[jklontz]

The AWS Rekognition study I believe you are referring to was fundamentally flawed in two ways.

First, the threshold used was the default suggested for one-to-one matching (lower false reject rate, higher false accept rate). It doesn't make sense to use this threshold for search applications, and when the study was reproduced with a more appropriate (higher) threshold there were 0 mis-identifications.

Second, law enforcement use of face recognition doesn't even involve the algorithm making a lights-out identification decision. Instead, the most similar faces are presented to the user in ranked order by similarity (like a search engine). It's a tool for generating investigative leads, often preferable to publishing the face image of a wanted perpetrator on local news.

[reaperducer]

the threshold used was the default

What makes you think that a LEO would change the defaults? They're not computer scientists, they're people who bought a product to perform a specific function.

It's a tool for generating investigative leads

So were photo books, lie detectors, surveillance cameras, and other tools at first. They became a close enough proxy for "guilt" in the eyes of the users for people to automatically be treated as criminals, even if they'd done nothing wrong.

[manigandham]

Was this the actual product used by law enforcement or just the same software workflow used for a study? Those are entirely different environments and there's no reason to expect the final offering to behave like the academic setup used in a specific report.

[HALtheWise]

I believe that there were multiple defaults, and they intentionally chose the wrong one to make a point. Your quote cropping is embarrassing.

The analogy would be blaming an anchor manufacturer when the "default anchor" for rowboats doesn't hold an oil tanker.

[skybrian]

The question is how do you prevent abuse without handicapping investigators too much, since presumably you do want cases to be solved if possible?