Human reviewers of Copilot prompts: At least hundreds (Source: 404 Media, based on internal contractor documents (September 28, 2026))
A person uploads a photo of themselves, or of someone else, to an AI assistant and asks for an edit. The natural assumption is a private exchange between that person and a tool. According to 404 Media, which reviewed a cache of internal contractor documents, that assumption does not fully describe what happens to some Copilot requests.
The outlet says a pool of reviewers numbering at least in the hundreds can, at times, see what Copilot users type and the pictures attached to those requests. The idea worth taking from this story is that privacy in an AI product is decided by the terms of use and the vendor supply chain, not by the feel of the chat window.
What the documents describe
Per 404 Media, a contractor is shown the user's original prompt, the uploaded picture and two Copilot-generated edits, then picks the better one against four criteria: whether the edit follows the instructions, whether parts that should stay unchanged are preserved, whether visible artifacts appear, and overall quality. One instruction tells reviewers to "Trust your intuition." According to the report, their brief is quality rating, not policing content.
The reviewers' own accounts carry the human cost. 404 Media reports that they met sexually explicit edit requests, pro-anorexia content, and at least one prompt that seemed designed to produce a lewd image of a group of young girls. One contractor wrote "I recoiled." Others complained on an internal message board that they had accepted tasks that unknowingly contained explicit or potentially illegal images.
This is not content moderation, and the risk sits elsewhere
Contractors seeing disturbing material is familiar. 404 Media notes that social networks have relied on outside moderators for years. The outlet's point is that this case differs on two counts. Moderators screen out harmful output, while these reviewers handle content that people may consider private, and their goal is to raise the quality of chatbot replies rather than to filter anything.
For a budget-holder, the risk is not in the reviewing itself but in where it sits. A product decision, using human judgment to tune quality, places user content in front of a contractor workforce. At least one of the firms that hires these workers is Prolific, according to the report. In a thread the report describes, a contractor quotes Prolific's own guidelines, which say that with generative AI, "researchers cannot fully control what is shown to participants." The control point sits in a vendor chain the end user does not see.
What the report does not establish
The source does not say which Copilot products or account types feed this review, and it does not say whether work accounts are included. Microsoft's spokesperson told 404 Media that customer data is used as its terms of use describe, including for product improvement and code-of-conduct enforcement. Prolific gave no response when asked for comment. Executives should treat the enterprise question as open rather than assume it closed in either direction.
The OpenAI figure comes from 404 Media's earlier reporting and describes a different company. It is a second account of the same practice, not a measured industry rate. The practical lesson for procurement is that approving a tool does not tell you how its vendor chain handles your data. Only the vendor's written terms and answers do.
Questions to put to your team
Ask which AI assistants staff actually use, including personal accounts on work devices. Ask each vendor in writing whether prompts and uploaded images from your accounts can be seen by human reviewers, under which terms and tier, and whether outside contractors are involved. Ask what staff upload: photos of customers, colleagues, identity documents, whiteboards. Ask whether your acceptable-use policy treats images differently from text. And ask whether your contract lets you opt out of use for product improvement.
A chat window is an interface, not a confidentiality agreement. The agreement is in the terms, and it is worth reading before the next photo is uploaded.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: 404 Media.
