The Scarcity Shift

Evidence Ledger

The numbers behind the shift

The Scarcity Shift is my point of view, but it's built on evidence, and this is where I keep it. Each figure links to its source, notes how closely I read it, and flags its caveats.

Every claim should be checkable, so this is where you keep me honest. When a figure changes, this page changes with it, and the changelog at the bottom says what moved and when. Some entries cut against my own argument. They stay.

Last updated ยท 14 entries

Knowledge

−17%

on the exam for students who practiced math with an answer-giving AI

About 1,000 high school students practiced with AI. A plain chat assistant lifted their practice scores by 48%, then left them 17% worse on the exam once it was gone. A version that gave hints instead of answers lifted practice by 127% and did no harm. Same technology, different design.

Source
Bastani et al., Generative AI Without Guardrails Can Harm Learning, PNAS, 2025 (author copy)
Read
In full
Caveat
One country, one subject, high school students.

Knowledge

319

knowledge workers, 936 real examples of AI use at work

The more people trusted the AI, the less critical thinking they reported doing. The more they trusted themselves, the more they did.

Source
Lee et al., The Impact of Generative AI on Critical Thinking, Microsoft Research and Carnegie Mellon, CHI 2025
Read
In full
Caveat
Self-reported survey. It shows a link, not a cause.

The scarcity shift

19%

below trend: employment for 22 to 25 year olds in the most AI-exposed jobs

The drop shows up where AI automates the work, and it comes through fewer hires rather than layoffs. Where AI complements the work, employment held flat or rose. The first rung of the ladder is where the shift lands first.

Source
Brynjolfsson, Chandar and Chen, Canaries in the Coal Mine? Six Facts, Stanford Digital Economy Lab, revised Aug 12, 2026
Read
In full
Caveat
Working paper based on payroll records. Not yet peer reviewed.

The scarcity shift

33 months

after ChatGPT launched, Yale found no discernible disruption to the overall US labor market

This one argues against my thesis, which is why it is here. At the level of the whole economy, the change was not yet visible in late 2025. The shift may be real and still be early.

Source
Evaluating the Impact of AI on the Labor Market, Yale Budget Lab, Oct 1, 2025
Read
In full
Caveat
Aggregate view. The Lab publishes updates, and a newer one may say something different.

Value

92%

of S&P 500 market value now sits in intangible assets, up from 17% in 1975

Most of what companies are worth was already in knowledge, relationships, brand and know-how before AI arrived. AI raises the stakes on who owns those assets and who carries them out the door at night.

Source
Intangible Asset Market Value Study 2025, Ocean Tomo
Read
In full
Caveat
Market value minus book value. A proxy, not an audit.

Value

9x to 900x

per year: how fast the price of AI falls for the same level of performance, depending on the task

Answers are getting cheaper at a rate no other input to business has ever seen. That is what abundance looks like when you measure it.

Source
LLM Inference Prices Have Fallen Rapidly but Unequally Across Tasks, Epoch AI, Mar 12, 2025
Read
In full
Caveat
Price per unit, not total spend. Company AI bills can still go up, and many have.

Value

−19 pts

accuracy for consultants using AI on a task just outside what AI does well

Across 758 BCG consultants, AI on the right tasks meant 12.2% more tasks done, 25.1% faster, at over 40% higher quality. On a task just past AI's edge, they were 19 percentage points less likely to get it right. Knowing where the edge is turns out to be the skill.

Source
Dell'Acqua et al., Navigating the Jagged Technological Frontier, Harvard Business School and BCG, 2023
Read
In full
Caveat
One firm, designed tasks, 2023 models. The edge has moved since.

Ownership of outcomes

$0.99

per resolved customer issue, and nothing when the AI hands off without a resolution

Intercom prices its AI agent by outcome instead of by seat. When software is paid for results, someone has to define what counts as a result, and own it.

Source
Fin AI Agent pricing, Intercom, 2026
Read
In full (vendor page)
Caveat
List price at time of reading. Vendor prices change; check the page.

Ownership of outcomes

Human only

US copyright requires a human author, and the Supreme Court let that stand

A federal appeals court ruled in March 2025 that a work made by AI alone cannot be copyrighted. The Supreme Court declined to take the case on March 2, 2026. If you want to own the output, a person has to shape it.

Source
Supreme Court Denies Certiorari in Thaler v. Perlmutter, Baker Donelson, Mar 2, 2026
Read
Summary read (law firm analysis of the ruling)
Caveat
US law only. Work made with AI assistance can still qualify when a person selects, arranges or modifies it.

Human capital

7.8%

less predictive: what a freelancer's track record says about whether they get hired

Across about 50,000 freelancers on Upwork, traditional human-capital signals carried less weight in hiring decisions, and price carried more. The resume is losing pricing power.

Source
Siddiq and Zhang, The Declining Power of Your Human Capital in the Age of AI, UCLA Anderson Review, Aug 12, 2026
Read
Summary read (the school's write-up of the research)
Caveat
One platform. Freelance markets move faster than payrolls.

Human capital

170M / 92M

jobs created and displaced by 2030, according to employers surveyed by the WEF

A net gain of 78 million jobs, if employers are right. The catch: 59 out of every 100 workers will need training to get there, and 41% of employers plan reductions where AI can do the work.

Source
Future of Jobs Report 2025, World Economic Forum, Jan 2025
Read
Summary read (press release; full report linked from it)
Caveat
Employer projections, not measurements.

The human agent

19% slower

for experienced developers using AI, who believed it made them 20% faster

Sixteen seasoned open-source developers worked 246 real issues on code they knew well. With AI tools they took longer, and came away convinced they had been quicker. How it feels is not a measurement.

Source
Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, METR, Jul 10, 2025
Read
In full
Caveat
Small sample, early-2025 tools, experts in familiar code.

The human agent

+34%

productivity for the least experienced support agents with AI help, and almost nothing for the best

Across 5,179 customer support agents, AI raised productivity 14% on average. It helped novices most and experts least, because it was spreading what the experts already knew. Which raises the question of who teaches the AI next.

Source
Brynjolfsson, Li and Raymond, Generative AI at Work, NBER and Quarterly Journal of Economics, 2023 and 2025
Read
In full
Caveat
One company's support operation.

Changelog

  1. Ledger opened with 14 entries across the six threads.

Found a newer number?

Reply to any issue of the newsletter with the source. If it holds up, it goes here with a note in the changelog.

Subscribe free