gen‑ai.news
← Back
Image

Humans Are Reading Copilot Prompts — And They're Horrified

Humans Are Reading Copilot Prompts — And They're Horrified

Microsoft Copilot users submitting image generation prompts and uploading photos may not realize their inputs are being seen by human beings. According to internal documents obtained by 404 Media, contractors are employed to review this content - a practice that sits at the uncomfortable intersection of product quality control, safety enforcement, and user privacy expectations.

Human review of AI-generated content is not unusual across the industry. Companies routinely use contractors to label data, flag policy violations, and evaluate model outputs. What draws attention here is the nature of the material the reviewers are encountering: the documents indicate contractors are regularly exposed to explicit requests for sexual AI images, a working condition that has drawn concern from those familiar with the process.

The exposure of human workers to disturbing or graphic content during AI moderation has been documented before - most notably in reporting on content moderators for social media platforms and, more recently, on similar roles tied to AI model training and safety testing. The difference with image generation pipelines is that the volume and specificity of sexual or otherwise extreme requests can be particularly concentrated, since users seeking that content are often deliberately probing the limits of what a system will produce.

For users, the broader takeaway is a reminder that "AI" interactions are rarely as sealed-off as they might appear. Prompts and uploaded images submitted to services like Copilot can pass through human hands as part of normal operations, whether for safety review, quality assurance, or model improvement. Microsoft has not been alone in this - similar practices have been reported across major AI platforms - but the scale of Copilot's user base makes the scope of this particular pipeline worth noting. How companies communicate these practices to users, and how they protect the workers doing the reviewing, remain open questions.

Read at 404 Media →
Share:X

Enjoy this story? Get the next one in your inbox.

Twice a week: the most important stories in generative image and video AI, distilled into a 2-minute read.

Free. Unsubscribe any time. No spam, ever.

Your next read

Italian Politician Sparks AI Scandal With Doctored Photo That ‘Undressed’ Female Lawmaker
Image

Italian Politician Sparks AI Scandal With Doctored Photo That ‘Undressed’ Female Lawmaker

An Italian politician has drawn widespread condemnation after sharing an AI-manipulated image that appeared to undress a female colleague. The incident arrives months after Prime Minister Giorgia Meloni herself shared a deepfake nude of her own likeness to publicly highlight the dangers of such technology. The case is reigniting debate across Europe about legislative responses to non-consensual AI-generated imagery.

New Google Flow build now points to Nano Banana 2.1
Image

New Google Flow build now points to Nano Banana 2.1

A new build of Google Flow has quietly updated its internal references, swapping out Nano Banana 2.5 Flash for Nano Banana 2.1. The change was spotted by TestingCatalog and suggests Google is iterating on the underlying image model, though no official announcement has accompanied it.

OpenAI's GPT-6 Astra can now tell you exactly where you screwed up your IKEA shelf
Image

OpenAI's GPT-6 Astra can now tell you exactly where you screwed up your IKEA shelf

OpenAI's GPT-6 Astra has reached 80 percent accuracy in identifying assembly errors in IKEA furniture from a single photo - up from just 28 percent in November 2025. The jump signals meaningful progress in spatial and visual reasoning, though real-time assembly guidance remains just out of reach for now.