AI deepfake detection systems show reduced accuracy in non-Western contexts
AI systems designed to detect deepfakes and synthetic media perform significantly worse on content from non-Western regions, featuring non-white subjects, or in non-English languages, due to biases in their training data. This detection gap leaves Global South populations vulnerable to AI-generated disinformation during elections and other sensitive periods.
What happened
AI-powered deepfake detection tools currently available on the market show confidence rates of 85-90 percent when determining whether content was made with AI, but this accuracy significantly declines when analyzing content from non-Western regions or featuring non-white subjects or non-English speakers.
The reduced accuracy stems from training data biases, as AI models were primarily trained on data from Western markets using English-language content and faces predominant in the Western world, limiting their ability to recognize content falling outside these parameters.
Reported impact
Journalists and researchers in much of the Global South have few resources to detect AI-generated disinformation, as detection tools are less effective for content originating from regions like Bangladesh and Senegal.
The detection gap leaves populations in non-Western regions vulnerable to synthetic media used for political purposes, including during elections.
Index analysis
The disparity in detection accuracy reflects underlying biases in how AI models were developed and trained, prioritizing English-language content and Western-appearing faces while excluding data from African and other Global South regions.
Digital data scarcity in some regions, where records exist primarily in hard copy rather than digitized form, has prevented AI models from being trained on representative datasets from these areas.
Governance controls
There are more readily accessible tools for creating synthetic media than for detecting it, creating an imbalance in the availability of technology for content generation versus verification.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.