Up until that point I'd have thought I would be a reliable witness (I didn't, but this proved it definitively!)
It was really eye-opening for me. Thinking about it is still kind of eerie. I remember the accident. I can see it in my head. But apparently the memory is not exactly correct.
I served on a jury where an attractive, sharp defense attorney just ripped a witness to shreds. The person testifying was a customer engineer who installed and repaired a particular piece of equipment, whose entire function was dependent on time. He was testifying about what he did every day, not some nuance about direction of travel.
She built the poor guy up and then threw him off a cliff. At the end he ended up testifying that time is impossible to accurately measure. The issue was pretty obvious - the incompetent DAs didn’t prepare the witness for cross examination.
There was nothing clever about the question. Maybe they had a few up their sleeve for subsequent questions, but we never got that far, as I was immediately dismissed. I spent maybe 60 seconds in the courtroom.
It really was a case of casual observer not having useful information. Just because I saw and heard the impact doesn’t mean that noticed any of the details that would have made a difference to anyone.
There's plenty of research about the unreliability of memory.
Presumably lawyers are very good at finding exactly the right questions to ask to exploit the weaknesses of human memory and make a witness look unreliable.
Big citation needed there...
> we should release every person who is in prison because an eyewitness ID'd them.
in spite of my previous statement this is not wrong if an eyewitness ID (from a stranger to the defendant) was the only thing that led to conviction
People might still be better at recognising faces they have learnt, and they might be worse in terms of percentage correctness, but have less bad failure modes.
Do you have anything other than an assertion for that?
Otherwise AI is scalable and can recognize millions of faces.
So when it comes to even tiny recognition for safety or security - it doesn't matter how good are humans at recognizing a few faces they can remember.
The metric that is used is extremely important. If the metric is "identify lots of faces" then AI is superior. If "identity faces correctly" is the metric, then it is far from certain or obvious that an AI is superior to a human.