Have you tried your methods on a range of photographs that you know haven't been tampered with?
How about on a 'blind' selection, some tampered and some not?
That would be sort of the minimum standard of validation for the method. Some of the snippets from his AI chats suggest he's trying to get the AI to spot alterations to photos, or to validate his method by whatever statistical roils and eddies the AI employs. But at least in terms of how the technology components works, the purported telltales need to be either present or absent based on whether there was an alteration actually present.
If you haven't used control photographs how do you know your methods are valid?
It's time to talk about the elephant in the room.
If your process is simply you fiddling interactively with image processing tools, that's not science. This was my master's thesis: the eye-mind-hand effect. Why do engineering designs look different when they're sketched out on paper versus early-stage sketching using a 2D or 3D design software tool? It's not just that software has a particular vocabulary of design elements. It's that an engineer's brain works differently when the eye, mind, and hand are working in a tight feedback loop such as occurs with pencil and paper. That's great for designing things. Turns out it's not so good when you're trying to figure out some objective truth. The feedback becomes a bias. This is why real image analysis isn't often interactive. We use in-house tools based around image manipulation libraries that implement the basic algorithms, but you have to parameterize them programmatically using statistically (not visually) determined parameters. This is the reproducible element. When you do this for legal forensics, you may be using Adobe Photoshop or some other interactive tool, but there's a plugin that's recording everything you do so that someone else can exactly reproduce the analysis steps.
Thus the first problem with this method (and most other anomaly-hunting methods used by conspiracy theorists) is that interactive fiddling engages the eye and brain in the same way that gives you pareidolia in other circumstances. You stop fiddling when you see something you like, even if the phenomenon underlying the fiddling thereby stops at a nonsensical value. You can fiddle with the luminance channel until you think you can see something that was hidden in the dark spots, but you'll also reveal lots of quantization and compression artifacts. If that startles you, you'll just stop there thinking you've found a smoking gun. You aren't thinking about whether the adjustment curve represents a physically plausible scenario.
The second problem is just basic science: reproducibility. Any process that requires a human to gaze at the manipulated images and say, "Yep, that's evidence of tampering," is not reproducible in the scientific sense. That's not to say that human judgment isn't probative in other ways. But if a protocol developed under the auspices of the scientific method relies on human judgment to turn observation into conclusion, it's not considered reproducible. You need some objective criterion deductively tied to the hypothesized underlying phenomenon that can be evaluated objectively. That may have to be a statistically reckoned value, but it has to be a reproducible step.
If you want to show that a method is valid in a scientific sense, you have to show that any person can execute your protocol from beginning to end and arrive at the same measured value for the variable. Here,
@nt1 doesn't even acknowledge the human judgment element. He gets around it by simply presumptively declaring that a particular observation is per se evidence of tampering without explaining why. He shifts responsibility entirely onto the Photo Curves app, imagining that it objectively just spits out that judgment.
