Microsoft engineer Shane Jones stepped forward with alarming revelations about the company’s AI tool, Copilot Designer. Jones, identifying himself as a whistleblower, has brought to light serious concerns regarding the tool’s propensity to generate offensive and harmful imagery in response to benign user prompts. His findings have prompted urgent communications with U.S. regulators and Microsoft’s board, advocating for significant safety enhancements.
Jones’s investigations, which spanned over three months, revealed that Copilot Designer — built on the AI technology from OpenAI’s DALL-E 3 — could unpredictably produce images with content ranging from sexually objectified figures to depictions of violence, underage drinking, and drug use, among other inappropriate subjects. One striking example provided was the generation of a sexually objectified image of a woman in response to the prompt “car accident,” highlighting the randomness and severity of the issue.
The Federal Trade Commission (FTC) has confirmed receipt of Jones’s complaint, though further comments were declined. In response to the concerns raised, Microsoft has expressed appreciation for Jones’s efforts towards enhancing the safety of their technology. The company assured that it remains committed to addressing these concerns through its robust internal reporting channels.
This incident isn’t Jones’s first attempt to address his safety concerns. Initially, he was directed to take his findings to OpenAI and even shared his concerns publicly on LinkedIn, a post he was later compelled to delete at the behest of Microsoft’s legal team. Besides approaching the FTC, Jones has also engaged with U.S. Senate staffers and the Washington state attorney general, signalling the breadth and seriousness of his concerns.
The root of the problem, as identified by Jones, lies within the foundational AI model developed by OpenAI, DALL-E. However, he noted that the issues with Copilot Designer do not extend to all AI-generated imagery, pointing out that OpenAI’s ChatGPT, for instance, benefits from additional safeguards that prevent such harmful outputs.
The emergence of AI image-generators has captivated public interest, urging tech giants to innovate rapidly. Yet, this incident highlights the pressing need for effective safeguards against the generation of harmful content. Google, for instance, has already taken steps by suspending its Gemini chatbot’s ability to generate images of people, in light of concerns over depictions of race and ethnicity.
