“There doesn't seem to be a very good accountability mechanism here,” said Victoria resident and frequent police critic Stephen Harrison, who obtained the documents. “Literally nobody other than the individual officer has access — you wouldn't even know that they're using it.”
I think there is still a significant issue with accountability and accuracy.
Are they more or less accurate than a trained human legal expert? Maybe equivalent, but they’re non-deterministic so will frequently have different answers to the same question asked at different times. They will typically bull ahead with incorrect information rather than output that they’re unsure and need to investigate which you would expect a human expert to say.
LLMs also have no accountability so all decisions must be verified by an accountable expert. Given the time it takes to read and absorb a potentially wrong advice, and given you need to ask the legal expert anyway, you might as well ask the expert in the first instance.
yeah, well, this is what I mean by the inherent problems. The open models still hallucinate, they are still probability-driven. Even when guided by RAG and other precautions, it still is only as good as those supporting elements can offer (assuming they are high-quality themselves).
On balance, I do not understand how this kind of tool will help officers in the field unless it is to advise on procedures for the officer to follow. Not trying to interpret situations as offenses or not, not trying to be a pocket legal interpreter.
I only suggest that whomever decided to try this for VicPD is being supported by a LLM geek that probably at least cares how well the models are performing. They seem to understand what kind of safeguards should exist for this type of thing to exist in this space… I personally doubt its good enough, but I rather see this than ChatGPT.
You’re wrong on the first count primarily because you presume RAG retrieval isn’t deterministic. Which it is, even if you format the question differently - as long as the request is coherent and covers the same topic, RAG lookup should be roughly the same, and that should result in the same RAG entries surfaced.
At which point it’s the prompt (the hidden/system prompt, not the end user’s basic question) that determines just how well that RAG data gets displayed and potentially reworded. But as long as it’s essentially used as a human language lookup system + summarisation, its reliability should be generally pretty good, on par with having an on-site paralegal.