New Dark Output: The Work Nobody Did Before AI Made It Cheap
When a literature review drops from about $2,000 to about $2, we don’t do the same amount of work—we do vastly more. That value barely shows up beyond token bills.
Not all invisible AI value comes from replacing a lawyer or an analyst. The longer-run story may be the opposite: work that nobody would have paid a human to do, because the price was too high—until tokens made it almost free. That is new dark output: real usefulness with almost no clear trail in GDP.
When price falls a thousandfold, volume explodes
A full academic literature review once meant a PhD student and weeks of work—on the order of a couple thousand dollars—so you did it mainly at the start of a big project. With AI, a similarly broad pass can cost about two dollars. We do not keep the old quantity and bank the savings. We review before every small project. We summarize a candidate’s papers before an interview. We scan five years of a partner’s public filings before a meeting. Each task improves decisions and saves time. Beyond a few dollars of tokens, there is often no invoice, no contract, no payroll line.
The scarier property: more value, more invisibility
Anecdotes suggest a large share of token spend goes to this kind of new work, not to one-for-one substitution of existing jobs. Scale is unknown: transactions hide behind the anonymity of tokens. Even with full chat logs, valuing a given task is hard. National accounts mostly see AI vendors’ revenue—not the surplus created on the user’s side. If the task never existed in the priced economy before, the statistical system has no category waiting to receive it.
A fingerprint: heavy tokens, no mass unemployment
Anthropic’s economic index (March 2026 discussion in the source talk) reported that about 37% of tokens went to computer and math domains—yet software investment’s contribution to GDP had not clearly broken from prior trends. One reading: many tokens are not replacing software engineers one-for-one; they fund work that would not have been commissioned at old prices. Industries can burn tokens without a matching wave of layoffs when AI mostly expands what people can attempt.
- New dark output is not “fake work”—it is useful work that used to fail a cost–benefit test.
- GDP may register only the token bill, not the decision quality or hours saved on the user’s side.
- If critics only count visible costs and visible jobs, they systematically miss this surplus.
The more AI creates work that never had a line item, the more an AI boom can look, in the official tables, like nothing much happening—except a rising bill for tokens.