3 ms·

The software being trusted to pick out similar faces is a GREAT ADVANCE IN SOLVING CRIMES because it scales the process going on in a photo lineup. But it's on

by mapt 3y ago

The software being trusted to pick out similar faces is a GREAT ADVANCE IN SOLVING CRIMES because it scales the process going on in a photo lineup. But it's only being trusted to pick out similar faces. There will ALWAYS be similar faces. It's doing the same part of the task as a photo lineup does. The point is that there has to be MORE than "This person has a similar face". It's a confirmation step for suspects that already exist, or for whom you can find other strong circumstantial evidence for. You have an unlimited supply of similar faces and you don't get to just choose one of them. The great problems with AI tools mostly arise from the fact that the supposed expert users were never taught what the tools do, or how causal inference works - how to discern reality from the evidence available, how to test hypotheses, how to change your reasoning to be valid in the face of wildly disparate scales. If you don't have at least the beginning of Bayesian probability, and you're feeding large datasets into policing, you have no fucking clue what you're doing and your ignorance is immediately dangerous.

woodruffw 3y ago
No crimes were solved here. We can't take a thing that literally did nothing except cost taxpayer money, police time, and stress for a pregnant woman and declare victory for it. > It's doing the same part of the task as a photo lineup does. It's doing something subtly different, and arguably pernicious: a lineup is meant to contain a variety of individuals of varying similarities, so that the person selecting from the lineup are themselves tested in their recollection. We have no evidence that this system is being used in this way; instead, it seems to be finding the closest match in an existing dataset under the assumption that anybody who's ever been photographed by the police is permanently a criminal suspect.
- mapt 3y ago
  >instead, it seems to be finding the closest match in an existing dataset under the assumption that anybody who's ever been photographed by the police is permanently a criminal suspect. So is a witness, for the most part, they're just logistically challenged. Which is why photo lineups are, in my alternate universe where everybody understands probability and causal inference (because we require 4-year degrees in the science of criminology), a specific tool to discover specific things rather than an instantaneous indictment. There are myriad things wrong with using them, and if the detective doesn't understand that or worse wants to exploit that, the detective should be regarded no differently than the beat cop shooting their gun into the crowd to scare it off. https://lawcomic.net/guide/?page_id=5 https://lawcomic.net/guide/?page_id=5 Similarly: A detective's hunch about whether the person being interrogated committed a crime is a powerful tool. We just don't throw people in prison based on that alone. We're getting to the point that we could train an AI to develop such hunches, and that could be valuable. But we can't just throw people in prison based on that alone. We need to confirm such things with external evidence. Just like we need to confirm a single-eyewitness recollection or a rough photo, rather than just measure that recollection & numerically characterize that photo without any indication (or more commonly, with a 100% fraudulent indication) of probability of matching.

by mapt 3y ago

woodruffw 3y ago

No crimes were solved here. We can't take a thing that literally did nothing except cost taxpayer money, police time, and stress for a pregnant woman and declare victory for it. > It's doing the same part of the task as a photo lineup does. It's doing something subtly different, and arguably pernicious: a lineup is meant to contain a variety of individuals of varying similarities, so that the person selecting from the lineup are themselves tested in their recollection. We have no evidence that this system is being used in this way; instead, it seems to be finding the closest match in an existing dataset under the assumption that anybody who's ever been photographed by the police is permanently a criminal suspect.

mapt 3y ago
>instead, it seems to be finding the closest match in an existing dataset under the assumption that anybody who's ever been photographed by the police is permanently a criminal suspect. So is a witness, for the most part, they're just logistically challenged. Which is why photo lineups are, in my alternate universe where everybody understands probability and causal inference (because we require 4-year degrees in the science of criminology), a specific tool to discover specific things rather than an instantaneous indictment. There are myriad things wrong with using them, and if the detective doesn't understand that or worse wants to exploit that, the detective should be regarded no differently than the beat cop shooting their gun into the crowd to scare it off. https://lawcomic.net/guide/?page_id=5 https://lawcomic.net/guide/?page_id=5 Similarly: A detective's hunch about whether the person being interrogated committed a crime is a powerful tool. We just don't throw people in prison based on that alone. We're getting to the point that we could train an AI to develop such hunches, and that could be valuable. But we can't just throw people in prison based on that alone. We need to confirm such things with external evidence. Just like we need to confirm a single-eyewitness recollection or a rough photo, rather than just measure that recollection & numerically characterize that photo without any indication (or more commonly, with a 100% fraudulent indication) of probability of matching.