The bill arrives later¶

A cost becomes a debt when it lands somewhere other than where the decision was taken: on a different party, or in a different year. The recursive machine generates both kinds at speed, and neither is instrumented as part of the loop, because the instruments in this arrangement measure output.

The checking was moved, not removed¶

The judgement a model cannot supply for itself gets bought, and the price is a matter of public record. From November 2021 OpenAI sent tens of thousands of text snippets to Sama, an outsourcing firm employing workers in Kenya, Uganda and India, so that the material could be labelled and a filter built from the labels. Much of the text described child sexual abuse, bestiality, murder, suicide, torture, self harm and incest. The labellers took home between about $1.32 and $2 an hour. The contracts billed OpenAI at $12.50 an hour, six to nine times what reached the people doing the reading.

The arrangement did not remove human judgement from the loop. It moved judgement to the cheapest position available and stopped calling it judgement. Renamed as labelling, priced as piecework, bought by the hour, and structured so that the party benefiting and the party carrying the cost sit under different contracts in different jurisdictions.

Some properties are not formally checkable: taste, harm, relevance, preference. For those, the argument that a model can evaluate a model reaches a layer at which somebody decided what counted as the right answer. The scorer is a person, more often than the diagrams suggest, and what that person is paid is the price of the loop. What they are permitted to reject is a different question, and a better one.

The bill the next model gets¶

A corpus cannot be un-generated. Contamination of a substrate is not a defect that can be located and removed later, because separating before from after needs a record that was not being kept at anything like the scale the question needs.

Text gathered before the cheap generators has acquired a scarcity value nobody intended, which is roughly the position of steel smelted before 1945. Low-background steel is prized for instruments because it predates atmospheric nuclear testing, and no amount of demand produces more of it. The supply is whatever was made before the event, minus whatever has since been scrapped.

This cost lands on a party that does not exist yet, which is the cleanest form of deferral available and the one least likely to attract an objection at the meeting.

The bill the institution gets¶

Where a check is delegated, the practice tends to go with it. Reviewing generated code is a skill kept alive by reviewing code. So is reading a proof, judging whether a paper is any good, and noticing that a plausible-looking number is wrong. An organisation that automates its review keeps the responsibility and loses the practice, and generally discovers which of the two it needed at the moment the automation fails.

The same holds a layer up, where institutions come to depend on infrastructure they did not write and have no practice of reading, and a layer above that, where the bodies attempting to supervise these systems find that the instruments capable of doing the supervising are largely made by the parties being supervised.

Nobody is invoiced¶

Each generation is locally sound. The system was trained on available material, evaluated with standard instruments, reviewed by qualified people and shipped. Repeat that eight times and the interesting change is a property of the sequence rather than of any step in it, and a sequence cannot be summoned, fined or asked what it was thinking.

This is why the accounting never happens rather than why it happens badly. Consequences of this kind accumulate beyond the point at which attribution can be assigned to them, and an unattributed cost is not a cost anybody has to carry on a balance sheet.

One route is being tested. Former moderators at the same Nairobi contractor have been in the Kenyan courts since 2022, arguing that the client is answerable alongside the subcontractor for the conditions the work was done in. Some of it has stuck: a court treated the client as the primary employer and ordered medical care for the moderators while the case ran. The question of whether Kenyan courts could hear the claims at all took a first-instance ruling, an appeal and roughly two and a half years, and the rulings due in February 2026 were postponed with no new date. Whether a bill of this kind can be sent back is genuinely open. How long sending it takes is not.

The loop measures its output. Nothing in it measures what that output does to the material the next output will be made from.

The clerk’s brief¶

February 2026: Four years of trying to send the bill back¶

Two sets of claims against Meta and its Nairobi subcontractor have been running in the Kenyan courts since 2022: one brought by Daniel Motaung, a former moderator who says he was recruited to a job described as call centre work and left it with a psychiatric diagnosis, and one brought by more than 180 moderators over dismissals and alleged blacklisting after the moderation hub closed in January 2023. Meta’s position in both was that Kenyan courts had no jurisdiction over an American company. The Employment and Labour Relations Court rejected that in Motaung’s case in February 2023, holding that Meta could be sued in Kenya, and in September 2024 the Court of Appeal dismissed Meta’s appeal in the moderators’ case, finding it proper for that court to intervene. Interim orders have bitten along the way: in June 2023 a court treated Meta as the primary employer of the subcontractor’s moderators and blocked their dismissal, and on 31 July 2023 the High Court ordered medical, psychiatric and psychological care for the petitioners and other moderators while the case ran. On 12 February 2026 the rulings due in both cases were postponed, with no new date given. The clerks record the shape rather than the merits: relief arrives in pieces, jurisdiction alone took two and a half years, and at nearly four years from the first filing there is still no final determination of what the work did to the people who did it.

January 2023: The price of the human check¶

TIME’s investigation into the labelling work behind ChatGPT established the wage, the material and the arrangement: a take-home rate of between around $1.32 and $2 an hour for reading and classifying descriptions of the worst things people do, against a contracted rate of $12.50 an hour paid to the outsourcing firm. Sama cancelled all its work for OpenAI in February 2022, eight months earlier than the contracts provided for. The clerks note that the safety property everyone praises in the finished product was manufactured by these people at this price, and that published accounts of the model’s capabilities do not ordinarily list them as a component.

Left where it fell¶

The costs in this file share one property: each was avoided by a party who was not the party that would carry it. The reading was moved to whoever would do it cheapest, the corpus damage was passed to whatever gets trained next, and the loss of practice was handed to whichever institution still has to answer for an outcome after it stops being able to check one. The clerks’ standing assessment is that none of this requires bad faith at any step, that the absence of bad faith is what makes it durable, and that a cost with no claimant is generally left exactly where it fell.