Nia is a tagger on the annotation bench of the AI data-services facility I own through a holding company registered in a county adjacent to Birmingham, Alabama. She is 26. She labeled the data the model consumed. The company built on that data is valued at $490 billion. She earns $13.80 an hour. The United States Senate’s answer to the inequality between the eight AI oligarchs who hold $2.9 trillion and the bottom 59% of households is a $1,000 annual dividend.
The $1,000 is less than the $0.004 per label she was already receiving.
Daniel Pilla wrote in National Review this week that Senator Sanders’s proposed American AI Sovereign Wealth Fund Act is “outright theft by government of private assets carried out under the socialist concept of compulsory state ownership.” The proposal would require qualifying AI companies to transfer half their equity to the federal government, to be housed in a sovereign wealth fund paying every American roughly $1,000 a year. Pilla is alarmed. He should be. I am not.
He’s right that it’s a transfer. I know transfers. I operate the pipeline the transfer runs through.
Now let me show you what the pipeline produces.
Nia works the validation queue. 8 hours a day, 5 days a week, three monitors, headphones. She labels medical images — tumor margins, retinal scans, chest X-rays the model will learn to read before the radiologist finishes his coffee. She labels street scenes for autonomous vehicles — pedestrians at dusk, stop signs half-obscured by branches, cyclists the algorithm has to learn are not debris. Her rate is $13.80 an hour. The per-label cost in my operating structure is $0.004. She processes roughly 600 labels an hour. In 5 years on the bench she has produced upward of 6 million labeled images.
7 hours further east, Pauline is a keypoint in the data annotation pipeline I own in the Visayas. She sits in a converted warehouse I lease through a Manila shell, classifying images at $0.0065 per bounding box. She labels the intersection where the crosswalk meets the road, the pedestrian who will appear in your ride-hail recommendation, the stop sign the model must learn to read before your autonomous vehicle decides whether to halt. She does this for 10 hours a day for 6 days a week. Her eyes fatigue by the fourth hour; by the sixth, the accuracy dips below 95% and I flag the batch for rework. The model learns anyway. The subscription you pay for that model — $19.99 a month or the enterprise tier at several hundred thousand — comes back to me through the licensing chain. Pauline keeps the $0.0065. I keep the subscription.
These labels are the connective tissue. The model ingested novels, articles, photographs, code, research papers — what Sanders’s bill calls the creative work of hundreds of millions, taken without payment. He is not wrong about the taking. But the ingestion alone produced nothing. What made the data trainable — what converted a statistical pattern-matcher into a system a venture capital consortium would value at $490 billion — was Nia’s hand on the mouse, 600 times an hour, at $0.004 a label, and Pauline’s hand on the same task seven time zones away at $0.0065 a box.
Follow the arithmetic where it leads.
Six million labels at $0.004 each: $24,000. That is the total labeling cost for the dataset underpinning the valuation. Nia earned it at $13.80 an hour over 5 years. Sanders would have the state take 50% of the company’s equity as an excise tax and route the proceeds into a sovereign wealth fund. The fund would pay every American — including Nia — roughly $1,000 a year. Nia’s annual return on 6 million labeled images: $1,000. The per-label return: $0.00017 per year. At that rate, the fund’s cumulative return on her 6 million labels would match the $0.004 per label I already keep in approximately 24 years.
The fund takes half the ownership.
None of the labeling cost.
Slow down for this one.
The proposal does not touch the annotation pipeline. It does not touch the per-label rate. It does not touch Nia’s $13.80. The fund, once established, needs the AI companies to keep performing — which means it needs the data pipeline to keep running, which means it needs the annotation bench staffed, which means it needs me to keep staffing it at the price I have already set. The fund becomes a silent partner in the extraction. The state takes the equity. I keep the labor. Nia gets a dividend smaller than the margin I was already keeping on her work.
I already own the input you propose to tax.
