We built an AI that says "I don't know"

Point a camera at something and the listing writes itself. That's the easy part and everybody has it. The hard part is the moment the model doesn't actually know what it's looking at.

What we found when we measured it

It came back "high confidence" nineteen times in twenty. Including every single time it was wrong. That's the worst arrangement available, because from where the seller is sitting the two look identical.

What we changed

We stopped asking how confident it felt and started asking something it can actually answer: did you read the model designation off the item, or work it out from the shape? High confidence has to be earned by the first one now.

• When it reads a model number it quotes the characters back, so you can check them against the thing in your hands.
• When it can't, it names the two or three products it might be and prices each, instead of picking one.
• Photograph a tablet face down with the screen off and no logo in shot, and it refuses to name the brand and asks for the back panel.

The fault we haven't fixed, on the record

About one run in three it can still tell you it read a string that isn't on the item. We haven't solved that. What we've done is make sure the string is always shown to you, so the check happens somewhere it actually can.

We'd rather hold you up for ten seconds than put somebody else's model number on your laptop.