But it’s not just that the models can’t recognize accents, languages, syntax, and faces that aren’t as common in Western countries. “Many of the early deepfake detection tools were trained on high-quality media,” Gregory says. But in many parts of the world, including Africa, cheap Chinese smartphone brands with stripped-down features dominate the market. The photos and videos these smartphones can produce are of much lower quality, further confounding the detection models, Ngamita says.
Gregory said some models are very sensitive and that even background noise in audio or compressed video for social media can lead to false positives and negatives. “But that’s exactly the situation you’ll encounter in the real world, it’s brute force detection,” he said. The free public tools available to most journalists, fact checkers, and members of civil society are also “highly inaccurate in terms of addressing both the inequity of who’s reflected in the training data and the challenge of working with this low-quality material.”
Generative AI is not the only way to create manipulated media: so-called cheap fakes – media manipulated by adding misleading labels or simply slowing down or editing audio or video – are also very common in the Global South, but can sometimes be falsely flagged as AI-manipulated by flawed models or untrained researchers.
Diya worries that groups using tools that are more likely to flag content outside the U.S. and Europe as AI-generated could have serious implications at the policy level, prompting lawmakers to crack down on fictitious issues. “There’s a big risk in terms of inflating these numbers,” she says. And developing new tools isn’t just a matter of pressing a button.
Like any other form of AI, building, testing and running detection models requires access to energy and data centers that aren’t available in most parts of the world. “When we talk about AI and local solutions here, without the compute side, it’s almost impossible to even run the models we’re trying to come up with,” says Ngamita, who is based in Ghana. Without local alternatives, researchers like Ngamita have no choice but to pay for access to off-the-shelf tools like those offered by Reality Defender (which can be cost-prohibitive), use inaccurate free tools, or try to access them through academic institutions.
For now, Ngamita said, the team has had to partner with universities in Europe where they can send content for verification. Ngamita’s team has compiled a dataset of possible deepfake cases from across the continent, which he says will be valuable for academics and researchers looking to diversify their model datasets.
But sending data to others also has drawbacks: “The time lag is pretty big,” Diya says. “It takes at least a few weeks before someone can confidently say this is AI-generated, and by that time the content is already damaged.”
Gregory said Witness, which runs its own rapid response detection programme, was receiving a “huge number” of cases. “It’s already challenging for frontline journalists to process the volume of cases they’re starting to face within the timeframes they need,” he said.
But Diya points out that too much focus on detection could divert funding and support away from organizations and institutions that are building a more resilient information ecosystem overall. Funds, she says, should be directed toward news organizations and private groups that can generate public trust. “I don’t think the money is being spent there,” she says. “I think the money is being spent more on detection.”
