> no trouble uploading all of that data in less than a week
When you're doing e-discovery, deadlines are often measured in days - not just for the upload time, but for the analysis and finding the needle in the haystack.
When you're doing e-discovery, deadlines are often measured in days - not just for the upload time, but for the analysis and finding the needle in the haystack.
I assume you’re right and AI now does some of the work but I doubt all of it. Also how reliable would the AI be… you’d hate to not have critical evidence at trial because you trusted the AI fully and it missed something.
Discovery data includes audio, video, social site data, as well as the usual documents and emails.
An example from the Karen Reed case, the police, somehow, uploaded a video that had been put through a "mirror filter" and thus showed a vehicle in the opposite orientation from reality. Is your LLM going to notice that?
You really shouldn't take the absence of evidence as any sort of evidence itself.
> You seem to have your mind made up anyways.
I haven't made up my mind about anything; it's you who claimed that using an LLM "is a great way to get sued for negligence." It's a fundamental rule of debate that person who makes the argument bears the burden of supporting it.
You seem to be making the implicit assumption that using an LLM to assist with the process will probably be found to constitute negligence. Again, why should anyone believe you, especially if it hasn’t happened yet? Your argument is just FUD, pure and simple.
As an attorney I can tell you these questions just aren't that simple. You can get sued for anything. But that's not really all that important. What matters is whether LLMs would do a worse job of performing document review than human review would. The answer to that question will depend on the specific facts of the case and the current state of the art.
We simply don't know yet what the error rate is of using an LLM; and the tech is improving rapidly. One should expect enterprising attorneys to test them out experimentally to build trust. For example, they can easily be tested vs. human review on small document corpuses.