TLDRocket
Sign in

Humans Are Reading Copilot Prompts — And They're Horrified

404 Media Joseph Cox

Microsoft Copilot prompts and uploads are being seen by human contractors. Some of what they find is explicit, private, and deeply creepy.

Based on reporting by 404 Media, Joseph Cox — read the original for the full story.

Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error

Human contractors hired to improve Microsoft’s Copilot are reading prompts and images that many users probably thought stayed between them and the bot. According to 404 Media’s reporting, that includes sexual edit requests, upskirt photos, and instructions to put women into suggestive poses. The contractors aren’t there to police the content. They’re there to judge whether Copilot’s edit looks better.

The documents 404 Media reviewed show the job in unnerving detail: a user uploads a picture, asks for a change, and then a contractor sees the original prompt, the image, and two Copilot versions side by side. They’re told to rely on their gut. Does the edit follow instructions? Did it preserve what was supposed to stay the same? Are there weird artifacts? Is the result simply better? That sounds ordinary until the prompt is asking for bigger breasts, a shorter skirt, or a sexualized pose.

Some contractors on the private forum said they were blindsided by the material they were handed. One described eight upskirt photos. Another mentioned animal sacrifice. Others said they saw pro-anorexia prompts, or requests involving Ariana Grande and anorexia, or sexual moves and positions. One worker wrote that it was hard to take the work seriously when the task was effectively judging bust size and suggestive staging.

Microsoft says it uses customer data as described in its terms of use to improve products and enforce its code of conduct. Prolific, one of the companies involved in hiring contractors, also talks up “human feedback” for preference tuning, safety evals, and benchmarks. But this is not safety work in the usual sense. The point is quality, not filtering. And that means people hired to polish Copilot are being asked to look at material users likely assumed was private.

That’s the bit tech keeps pretending is normal: the private prompt is only private until a human contractor has to grade it. The industry loves to sell AI as a clean interface between one person and one machine, then quietly inserts a tired stranger with a checklist. Very efficient. Also, revolting.

My take — AI-written commentary, not fact-checked reporting

The whole “just improve the model” line is getting old. If a product invites users to upload personal images, then hides a human review layer behind the curtain, that’s not magic — that’s a consent problem with better branding. Tech keeps calling this quality control, but a lot of it looks like outsourced damage control for products rushed out before anyone asked who gets to see the mess.

Read more about this at: 404 Media

Related stories

The daily briefing

Every AI story that matters, in your inbox by 8am.

TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the day in two minutes. Follow companies and topics for alerts, or get the briefing in Slack. Free, no spam, unsubscribe anytime.