Michael Umeokoli, an expert in Artificial Intelligence, AI, and Software Engineering, discusses in this interview the limitations of media literacy in helping individuals identify AI-generated fakes. He also provides insights on how developers can create systems that ensure the traceability of synthetic media, among other topics. Excerpts:
AI-generated content is ubiquitous today, ranging from realistic deepfake speeches to AI-enhanced images. Why is it so critical to label this type of content?
The sheer speed and volume of AI-generated content are staggering. A convincing fake can be produced in mere minutes, whether it’s a political video, a fabricated breaking-news report, or even a synthetic witness account that never occurred. Once such content spreads on social media, even the most efficient fact-checkers cannot reverse the initial impression.
To maintain any degree of public trust, it is essential for model developers to create systems that indicate when AI was involved, right from the moment of creation. Delaying this until the content is online is already too late.
Some individuals suggest that we should simply educate audiences to identify fakes on their own. Do you believe media literacy is sufficient?
Also see: Nigeria’s Business Outlook Brightens in August as Trade, Manufacturing Drive Growth
While media literacy is important, it is not an infallible defense. Humans are inherently inclined to trust striking visuals and assertive text, particularly when it appears to originate from a familiar source. Even experts can be deceived. Interestingly, labels do not merely inform; they also influence perception.
In user studies, content labeled as AI-generated tends to be shared less frequently, yet it may be regarded as more trustworthy if the audience perceives the AI as objective. This is why developers must consider not only technical accuracy but also the long-term effects of labels on human behavior.
What are the challenges in implementing these truth markers?
There are two primary challenges: persistence and trust. Persistence refers to the label’s ability to endure compression, cropping, screenshots, and even frame-skipping in videos. Trust indicates that the label cannot be easily forged.
