The issue of deepfake detection goes beyond just the ability of models to recognize accents, languages, syntax, or faces that are less common in Western countries. The initial deepfake detection tools were trained on high quality media, which poses a challenge when considering the prevalence of cheap Chinese smartphone brands in regions like Africa. These phones produce lower quality photos and videos, creating confusion for detection models. Background noise in audio or compressing videos for social media can also lead to false positives or negatives, highlighting the discrepancies between models trained on high quality media and the reality of media production in many parts of the world.
Generative AI is not the only method used to create manipulated media. Cheapfakes, which involve manipulating media by adding misleading labels or modifying audio and video, are common in the Global South. However, faulty models or untrained researchers may mistakenly flag these cheapfakes as AI-manipulated content. This could have serious repercussions on a policy level, potentially leading legislators to address non-existent issues. The risk of inflating numbers and misidentifying content underscores the importance of developing accurate and region-specific detection tools.
Building, testing, and running detection models require access to energy and data centers, which are often lacking in many parts of the world. Without local alternatives, researchers face challenges in accessing reliable detection tools. Options include paying for expensive off-the-shelf tools, using inaccurate free tools, or seeking access through academic institutions. The lack of local compute resources makes it nearly impossible to develop and deploy detection models independently, forcing researchers to rely on partnerships with institutions in other regions.
Sending data to external entities for verification introduces significant lag time, with delays of several weeks before confirming whether content is AI-generated. This lag can be detrimental, as the damage from manipulated content is often already done by the time verification is complete. Organizations like Witness receive a high volume of cases and struggle to respond within the timeframes required by frontline journalists. While detection efforts are crucial, there is a risk that excessive focus on detection could divert funding and support away from institutions that promote a more resilient information ecosystem.
Instead of solely investing in detection technologies, funding should also be allocated to news outlets and civil society organizations that foster public trust and promote media literacy. By strengthening institutions that uphold journalistic integrity and combat misinformation, the overall information ecosystem can become more resilient to the threat of deepfakes. Redirecting resources towards organizations that prioritize transparency and accuracy can help mitigate the impact of manipulated media on society.
Addressing the challenges of deepfake detection requires a global perspective that considers the quality of media, access to resources, response times, and the broader information ecosystem. By recognizing the limitations of current detection models and advocating for greater support for reliable institutions, we can collectively work towards building a more resilient defense against the proliferation of manipulated media.
Leave a Reply