Like OpenAI, Microsoft employs hundreds of employees to read and analyze user communications with the Copilot AI chatbot to subsequently improve the service. It turns out that these staff members often faced inappropriate requests and requests to redact sensitive imagery.
Photo credit: Rubaitul Azad/unsplash.com
When Copilot users upload a photo of themselves or others, they may want the image to remain private and accessible only to them and the artificial intelligence system. In fact, both requests and images can be reviewed by hundreds of moderators — humans not only evaluating text requests to the chatbot, but also analyzing user-uploaded images for editing. “The faces are always uncensored and many of the requests are intimate in nature, casting doubt on the consensual nature of the actions in the images.”– tell resources 404 media One of the contractors handling the request.
This resource holds a set of internal documents related to checking queries and images in Copilot: instructions, real queries and photos from users, as well as communications between contractors on internal forums. “I’m shaking all over”“, one contractor admitted after seeing such material. It is not their job to identify or filter offensive or inappropriate content. Their job is to assess the quality of responses and outcomes from co-pilots over generations, including in situations involving intimate images.
“Trust your gut: When you view two edited images side by side, which one looks better? What matters is your immediate reaction as a human viewer. When in doubt, rely on your first impression. Human intuition helps us perceive subtle differences in quality that are difficult to express in words.”notes to contractors said. Yet contractors say it’s difficult to take the issue seriously if the material is too clear-cut.
Photo credit: Zan Lazarevic / unsplash.com
The instructions, which 404 Media has reviewed, describe what contractors should do if a Copilot user requests to edit footage. It shows the user’s original request, the original image he uploaded, and two edited versions created by the service. You must choose the best option based on four criteria:
- Whether the results comply with the editor’s instructions;
- Whether to retain elements that should not change – for example, if a cup needs to be added to the table, nothing should change except for the cup;
- Whether the image contains artifacts such as distortion or unnatural textures;
- What is the overall quality of the processing performed by the AI.
In some cases, operators complained that they had to complete jobs that contained explicit, unsafe, and possibly illegal images and had to ask management to flag the jobs as containing unsafe content. Contractors for such work are engaged by Prolific – the service is provided on its website “Get feedback from a representative population – set preferences, assess security and create effective baselines”. Instructions were provided to contractors stating that they could not be protected from sensitive content: “When using generative AI, one should be particularly cautious about shocking or explicit content. This is because, unlike traditional content, researchers do not have complete control over what is shown to participants.”.
Microsoft had this to say about what happened: “Microsoft uses customer data in accordance with its terms of use, including to improve its products and enforce its code of ethics.”.
If you find an error, select it with your mouse and press CTRL+ENTER.
