
What happened
Internal documents obtained by 404 Media say at least several hundred contract staff review Copilot image edits, comparing the user's prompt and uploaded image against two AI-generated results.
Why it matters
Uploaded images and files are treated like Copilot conversations, so the images users hand over for editing appear to reach human reviewers, not just automated systems.
What to watch
Microsoft told 404 Media it uses customer data as stated in its terms of service, so the test is how far the published privacy FAQ is read to cover uploaded images.
WHO IT HITSThis lands hardest on people who upload personal or sensitive photos to Copilot for editing, and on the contract reviewers who screen that material; enterprises weighing Copilot for such work may also want to check the terms.
Summaries like this, in your inbox every morning.
Copilot's image-editing feature already asks a lot of automated quality checks. Users upload a photo and ask for changes like removing a person in the background or altering clothing. Judging whether the edit followed the instruction, left untouched areas alone, and avoided distortion is a fine-grained call, and Microsoft has brought human judgment into that evaluation.
According to 404 Media, the setup puts original prompts, original images, and two edited outputs side by side in front of reviewers. Some of those images arrived without blurring on faces, and many prompts requested sexual edits, including shortening a skirt or shifting a person into a sexual pose — meaning reviewers encountered sensitive material in the course of the job.
Microsoft has not hidden that humans may look at Copilot conversations. Its privacy FAQ says some conversations are subject to automated processing and human review for product improvement and safety, and that trained AI experts may review them; uploaded images and files are described as handled the same way. That published framing is likely to be the reference point as this story develops, since the dispute looks less about whether human review happens and more about how far users understand it to extend.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
Circlia started a proof-of-concept for TASTE SIGNAL: visitors to Holland Village Cafe FUKUOKA taste cookies an…

In a webinar hosted by French Mathematical Society president Isabelle Gallagher, Fields Medal winner Hugo Dumi…

NVIDIA released Nemotron 3.5 Super VL, an open-weight model with 120 billion parameters that handles text, ima…

NVIDIA released Nemotron 3.5 Bolt 74B-A9B 24GB, an AI model for local agents that hits over 80 tokens per seco…

Anthropic said on October 15 it has started giving METR's six staff company devices, office space and access t…

NVIDIA published a blog post titled "AI Security Is an Engineering Problem - How to Solve It at Every Layer of…
