Turn on the Swan mirror and the first thing it does is flatter you. A ring of adjustable LEDs wraps the 15.6-inch screen and throws the even, shadowless light that fills under-eye hollows and softens the pores a window would expose. That ring of light, not the artificial intelligence behind the glass, is what people are paying $795 for. Only once the flattering is done does a 4K camera photograph you and hand back a skin score out of 100. The light is there to make you look your best; the camera is there to find what is wrong.

Flattering light is the oldest trick in the room. Put a ring of bulbs around a mirror and you have the look of every theatre dressing room; build that ring into a camera and you have an invention with a precise birthday. In 1952, Lester A. Dine made the ring flash so dentists could photograph a mouth without shadows. Portrait and fashion photographers followed, because it nearly eliminates shadow and sets a clean circular catchlight in the eye. Swan’s move is to centre a camera inside a large display-ring rather than mount a ring on the end of a lens, and to file the result under intelligence.

The intelligence is the part that wobbles. A Coveteur reviewer scanned her face three or four times and watched the scores drift five to ten points between near-identical selfies, then collected a flawless 100 for acne while she was visibly breaking out. The ingredient guidance, she noted, was “genuinely helpful”; the device “definitely looks better on my vanity” than the generic ring-light mirror beside it. That sentence is the product. You pay $795 for an object that looks good on a bathroom shelf and a light that improves your face on camera; the skin score is the reason you give yourself for wanting it.

The score is not where the category competes anyway. At CES 2026, Amorepacific and Samsung showed a beauty mirror whose scan is trained on more than 450,000 cases, alongside Perfect Corp’s instant color diagnosis and L’Oréal’s wavelength-tuned LED mask. Skin analysis has become table stakes, shipped by firms with datasets Swan will never assemble. What Swan owns that a 450,000-case model does not is an object people want to be filmed beside, and that is a design problem, not a machine-learning one.

The camera is the only new thing in the frame, and it is the part that works least well.

This is the tell for where AI beauty hardware is going. Rack & Reason has already traced how a viral bachelorette made Swan a name; what that audience bought is, under the software, an old idea about light. The next launch will train a larger model and write a steadier scan, and it will still rise or fall on whether the thing looks good switched on, on a vanity, in a phone’s front camera. Build the glow first. The intelligence can follow the light, the way it always has.