The reading the product does not show
Documents seen by 404 Media show OpenAI hiring hundreds of contractors who read a large stream of real ChatGPT prompts, at times whole conversations. The teams score and critique replies. The documents show that work being used to stop the model presenting itself as human and to cut sycophancy.[1]
The contractors do not see usernames. OpenAI said it tries to strip personal information before review and acknowledged that sensitive details can still pass. Anthropic confirmed human review of Claude replies for users who have the improve-our-models setting on, with account details stripped. Someone who works with the prompts said no when asked whether users know that humans are reading them.[1]
Where the gain is booked
The setup is a real production step: a scored reply becomes input to the next model. The same step is a privacy risk for the user and repetitive reading work for the contractor. The improvement is booked to OpenAI's product; the reading hours stay on an outside contract. The user does not see that share as a shift. It appears as a line in a privacy notice.[1]
In a July column I wrote that a compute pledge left maintenance, review and false-alarm work off the ledger. Here too, the impression that the model improves by itself hides a paid reading layer. The counterpoint stands: the reading can be a real check against sycophancy and a false persona. That does not change the allocation. It only shows who the work is booked to.[1], [2]
The one line that would make it visible
OpenAI and Anthropic accept that the review exists. Headcount, pay, shift length and whether the user actually consented are not in the documents. A published headcount, contract type or consent wording in the next quarter would make it measurable whether the reading is still parked outside. If none appears, the shift stays behind the product.[1]