Or we find out that certain population groups have different blood flow patterns, which the system incorrectly identifies as proof of fakery. Or perhaps for some, it's simply not detectible even though they are real live people.
Yeah, neither deepfakes nor deepfake detectors will end epistemology. We'll need to use a multiplicity of tools, with strengths and weaknesses known and unknown, and come to a conclusion based on the preponderance of evidence knowing full well we will sometimes get it wrong.