AI image generators have the potential to spur creativity and revolutionize content material creation for the higher. Nevertheless, when misused, they will trigger actual hurt by means of the unfold of misinformation and reputational harm. Microsoft is hoping to forestall additional misuse of its generative AI instruments by means of implementing new protections.
Final week, AI-generated deepfakes sexualizing Taylor Swift went viral on Twitter. The photographs had been reportedly shared through 4chan and a Telegram channel the place customers share AI-generated photographs of celebrities created with Microsoft Designer.
Microsoft Designer is Microsoft’s graphic design app that features Image Creator, the corporate’s AI picture generator that leverages DALLE-3 to generate sensible photographs. The generator had guardrails that prevented inappropriate prompts that explicitly talked about nudity or public figures.
Nevertheless, customers discovered loopholes resembling misspelling celeb names and describing photographs that did not explicitly use sexual phrases however generated the identical end result, in accordance with the report.
Microsoft has now addressed these loopholes, making it not possible to generate photographs of celebrities. I attempted getting into the immediate, “Selena Gomez enjoying golf,” on Picture Creator and obtained an alert that mentioned my immediate was blocked. I additionally tried misspelling her title and obtained the identical alert.
“We’re dedicated to offering a secure and respectful expertise for everybody,” mentioned a Microsoft spokesperson to ZDNET. “We’re persevering with to research these photographs and have strengthened our present security methods to additional forestall our companies from being misused to assist generate photographs like them.”
As well as, the Microsoft Designer Code of Conduct explicitly prohibits the creation of grownup or non-consensual intimate content material, and a violation of that coverage can lead to shedding entry to the service totally.
Some customers have already expressed curiosity to find a workaround to those new protections within the Telegram channel, in accordance with the report. So, that is prone to be a cat and mouse recreation between dangerous actors discovering and exploiting loopholes in generative AI instruments and the businesses behind these instruments dashing to repair them for a very long time.