The Ninth Salon Journal · By Sophie, founder · Published 31 August 2026
A disclosure before anything else, because it changes how you should read this: our studio makes AI portraits and says so on every page. That gives us an unusual view of this question. We know what these systems do well, we know what they still get wrong, and we have no interest in pretending the tells do not exist.
The short version: in 2026, visual tells still work but they are shrinking, detector tools are unreliable in both directions, and the question worth asking is usually not "was AI involved" but "did anyone say so".
The oldest tell and still the most useful. Fingers that number wrong or bend where no joint exists, teeth that merge into one another, too many legs on a chair. Anything the eye counts without meaning to is where these systems slip.
Signage, book spines and labels in the background often come out as confident-looking letterforms that spell nothing. Zoom into any writing in the image before you zoom into the face.
Follow a shadow back to its light source. In generated images the shadows are often beautiful and mutually contradictory: two suns, or a reflection that shows something the scene does not contain. Jewellery, glasses and earrings are worth a second look too, since matched pairs frequently fail to match.
This is the subtlest one and, to us, the most reliable. Real photographs and real paintings contain accidents: a stray hair across the cheek, a thumbprint in the varnish, a fold that falls awkwardly. Generated images tend toward a plausible average of everything, so surfaces come out slightly too even and compositions slightly too obliging.
Automated "AI or not" detectors are worth using and never worth trusting alone. They produce false negatives on recent models and, more damagingly, false positives on genuine human work, which is how real illustrators have found themselves accused of fraud by a percentage score. Treat any detector output as one weak signal among several, never as a verdict.
The better long-term answer is not detection but provenance: files that carry a signed record of how they were made. Industry efforts around content credentials are moving that way, and some generators embed invisible watermarks in what they produce. The weakness is practical: screenshotting, re-encoding or uploading to many platforms can strip that information, so its absence proves nothing at all. Present and intact, it tells you a lot. Missing, it tells you nothing.
Most people asking "is this AI" are not conducting forensics. They are asking something more human: was I told the truth about what I am looking at, and about what I am paying for. That reframes the whole thing usefully. A seller who describes their method plainly is behaving well whether they paint by hand or by prompt. A seller who is vague about it has answered your question already.
So the practical checks are not visual at all. Does the page say how the work is made? If it is sold as hand-painted, is there process evidence: work in progress, brush texture under raking light, a named artist with a body of work? And what exactly are you receiving: a unique object, or a file that can be regenerated?
Ninth Salon generates portraits with AI from a photo you upload. We say it on the homepage, in the FAQ and in every article in this journal. What we sell is not the illusion of a human brush; it is a transformation of your own photograph into a classical style, at a price and a speed no commissioned painter can match, printable on canvas if you want it on a wall. When we compare that honestly against a real commission, we do it in our guide to what a custom portrait costs, including the cases where paying a painter is the better answer.
The distinction that matters is not human versus machine. It is disclosed versus concealed. Everything in this article works better once you sort by that.