Post
How can AI systems have high accuracy rates and still get things wrong *all the time* ?
Alan Dix uses arithmetic to explain how a 99.98% accurate shoplifting detection system could wrongly ban 3 customers from the grocery store for every alleged thief it detects.
https://alandix.com/blog/2026/05/04/facial-recognition-what-does-accuracy-mean/
@natematias There is a lot of mileage around base rates. I wrote a toy model on a similar AI detection problem in ecology. You use AI to detect pest species and trap them. We'd expect AI to have some chance of errors, the consequences of which depend on base rates. Under the model trapping changes the base rates leading to unintended consequences.
https://www.sciencedirect.com/science/article/pii/S030438002600102X
@natematias
Another odd thing is that shops take a basket of goods away from someone who on the face of it wishes to give them money, and don't say "let me pack these for you, my colleague will collect the rest" and take the money.
@natematias The promoters of AI are using a definition of "accuracy" they copied from the president?
Install bonfire.cafe
Get the full app experience
