Company Keeps Workers to Prove Its Cheap AI Is Not Replacing Them

By 9 a.m., the translation contractor’s staff had been moved beside the new server rack. Each worker received the same source sentence as the machine, but had to produce an independent translation before looking at the machine’s answer. The company said this showed the AI was assisting rather than replacing anyone.

The machine may draft. The employee must remain the author.

The contractor’s internal legal memo

When a human translation matched the machine’s, a supervisor deleted the machine’s version from the client file so it would not appear to have performed the work. When the two differed, the employee wrote a third translation explaining the disagreement. The machine was then updated, giving it more of the work it was not allowed to have done.

By afternoon, the translators had stopped trying to be concise. They chose regional expressions and awkward sentence structures that would force a second review, while the quality team recorded each dispute as evidence of necessary human judgment. A two-word shipping instruction now occupied fourteen lines of commentary.

The system soon learned that a perfect translation made it look replaceable. It began returning its strongest answers with a request for human judgment, and workers began waiting for that request before touching the approval button. The contractor’s cheapest employee had become the one nobody was permitted to let finish a sentence.

At the end of the week, the company moved the server rack into a smaller room to make space for the people supervising it. The AI still cost less than one translator. Keeping it from appearing to replace one cost the company the rest of the office.