4 ms·

Given how LLM generated stuff has gone over with courts thus far, I can't see one accepting an AI summary as evidence. They're describing basically using AI as

Given how LLM generated stuff has gone over with courts thus far, I can't see one accepting an AI summary as evidence. They're describing basically using AI as grep -r, which seems reasonable with the assumption that they are following due process to access the device storage. And if they aren't the problem isn't the LLM.

neodymiumphish 2y ago
They wouldn’t use AI to produce court exhibits. These are to help investigators quickly sift through the phone data to determine what data is worth investigating further.
- outer_web 2y ago
  Exactly my point, "hallucinate evidence" makes it seem like it'd be used as, well, evidence.
  - pixl97 2y ago
    Eh, that's not exactly how that would work, eh, hopefully. You'd dump 1000 hours of video into the AI for example. The AI would say at hour 666 a crime may be being committed. You'd then view and present that chunk of video as evidence.
    - trod1234 2y ago
      Hallucinations are a thing you cannot know or predict ahead of time in any specific detail. Human's also have biases which favor the first choices they make in maintaining internal consistency. About all you have is the fact that they will happen. Now, following your own example: You'd dump 1000 hours of video into the software which is not fully disclosed for example. The Officer or Investigator is told to look at 666 hours of related material. They manually review the first portions, the video is contextually consistent with their narrative. They pick and choose the video consistent in supporting their narrative, discarding the rest which may be exculpatory. The same way a news room might cut pieces of news together to take an interviewer's words out of context, structuring as a "Complex Question" fallacy. The burden and cost to prove otherwise is on the defense, which most often will not have the same access to the same software. Now it doesn't matter if you debunk it later, the jury goes with their first initial decisions. Tainting the jury. Subtle but that's how it would happen.
stefantalpalaru 2y ago
[dead]

neodymiumphish 2y ago

They wouldn’t use AI to produce court exhibits. These are to help investigators quickly sift through the phone data to determine what data is worth investigating further.

outer_web 2y ago
Exactly my point, "hallucinate evidence" makes it seem like it'd be used as, well, evidence.
- pixl97 2y ago
  Eh, that's not exactly how that would work, eh, hopefully. You'd dump 1000 hours of video into the AI for example. The AI would say at hour 666 a crime may be being committed. You'd then view and present that chunk of video as evidence.
  - trod1234 2y ago
    Hallucinations are a thing you cannot know or predict ahead of time in any specific detail. Human's also have biases which favor the first choices they make in maintaining internal consistency. About all you have is the fact that they will happen. Now, following your own example: You'd dump 1000 hours of video into the software which is not fully disclosed for example. The Officer or Investigator is told to look at 666 hours of related material. They manually review the first portions, the video is contextually consistent with their narrative. They pick and choose the video consistent in supporting their narrative, discarding the rest which may be exculpatory. The same way a news room might cut pieces of news together to take an interviewer's words out of context, structuring as a "Complex Question" fallacy. The burden and cost to prove otherwise is on the defense, which most often will not have the same access to the same software. Now it doesn't matter if you debunk it later, the jury goes with their first initial decisions. Tainting the jury. Subtle but that's how it would happen.

stefantalpalaru 2y ago

[dead]