There are analysis techniques that can determine if AI generated an image. They aren’t certain, but the can provide a % degree of confidence.

The problem is that AI will get better at avoiding this (even if it’s not being designed to specifically), and AI “ enhancement” also leaves similar artifacts (things like clearing up sensor-tearing or compression artifacts; iOS does a bunch of this and I think Android too)

Detection will always be an arms race; the real problem is how to we verify those techniques? How do we determine that government witnesses are testifying to verifiable methods? If we cannot do that then everything falls apart, from responsible disclosure in news articles to trials in a court of law.

The same way you detect any other fraud. It can range from analysis to whistleblowers. The point is not even to catch all fraud, it's to catch it often enough that most people and companies will not attempt it.

As often, it's more of a liability thing than anything else.