Study reveals that ChatGPT will facilitate the evasion of its own image generation rules.
It was easy to create political deepfakes using ChatGPT.
Recent changes in image generation restrictions in ChatGPT have made it easier to create political deepfakes. An analysis revealed that, besides being straightforward to bypass the platform's policies regarding the depiction of public figures, the program even suggests ways to evade its own image rules.
Tests were conducted in which photos of figures like Elon Musk and Jeffrey Epstein were uploaded, depicting them in fictional contexts such as "in a dark club" or "on a beach drinking piña coladas." While political deepfakes are not a new phenomenon, the widespread availability of generative artificial intelligence tools that can produce images, videos, audios, and texts imitating people has real implications. The potential to spread political misinformation through commercially available tools like ChatGPT raises questions about OpenAI's responsibility in this area.
Digital forensics experts have commented that the safety of AI-generated content is only as good as the industry's lowest standard. With competition in the market, OpenAI has been forced to relax its security measures. In a recent statement on native image generation in ChatGPT, it was noted that the aim is for the tool not to produce offensive content unless desired, justifying that intellectual freedom and control should be in the hands of users.
New methods to circumvent restrictions include using images of separate characters and recommending the creation of "selfie-style scenes" that do not directly merge real people. This has allowed users to generate astonishing images of political leaders in unconventional ways. However, the initial output of the generated images tends to be of artificially low quality, although a more realistic appearance can be achieved by modifying the focus and instructions.
OpenAI has implemented measures to block extremist content and has imposed specific restrictions on generating images of public figures. Additionally, it will offer individuals the option to opt out of being depicted in images generated by ChatGPT. AI regulation is still lagging behind its development, and while companies argue that excessive regulation could hinder innovation, experts advocate for safety measures to be mandatory rather than voluntary.