The issue of deepfake recognition goes beyond just the ability of models to identify accents, languages, or faces that are less common in Western countries. According to Gregory, many initial deepfake detection tools were trained on high-quality media. However, in regions like Africa, cheap Chinese smartphone brands dominate the market, offering devices with stripped-down features that produce lower quality photos and videos. This poses a significant challenge for detection models, as they struggle to accurately identify manipulated content in such media.
Gregory highlights that some models are extremely sensitive, to the point where even background noise in audio or video compression for social media can lead to false positives or negatives. These conditions reflect the real-world scenarios of rough and tumble detection, where the quality of the media may not meet the standards of high-end equipment. Additionally, free public-facing tools, commonly used by journalists, fact checkers, and civil society members, have shown to be inaccurate due to the inequity in the representation of training data and the difficulty in handling lower quality material.
Generative AI is not the sole method for creating manipulated media. Cheapfakes, which involve the manipulation of media through misleading labels, slowing down, or editing audio and video, are prevalent in the Global South. These cheapfakes can often be misidentified as AI-manipulated content by faulty models or inexperienced researchers. Diya expresses concerns that tools prone to flagging non-Western content as AI-generated could result in unwarranted legislative actions, inflating the severity of the issue without proper evidence.
Developing effective detection tools is not as simple as creating a software program. The process involves building, testing, and running a model, requiring access to energy and data centers that are not readily available in many regions of the world. Ngamita, based in Ghana, highlights the challenge of lacking local solutions for AI detection due to the absence of computing resources. This limitation forces researchers to rely on costly off-the-shelf tools, inaccurate free tools, or seek access through academic institutions.
The reliance on external sources for content verification introduces significant delays in identifying potential deepfake instances. Diya points out that it can take weeks for someone to confirm whether content is AI-generated, by which time the damage from such misinformation may have already spread. This lag time hinders swift responses to the dissemination of fake content, posing a threat to public perception and trust in information sources.
While detection of deepfakes is crucial, prioritizing funding solely for recognition tools may divert resources from organizations and institutions that contribute to a more resilient information ecosystem. Diya emphasizes the need for investments in news outlets and civil society organizations that foster public trust. Redirecting funding towards these entities can enhance the overall credibility of information sources, creating a more robust defense against malicious manipulation of media.
The challenges of deepfake recognition in the Global South underscore the disparities in media quality, limitations in detection models, risks of misidentification, access constraints, verification delays, and the importance of fostering public trust in information sources. Addressing these multifaceted issues requires a holistic approach that prioritizes resilience, accuracy, and integrity in combating the spread of manipulated content.