4 ms·

Could it be worse ther police blatantly lying on their own?

by lupire 2y ago

more_corn 2y ago
I actually think it would be better. Ai would try to faithfully describe what happened. And if you instructed the ai to lie the prompt could be requested by the defense and the case would fall apart.
- Terr_ 2y ago
  > Ai would try to faithfully describe what happened. No, those haven't been invented yet. What we actually have are LLMs are still chock-full of biases and errors and hallucinations. Hell, we still haven't figured out how to stop them from being racist in nasty ways. What the AI here will actually do is try to "faithfully" make something that contains correlations it's seen before, even when the result is not what happened.
Terr_ 2y ago
My vote is on "worse", for several reasons. 1. Ghostwritten reports will make some officers seem like better witnesses than they actually are. The existence of half-assed report full of big mistakes could itself be important evidence in court someday, but the machine will replace that potential evidence with something which makes a jury think that the officer has good recall and clear thinking etc. 2. What officers may do (e.g. stops, searches) sometimes depends on what they were aware of and what they believed at the time. While there are already recipes to exploit that (e.g. "I smelled X therefore I can barge in") an LLM could easily learn to insert the normal post-hoc "explanations" automatically, streamlining the abuse. 3. It makes it easier for deliberately-lying officers to craft a falsehood, because they get a preview of "what a computer might easily detect" that they can work with to massage the result. 4. Even with the Perfect Cop, LLMs are chock-full of bias correlations and nobody knows how to reliably stop it. The not-so-objective computer could easily write a report that makes a suspect seem more hostile and dangerous simply because of the suspect's grammar. [0] 5. It creates an excuse for omissions and contradictions. "Oh, that part? That was by the software, I should have proof-read it more carefully but it's not something I did deliberately, trust me." [0] https://www.nature.com/articles/s41586-024-07856-5 https://www.nature.com/articles/s41586-024-07856-5
- avidiax 2y ago
  > 3. It makes it easier for deliberately-lying officers to craft a falsehood, because they get a preview of "what a computer might easily detect" that they can work with to massage the result. I think this is really key. The officer is going to see some boilerplate justifications and verbiage and not really consider whether those justifications really apply to this case. "Computer said so" will normalize all sorts of abuses, turning "is that marijuana I smell?" into "Responding officer detected a strong oder of illicit substances (marijuana) from the residence. Male subject #1 did not contest the finding, whereupon the responding officer established probable cause to a search of the residence..."
- lithos 2y ago
  Honestly a real lawyer is likely to have fun with an AI report. "Is the AI more competent than you?" Is a question that answered any way screws the person using the AI.

Could it be worse ther police blatantly lying on their own?

by lupire 2y ago

Could it be worse ther police blatantly lying on their own?

I actually think it would be better. Ai would try to faithfully describe what happened. And if you instructed the ai to lie the prompt could be requested by the defense and the case would fall apart.

Terr_ 2y ago
> Ai would try to faithfully describe what happened. No, those haven't been invented yet. What we actually have are LLMs are still chock-full of biases and errors and hallucinations. Hell, we still haven't figured out how to stop them from being racist in nasty ways. What the AI here will actually do is try to "faithfully" make something that contains correlations it's seen before, even when the result is not what happened.

Terr_ 2y ago

My vote is on "worse", for several reasons. 1. Ghostwritten reports will make some officers seem like better witnesses than they actually are. The existence of half-assed report full of big mistakes could itself be important evidence in court someday, but the machine will replace that potential evidence with something which makes a jury think that the officer has good recall and clear thinking etc. 2. What officers may do (e.g. stops, searches) sometimes depends on what they were aware of and what they believed at the time. While there are already recipes to exploit that (e.g. "I smelled X therefore I can barge in") an LLM could easily learn to insert the normal post-hoc "explanations" automatically, streamlining the abuse. 3. It makes it easier for deliberately-lying officers to craft a falsehood, because they get a preview of "what a computer might easily detect" that they can work with to massage the result. 4. Even with the Perfect Cop, LLMs are chock-full of bias correlations and nobody knows how to reliably stop it. The not-so-objective computer could easily write a report that makes a suspect seem more hostile and dangerous simply because of the suspect's grammar. [0] 5. It creates an excuse for omissions and contradictions. "Oh, that part? That was by the software, I should have proof-read it more carefully but it's not something I did deliberately, trust me." [0] https://www.nature.com/articles/s41586-024-07856-5 https://www.nature.com/articles/s41586-024-07856-5

avidiax 2y ago
> 3. It makes it easier for deliberately-lying officers to craft a falsehood, because they get a preview of "what a computer might easily detect" that they can work with to massage the result. I think this is really key. The officer is going to see some boilerplate justifications and verbiage and not really consider whether those justifications really apply to this case. "Computer said so" will normalize all sorts of abuses, turning "is that marijuana I smell?" into "Responding officer detected a strong oder of illicit substances (marijuana) from the residence. Male subject #1 did not contest the finding, whereupon the responding officer established probable cause to a search of the residence..."
lithos 2y ago
Honestly a real lawyer is likely to have fun with an AI report. "Is the AI more competent than you?" Is a question that answered any way screws the person using the AI.