@RepDonBeyer Published frameworks and incident reporting help. Someone still has to own the call when the model is wrong. I've written on that: https://t.co/CrDSNIgmfV
@ABAJournal If the agent picks the docs, does the work, and writes its own summary of what it did, you're grading its homework with its answer key. Firms need a log the agent didn't write.
I write about this here: https://t.co/CrDSNIgmfV
@thediary_keeper 100%. The scary part is a polished draft is easy to rubber-stamp. And juniors who never write the first draft never learn to spot what's wrong with one.
My research on this: https://t.co/CrDSNIgmfV
@ReutersLegal Uploading the doc was a policy fail. Telling the judge something false was a lawyer fail. The AI didn't sign the filing.
I've been researching exactly this: https://t.co/CrDSNIgmfV
@ryanmckeen Bans don't stop AI use. They just move it to personal phones where nobody can see what got uploaded.
More on why the review step matters: https://t.co/CrDSNIgmfV
@LuizaJarovsky For lawyers, transparency isn't abstract. If you can't explain what the tool did with client data, you can't explain it to the client, the board, or the judge.
I dig into this in my papers: https://t.co/CrDSNIgmfV
@ThisWeeknAI The 59:39 question is the one firms are least ready for. If the model does the first pass, the junior's job becomes checking it, and you can't check work you were never trained to do. Building that skill has to be someone's actual job, not a hope.
Privilege is usually the first question counsel asks. The quieter one comes right after it: when a draft goes out under a lawyer's name, what shows the lawyer did the legal reasoning? A purpose-built tool can help with the first. The second is a process each legal team has to own.
The tools keep getting better at the checkable parts: quotes, pin cites, deviations from a playbook. The supervision duty stays where it was. Formal Opinion 512 still expects the lawyer to understand what the tool did well enough to stand behind the work. Firms should decide now what that review looks like in practice.
@ryanmckeen Agree it's on the firm. The part firms will miss is measurement. A partner's markup used to show what a junior could actually do. Once a tool is in every first draft, you need a deliberate way to see their own work, or the gaps show up years later in front of a client.
@mebustelo@NISTcyber Consent fatigue is the right name. Lawyers know it from doc review: by hour six, everything looks responsive. 200 approvals a day isn't oversight.
@AlexTamkin Curious whether the 60% switch rate held when analysts had to write out their reasoning first. In law, that's what a memo is supposed to force.
@SINERD16 That BNY number is contract review, which worries me more than briefs. A fake cite in a brief gets caught by the judge. A missed clause in a contract might not show up for years.
@Keyon2046 Yeah. Whoever signs still owns it. What's gotten thin is the checking that used to happen before the signature. Your point about handoffs rings true, that's where most of these seem to slip through.
@vsachar The firms that already had AI policies are the part I'd want to dig into. My guess is the policy said "verify" and nobody was ever assigned to actually do it.
@LesOttolenghi Picking one owner and giving them real authority makes sense to me. I'd just want that owner to have someone who can spot a bad output, plus some record that they looked.
@Seeing_Redlines The revised invoice is the part that gets me. AI drops out of the line items and the total stays at $96k. So what was the "AI supervision" charge paying for?