have somebody else do it.

The bedrock could be an overbuild, and it would show up in utilisation rates and depreciation schedules long before it showed up in a headline. Some of that signal is already sitting in the filings. Microsoft's adjustment to asset useful life and lease treatment moves its 2026 number from about 190 billion to roughly 175 billion on a capex plus finance lease basis, and around two thirds of the spend is against short lived GPU and CPU assets.⁵ Rebound is a live hypothesis for AI compute rather than a settled law. Sasha Luccioni and colleagues argued the Jevons case at FAccT in 2025,³⁰ while Steve Sorrell's review concluded the evidence that efficiency actually raises total consumption is far from conclusive.³¹ I used it as framing, not as a prediction.

The middle layer only becomes an economy if agents finish transactions instead of handing back to a person before every decision. If bounded autonomy holds, this is a software feature and agent attention never gets a price at all.

The surface has the sharpest problem, and it threatens what I just argued more than anything else here. The first independent formal methods analysis of C2PA concludes the specifications fail to meet their own claimed security goals, and should not yet be relied on for journalism, legal evidence or financial disclosure.³² Tier A No scheme currently combines robustness, unforgeability and public detectability.³³ Researchers have demonstrated an integrity clash, where an asset carries a cryptographically valid manifest asserting human authorship while its pixels carry a watermark saying it is AI generated, and both signals pass their own checks in isolation.³⁴ Tier A So provenance cannot be handed off to a standard yet. Warrant stays editorial and human for now, which is exactly why it is scarce, and if that tooling matures quickly this whole layer collapses back into brand.

That is where I have got to. A compute buildout with an agent interface bolted on, and the only durable human position on top of it belongs to whoever can show where their work came from. I have been wrong about this twice while writing it. Version three is above.