You can. People do it every day. Here is exactly what you get when you do, and what you do not.
You ask a chatbot
It looks at your image and reasons about it, the way a person would — noticing hands, lighting, texture. That is observation, not measurement. It has no forensic detector running underneath.
REAL runs purpose-trained detectors built for one job: finding the statistical residue generative models leave in pixels, waveforms, and text. Several of them, independently, on the same file.
You ask it twice
You get two different answers. Language models are probabilistic by design — rephrase the question and the verdict can flip.
The score is computed by fixed rules in code. The same evidence returns the same result today, next week, and when someone challenges it. Reproducibility is not a nicety in forensics; it is the point.
You tell it what you are hoping for
It accommodates you. Chat models are trained to be agreeable, and the way you frame the question shapes the answer you get back.
Rules do not care what you were hoping for. And we will not say “authentic” even when you want us to — that verdict does not exist here.
The evidence is too poor to judge
It answers anyway. A chatbot will rarely tell you it cannot tell — it produces something confident-sounding from whatever it has.
Unable to evaluate is a real, first-class result here. If the face is never clearly visible, if the audio is too processed, you are told that plainly instead of handed a manufactured number.
Someone disputes you
You have a chat log. You ran it, you wrote the prompt, and you are quoting yourself. It persuades no one who does not already believe you.
You have an independent review by a separate party, with a case number, a timestamp, a fingerprint of the exact file, and a registry record the other person can check without taking your word for anything.
You upload private evidence
It goes to a general-purpose model, under whatever data terms that provider happens to run this quarter.
Your files are processed in an isolated environment and are never used to train or refine external generative models. Evidence is retained only as long as processing and support require, then deleted.
You want to show it to someone
You screenshot the conversation.
You get a REAL Check — a shareable mark carrying the result, the date, the case number, and a code anyone can verify in seconds.
An AI can tell you what it thinks.
It cannot hand you proof.
