"Human contractors hired to improve Microsoft’s Copilot AI chatbot are constantly bombarded with lewd or sexually explicit photo editing requests and images that users have uploaded, including upskirt photos or putting women into sexual positions. These contractors are then asked to review whether the generated image successfully fulfilled the prompt — such as, did the image generator make the woman’s AI-enlarged breasts big enough.

When a Copilot user uploads a picture of themselves or someone else, they probably expect that image to remain private between just them and the AI tool. In reality, a workforce of at least hundreds of human reviewers are sometimes looking at that prompt and whatever images they upload. And in many cases, those contractors are inundated with requests to make foot fetish images of children’s cartoon characters, shorten a real woman’s skirt, or put people into sexual positions.

The news, based on a cache of internal contractor documents seen by 404 Media, shows that AI companies are using human workers to review not just AI chatbot users’ text prompts, but the pictures they upload and wish to edit too."

404media.co/humans-reading-cop…

#AI #GenerativeAI #Microsoft #Copilot #ContentModeration

"In June 2025, Jesse Van Rootselaar received an email notification from OpenAI. Van Rootselaar had been using ChatGPT to discuss committing potential acts of violence—including a mass shooting at a shopping mall—and now OpenAI had banned that account. But the troubled teen living in Tumbler Ridge, British Columbia, was not deterred and soon logged back on with a second account.

What happened next was astonishing: Van Rootselaar told ChatGPT about the account ban, noting she had specified “a famous real-world mall as the target” and herself as the perpetrator, but also claiming the scenario hadn’t been real. In response, ChatGPT explained why such violent content would be flagged—and then offered tips on how to evade its own safeguards. It advised Van Rootselaar to always frame her content as fictional or hypothetical in order to “never get flagged again” by OpenAI’s moderation systems.

“Don’t use real-world locations,” ChatGPT said, later adding: “You can still be twisted. Just be clever about it.”

Over the next eight months, Van Rootselaar, who identified as transgender and used she/her pronouns, continued to use this second ChatGPT account to focus on violence. She discussed building and using firearms and homemade explosives, fantasized about becoming a notorious mass killer, and asked about the timing of previous real-world school shootings just before committing her attack, according to three separate sources with knowledge of the matter and material I reviewed from Van Rootselaar’s ChatGPT history."

motherjones.com/media/2026/09/…

#AI #GenerativeAI #LLMs #OpenAI #ContentModeration #ChatGPT #Canada

"OpenAI is hiring hundreds of contractors who read a massive stream of real users’ ChatGPT prompts, with the prompts sometimes including sensitive personal information, 404 Media has learned. The prompts these people review can include whole conversations between users and the chatbot, conversations that most of ChatGPT’s more than 900 million users probably don’t realize may be read by actual people.

The goal of these prompt review teams is to improve the responses ChatGPT gives to its users, with the contractors rating and critiquing the chatbot’s generated replies. Internal documents seen by 404 Media show contractors training ChatGPT to not anthropomorphize itself, and to be less sycophantic, a key problem for OpenAI whose over-sycophantic 4o model led in part to multiple peoples’ suicides, according to various lawsuits.

The news presents a major privacy risk for ChatGPT’s users, with people often using ChatGPT as a therapist, professional assistant, or digital friend, and providing it with all sorts of intimate details about their lives. The contractors don’t see ChatGPT usernames, and OpenAI says it tries to remove personal information before prompts reach the reviewers, but the company acknowledged sensitive details can still get through.

The news also dispels the misconception that these models are improving only because of OpenAI’s mass scraping of the internet, the talent of its well-paid engineering and AI teams, or the power of its newer models. An important and overlooked part are the outside contractors paid to read and review ChatGPT responses to real prompts over and over again. Anthropic confirmed to 404 Media it is also using human review to improve its models."

404media.co/inside-project-lil…

#AI #GenerativeAI #OpenAI #Privacy #DataProtection #ContentModeration

"As the worlds of social media and artificial intelligence collide, Meta’s Oversight Board is rethinking how it can best execute its regulatory role.

Meta’s Oversight Board is an independent, external regulatory body established in 2018 to review content moderation decisions across Facebook, Instagram, and Threads. The 21-member global team acts like a “supreme court” making binding decisions as well as suggestive recommendations when a user appeals Meta’s original decision. Its monthslong, case-based approach alone may not cut it in the age of AI, Sudhir Krishnaswamy, a legal scholar and academic, and the only Indian on the global board, told Rest of World.

“We are always trying to reduce time taken and trying to get more done. Maybe because of the gen-AI space, some of our work would be less individual case-based and more structured. That’s a possibility. And we’re very open to that,” Krishnaswamy said.

Already, seven in 10 social media images are AI-generated using tools like Midjourney or DALL-E, and eight in 10 content recommendations rely on AI. Nearly half of all social media content by businesses will be AI-generated in 2026. Meanwhile, AI tools are leaving behind non-English speakers."

restofworld.org/2026/mmeta-ove…

#SocialMedia #Meta #Facebook #ContentModeration #MetaOversightBoard

"Two independent analyses of social media content in the lead-up to the German federal election in 2025 have shown that extremist parties, in particular the right-wing Alternative for Germany (AfD), were disproportionately favored by X, TikTok, YouTube, and Instagram. A report prepared by the German nonprofit organization Bertelsmann Stiftung found that on TikTok, for example, 50% of all suggested political content was found to be AfD-related, with the mainstream conservatives a distant second at 15%. The outsized prominence of extremist content cannot be explained by the parties’ actions alone, because they all used very similar social media strategies. Another study, which has been shared by the authors via the pre-publication platform Arxiv, showed that the X algorithm disproportionately amplified content by extreme parties, especially on the extreme right. This selective amplification is particularly concerning in light of earlier research conducted by me and my team, which showed that German politicians from the extreme right and left share far more untrustworthy content on Twitter than politicians of the four mainstream parties.

A recent field experiment investigated the consequences of algorithmic amplification by re-ranking content favored by the X/Twitter algorithm that expressed antidemocratic attitudes and partisan animosity. When antidemocratic content was downranked, participants’ outgroup animosity declined compared to a control group that was exposed to the standard X/Twitter algorithm, both during the study and afterwards. Reduced exposure to antidemocratic content also reduced people’s negative emotions during the study. This is not an isolated finding but adds to existing evidence that social media causally contributes to hate crimes and xenophobia.

The DSA was designed to address such challenges."

science.org/doi/full/10.1126/s…

#SocialMedia #SocialNetworks #ContentModeration #Algorithms #AlgorithmicRecommendation #EU #DSA