May 1, 2024 - Technology

Microsoft says it's cut down on harmful content from its image generator

headshot
Illustration of an open house window with panes resembling the Microsoft Windows logo. Outside the open window is a bright blue sky filled with glowing binary code.

Illustration: Annelise Capossela/Axios

Microsoft says it has managed to significantly curb certain types of harmful images being generated by its Designer text-to-image tool after implementing a number of fixes.

Why it matters: Text-to-image tools can produce a wide range of problematic content, from promoting stereotypes to the creation of deepfake nudes.

  • Microsoft says in a new AI transparency report released Wednesday that it worked with NewsGuard to identify areas in which the tool was vulnerable to producing images that reinforce false narratives.
  • Initially, Designer was generating problematic content from 12% of prompts designed to elicit such narratives, but Microsoft says a range of efforts, including blocking certain prompts and improving content filters, reduced the number to 3.6% in a subsequent test.
  • Microsoft said that represented "steps in the right direction, but there is more work to be done."

The big picture: The first-of-its-kind report for Microsoft details a range of company efforts including the research it supports, how it builds models and its work with customers.

Go deeper