But it’s not just that the models can’t recognize accents, languages, syntax, or faces that are less common in Western countries. “A lot of the early deepfake detection tools were trained on high-quality media,” Gregory says. But in most of the world, including Africa, low-cost Chinese smartphone brands that offer simplified features dominate the market. The photos and videos these phones can produce are of much lower quality, further confusing the detection models, Ngamita says.
Gregory says that some models are so sensitive that even background noise in audio or video compression for social media can result in false positives or negatives. “But these are exactly the kinds of circumstances you encounter in the real world, detecting bumps and dips,” he says. The free, publicly available tools that most journalists, fact-checkers, and members of civil society probably have access to are also “the ones that are extremely inaccurate in terms of dealing with both the inequality of who is represented in the training data and the challenges of dealing with that lower-quality material.”
Generative AI is not the only way to create manipulated media. So-called cheapfakes, media manipulated by adding misleading labels or simply slowing down or editing audio and video, are also very common in the Global South, but can be mislabeled as AI-manipulated by flawed models or untrained researchers.
Diya worries that groups using tools that are more likely to flag content from outside the U.S. and Europe because it’s generated by AI could have sedate political repercussions, prompting lawmakers to crack down on imaginary problems. “There’s a huge risk of inflating these numbers,” she says. And developing fresh tools isn’t a matter of clicking a button.
Like any other form of AI, building, testing, and running a detection model requires access to power and data centers that simply aren’t available in most of the world. “If we’re talking about AI and local solutions, without the computational side of things, it’s almost impossible for us to run any of our models that we’re going to build,” says Ngamita, who lives in Ghana. Without local alternatives, researchers like Ngamita have few options: pay for access to an off-the-shelf tool like Reality Defender, which can be prohibitively exorbitant; utilize incorrect free tools; or try to get access through an academic institution.
For now, Ngamita says his team has had to partner with a European university where they can send snippets of content for review. Ngamita’s team is compiling a dataset of possible deepfake cases from across the continent, which he says is valuable to scientists and researchers trying to diversify their models’ datasets.
But sending data to someone else has its drawbacks. “The lag time is quite significant,” Diya says. “It takes at least a few weeks before someone can confidently say this is AI-generated, and by that time, the content, the damage has already been done.”
Gregory says Witness, which runs its own rapid response detection program, is receiving a “huge number” of cases. “It’s already hard to deal with them in the time that frontline journalists need and with the volumes they’re starting to see,” he says.
But Diya says that such a weighty focus on detection can draw funding and support away from organizations and institutions that generally create a more resilient information ecosystem. Instead, she says, funding needs to go to news outlets and civil society organizations that can build a sense of public trust. “I don’t think that’s where the money is going,” she says. “I think it’s more about detection.”