Dall-E, the AI-powered text-to-image generator from OpenAI has been adding numerous features of late. Most recently, it introduced the ability to upload images of faces and change the scene surroundings.
Eric Schwratz wrote in Voicebot.ai, “OpenAI had deliberately prevented faces from being included with the AI-powered synthetic media tool out of concern it would be misused, but now says upgraded safety software has mitigated those concerns.”
Trust & Safety
Trust and safety are significant issues for many corners of the AI world. The fear of misuse of the faces of real people in Dall-E output ranged from the risk of embarrassment to blackmail to fraud to potentially framing people for engaging in illegal activity. OpenAI’s solution only allowed for synthetically generated faces without the option to use real images until this week.
This is one of the key concerns about the open source Stable Diffusion. According to one industry insider, users are employing Stable Diffusion to create images that violate laws and are likely creating images with real people’s faces. Because it is open source, there may not be one person behind the text-to-image engine to assign blame to for any particular output. OpenAI is different. Everyone knows exactly who is behind every one of its models and their output.
While many people have been discussing whether the content produced by text-to-image generators should be viewed as original artwork, there are more complex and troubling questions. Some more advanced research is considering whether an image’s content can be segmented into relative weights of other creative works. Then there is trust and safety.
This final concern is the biggest risk to the widespread adoption of the technology today. OpenAI seems to think it has the angles covered. It is clear that there are other solutions with no one even thinking about covering angles for misuse. This is a topic to continue monitoring.