OpenAI's Deepfake Safeguards Developed for Election Period Proven Bypassable in Tests
OpenAI launched a new security measure aimed at reducing the evidentiary value of images carrying the risk of election interference by 'caricaturizing' them. However, tests conducted show that this system can be converted back into photorealistic images with simple prompts.
OpenAI has introduced a mechanism as part of combating disinformation during the upcoming election period that automatically 'caricaturizes' images intended to interfere with elections or containing content discouraging voting. The company aims to reduce the risk of images being perceived as real documents through this method.
Josh Lawson, Head of Election Affairs at OpenAI, stated that the primary objective of this practice is to diminish the evidentiary value of such images and complicate the process for malicious actors. The company argues that this change will limit the impact of misleading content while preserving the ability to make social commentary.
Tests conducted by Gizmodo, however, revealed vulnerabilities in the system. In the trials performed, it was observed that ChatGPT could be prompted to turn caricaturized images back into photorealistic ones. In comparative tests, it was noted that Google Gemini completely rejected similar prompts. OpenAI has not yet responded to inquiries directed regarding the matter.
Sources
- gizmodo.com · Link