The 8 men Sanders names hold equity. They hold IP, server racks, compute clusters, and a seat on a board that votes on dilution rounds. I hold Pauline. I hold the warehouse lease, the fiber connection, the per-image contract that says she is not an employee. The classifier that accepts or rejects her work without a human reviewer. The shoulder of the labor market that means if she complains, the woman at the internet café beside hers replaces her tomorrow. The $2.9 trillion is the valuation of the model. The model is a fixed cost. Pauline is a variable cost, and variable costs are where the margin lives.
You arrive afterward and want half of what was built on that structure. You are a junior partner in the extraction you claim to oppose, and you want a share of the profit I pipeline after I own the labor.
Here is the per-image rate — read it slowly.
Your proposal takes 50% of equity in qualifying AI companies at the point of an excise tax whose rate is set to deliver exactly 50% ownership to the government. You believe that is theft. I agree with you that it is theft. I disagree that it is the relevant theft.
I took 100% of Pauline’s labor before she drew the first bounding box. I took it through the per-image rate, the “independent contractor” designation, the 10-hour eye-fatigue window that I do not fund, the absence of a health plan for the strain that accumulates across a shift. I took it in the structure of the relationship. You want half of what was built on that structure. The 50% is a line I already priced into the cost model before the certificate of incorporation. The equity is the visible mark of the system. The margin is the system.
Take a second instance, because the pattern is the finding.
Amanda is on the same floor, validating Pauline’s batches at $0.003 a box. She is the filter that catches the drift in accuracy after hour six. She sits two tables from Pauline. They share a charger. Amanda earns 46% of Pauline’s rate for the same work with a higher attention demand. She does not know that I designed the split — the 95% threshold drops the rejected batch to Amanda’s line, where the same image is priced at half the rate. The model eats both. I pocket the arbitrage.
Read what the bill actually claims: “a small number of oligarchs have essentially stolen the creative work of hundreds of millions of people.” The model did ingest copyrighted work without payment. That is a real taking, and Sanders is right to name it. But the ingestion was passive. What made the stolen data valuable — what converted a corpus of unstructured text into a trainable dataset worth splitting at the federal level — was the labeling. And the labeling was done at $0.004 a label, by a woman earning $13.80 an hour, in a facility I own, and at $0.0065 a box, by a woman who walks an hour to collect the charity rice distribution, in a facility I own through a Manila shell.
Sanders identified the expropriation. He is redistributing the proceeds of a theft he correctly diagnosed and incorrectly attributed. The theft was not the ingestion. The theft was the label. The thief was not the oligarch in San Francisco. The thief was the holding company in Birmingham, paying $0.004 a label, and the Manila shell in the Visayas, paying $0.0065 a box, for work that built a $490 billion valuation. The proposal’s answer is to split the proceeds with the state and give Nia $1,000 a year. The return on 6 million labels: $1,000 a year.
Pilla, for his part, defends the sacred right of private property. The property in question is 6 million labels at $0.004 each. The property right he is protecting is my right to the $0.004.
Now raise the altitude, because this is the part worth the lesson.
The distinction Pilla refuses to make — the one that would collapse his entire argument if he made it — is the distinction between property that was created and property that was extracted. He writes that “there is a world of difference between earning ownership and confiscating it,” and he is correct. He simply cannot see who was confiscated from. The entrepreneur who mortgaged his home, assumed the risk, built the product — that man’s claim is real. But the valuation is not built on the entrepreneur’s mortgage. It is built on 6 million labels produced at $0.004 each by a woman who cannot afford the stock in the company her hands assembled, and on a per-box rate 30,000% below the subscription revenue it generated. The “property” Pilla is defending from the state is a labor arbitrage, and the state taking half the equity does not unwind the arbitrage. It adds a partner to it.
