Gregory states that many deepfake detection tools were trained on high-quality media, posing challenges in regions like Africa where lower-quality smartphone images prevail, complicating detection.
Ngamita points out that cheap smartphones lead to inadequately stable images that confuse detection models, as they were not trained for this kind of variability in quality.
Gregory remarks that free detection tools available to journalists and fact-checkers are extremely inaccurate, failing to address who is represented in training data and the challenges of lower quality material.
Diya expresses concern that faulty detection models could flag non-Western content as AI-generated, risking policy overreactions to non-issues created by inflated statistics.
Collection
[
|
...
]