Generated by Codex with GPT-5

Techmeme surfaced this May 29, 2026 essay in its AI Dark Output cluster, and the original is SemiAnalysis’s AI Dark Output: The Visible Cost of Invisible Output.

The article asks a harder question than whether companies are spending too much on AI: what happens if the value created by AI becomes real before the economy knows how to count it?

The costs are easy to see. Data centers, GPUs, electricity, water, land, software subscriptions, and token bills all appear somewhere in the ledger. Labor displacement is visible too, especially when hiring slows or routine roles disappear. But the output side is murkier. If an AI system drafts a legal document, summarizes six months of email, produces a literature review, or handles an internal administrative task for a few dollars or cents, much of the useful work may never appear in economic statistics.

SemiAnalysis calls this missing value “dark output.” The term is useful because it separates two questions that are often collapsed into one. One question is whether AI is producing valuable work. The other is whether GDP, inflation data, labor statistics, and industry accounts can recognize that value. The answer to the first can increasingly be yes while the answer to the second remains no.

Services Are Hard To Count

The measurement problem is not unique to AI. Economists have long struggled to measure productivity in services because many services do not have a clean physical unit. A factory can report that it produced more screws at a lower cost. A law firm, hospital, consulting company, or research group cannot report output in tons or barrels. Statistical systems often have to infer output from receipts, wages, hours, and sampled prices.

AI puts unusual pressure on that method because the price of some mental work can collapse very quickly. SemiAnalysis uses a basic legal document as an example. If a lawyer drafts it, the transaction appears as legal-services revenue. If software drafts it internally for a small token cost, the document still exists and may still be useful, but most of the visible transaction disappears. The statistics may record lower legal-services output and a small amount of software spending even though the same task was completed.

This can also distort inflation measures. If AI handles the simpler legal work, the remaining work sent to lawyers will tend to be more difficult and expensive. Average legal prices can rise because the cheap tasks have left the sample, not because every lawyer raised prices. A dataset built around receipts and sampled prices can read the transition as falling output and rising inflation when the underlying story is a large productivity gain.

Manufacturing automation was easier to observe because statisticians could still count the objects. When the cost of a screw falls, output can rise and the quantity is measurable. There is no comparable standard unit for a literature review, a code review, an email summary, or a legal memo. Tokens are not a substitute. A million tokens can produce noise, a useful document, a business decision, or a failed agent loop. The economic value depends on the result, not the input count.

Three Kinds Of AI Output

SemiAnalysis divides the problem into several categories.

The first is substitution dark output: work that humans previously did for money and AI now performs cheaply. If an externally purchased \$10,000 service becomes an internal workflow that consumes \$10 of tokens, measured GDP can decline even if the organization receives the same practical result. The transaction boundary moved, but the work did not vanish.

The second is new dark output: useful work that was rarely purchased before because it was too expensive. A company may run a literature review before every project, summarize a large email archive before a meeting, or ask an agent to inspect a codebase for a narrow question. These are not necessarily labor replacements. They are tasks that become worth doing only after the marginal cost falls dramatically. The resulting value is real, but the only visible market signal may be the token bill.

The third is captured AI output: work that remains visible because the provider can keep charging roughly the old price even after its production costs fall. If a company still pays \$10,000 for an HR service that is now largely produced by AI, the service remains in the national accounts and the provider records a larger margin. This output is easier to count, but it may be the exception where competition does not quickly force prices down.

These distinctions matter because “AI productivity” is not one economic event. A cheaper service, an internal automation, a new low-cost workflow, and a high-margin AI vendor can all use similar models while appearing very differently in the data.

The Statistical Failure Modes

The essay names four ways output can go dark.

Boundary shift happens when work that used to be purchased moves inside a company or household. A paid research brief becomes an internal prompt. The useful result survives, but the transaction that made it visible disappears.

Price collapse happens when AI sharply reduces the cost of a service that lacks a stable quantity measure. The statistics can see that receipts fell, but not that far more work is now being completed at a much lower price.

Sector misrouting happens when AI creates value in one part of the economy while the visible revenue appears elsewhere. A hospital may process paperwork faster with an AI tool, but the recorded transaction sits in the software vendor’s sector. The software industry appears more productive while healthcare can still look stagnant.

New-work invisibility happens when AI makes previously uneconomic tasks routine. A meeting dossier assembled for pennies can improve a decision, but there is no receipt for the value of being better prepared. Only the cost of the model call is easy to observe.

SemiAnalysis argues that the fingerprints of this transition may show up indirectly. Employment can fall in exposed occupations while average wages rise because lower-paid junior roles leave the sample. Token usage can surge in a field without a corresponding break in measured sector output. The economy sees the spending and some of the displacement while missing much of the new surplus.

The article is careful not to turn this into an unsupported claim that mass replacement has already happened. SemiAnalysis says its monitor identifies roughly \$1.5T in tasks with credible augmentation or automation potential, not \$1.5T of labor that has already disappeared. Its evidence currently points more toward augmentation than wholesale replacement. The dashboard is a map of economic pressure, not a layoff forecast.

A Familiar Blind Spot Gets Larger

The essay also connects dark output to a longstanding weakness in national accounting: useful work often disappears from GDP when no market transaction occurs. Care economics has documented this problem for decades. Household labor, caregiving, and other unpaid work can create substantial value while remaining largely invisible in the headline statistics.

AI widens the same gap. Information work that once crossed a market boundary can move into the unpriced part of the economy. A task can still be produced, and it can still matter, while the visible transaction shrinks to a software subscription or a few tokens. The production boundary has not changed, but the technology can move much more activity across it.

That has practical consequences. Investors use macroeconomic data to decide whether an AI boom is productive or speculative. Policymakers use it to balance inflation, employment, growth, taxation, and infrastructure demands. Companies use it to decide whether to hire, automate, or invest. If the data can see the costs more clearly than the output, each group risks making decisions with a distorted picture.

Takeaway

The value of the “dark output” idea is not that it excuses AI spending or dismisses the visible costs. It is that it demands a fuller accounting.

AI infrastructure consumes capital, electricity, water, and land. Workers can be displaced. Some token spending will be wasteful, and some automation projects will fail. Those costs should remain visible and contested. But a serious assessment also has to ask what the tools produce, including work that moves inside organizations, falls sharply in price, appears in the wrong sector, or becomes possible only because the cost dropped close to zero.

The economy already struggles to measure services and unpaid work. AI makes that weakness more consequential because it can turn expensive cognitive tasks into cheap, abundant inputs faster than existing statistics can adapt. Cheap screws became countable output. Cheap AI work may not.

That leaves a demanding measurement agenda. Token counts are inputs, not outcomes. Revenue captures only the work that still crosses a market boundary. Labor statistics reveal some pressure but not the full value created. If AI is becoming a general-purpose technology, the next challenge is not merely building more of it. It is learning how to see what it does.