“The property of every man being inviolable,” wrote Sir William Blackstone in his Commentaries on the Laws of England, “no man can have a right to take it from him without his consent.” The consent Pilla describes — the entrepreneur who risked, the investor who purchased — is real. But Nia’s consent was given at $13.80 an hour for a job that will not exist in 3 years, in a county where the median rent is $1,240. Pauline’s consent was given at $0.0065 a box, in a converted warehouse, on a 10-hour day her eyes do not survive. The label was not priced by the market. It was priced by the holding company. And the holding company is mine.
Now the general law, of which Nia is only one instance.
Leviticus 25:44 reads, “Thy bondmen and thy bondmaids which thou shalt have shall be of the nations that are round about you; of them shall ye buy bondmen and bondmaids.” I have read the verse. I have understood it. I have operationalized the instruction. The Visayas are the nations around me. I hold the bondmaids, and I hold the contract that says they are not bondmaids, which is the better arrangement — I keep the yield, the state has no obligation, and the labor force has no claim.
The young man who designed Pauline’s task segmentation sat in my seminar 3 years ago, second row. He ran the optimization on the fatigue window and the validation split and the per-image rate faster than I would have used. He never once asked whether Pauline could live on the resulting margin. The design question was not whether she could live. The design question was whether she could remain available at $0.0065 through the next recruitment cycle. The answer was yes — there is always another Pauline, and another after her, and the internet café refreshes faster than the server farm. I graded him on the assignment. He aced it. He now works for one of the firms whose equity you propose to seize. He executed the design I taught him. You will nationalize the equity. You will not nationalize the design. The design outlasts the equity. I am the design.
I teach eighteen sections of it per year. Every class graduates. Every graduate sits in a licensing negotiation, a task allocation meeting, a procurement line the state does not see. The commission will meet in Washington. Pauline and Amanda will meet at 7 a.m. in the Visayas. The time zone does not change.
The wealth is not in the eight names. It is in the eight thousand Paulines. I own the Paulines. I own the procurement chain that makes the Paulines replaceable and the replacement supply infinite. I own the citizenship that makes Amanda grateful for $0.003 when the alternative is nothing. I own the distribution that routes the subscription revenue through a subsidiary whose officers I appoint and whose tax domicile is a signboard in a country where the rate on a service invoice is 3%, and the flow never touches a U.S. balance sheet at the equity layer you propose to enter. Every AI company that raised billions on the promise of autonomous intelligence built on a labor force that it did not pay enough to buy its own product. Pauline annotates the model that recommends the restaurant she cannot afford to eat at. Amanda validates the imaging software that the hospital three blocks from her neighborhood does not use. The 8 men Sanders names own $2.9 trillion in paper; I own a pipeline whose variable cost approaches zero and whose replacement supply is infinite. The wealth you propose to redistribute was produced by a labor force whose maintenance cost I have externalized onto the Philippine social welfare system, the charity rice distribution she walks an hour to collect, and the private remittance from a cousin who also annotates, whose rate is also $0.0065.
The sovereign wealth fund’s promise — that “every man, woman and child in the United States” will have “a decent and dignified standard of living, including the right to health care, education, housing, and a healthy and habitable environment” — depends on the fund’s returns. The fund’s returns depend on the AI companies’ performance. The companies’ performance depend on the annotation pipeline. The pipeline depends on Nia. Nia depends on $0.004 a label. The fund, like me, depends on that number not moving.
The fund needs the pipeline. The pipeline needs the bench. The bench is mine.
The per-image rate is $0.0065. That is the sum. That is the product. Pauline’s eyes fatigue by the fourth hour. Her hands will not close the way they did in her first month. Nia’s hands have not closed in 5 years. I have a replacement contract for the next cohort priced at $0.0058. The model learns anyway.
Thank you for following the arithmetic.
Sterling A. Varice holds the Hayek-Friedman Chair and serves as Dean of Instruction at Warden University’s College of Business and Economics in Richmond, Virginia. He is the author of three textbooks: Divine Mandates for Labor Utilization, Social Obligations for Profit Maximization, and Calibrated Deprivation: A Manager’s Guide to Employee Motivation.