take

The review bottleneck moved from drafting to citation checking

I'm a paralegal at a 14-attorney litigation boutique. Eighteen months ago, I spent maybe 60% of my week drafting first passes of motions, discovery responses, and research memos. Now that's down to about 15%. The model handles it. What replaced it isn't free time, it's verification work nobody costed in. Every cite has to be pulled and read. Every quoted holding has to be checked against the actual opinion, not the model's paraphrase. I caught a fabricated parenthetical in a brief draft last month that would have made it past a junior associate because the case was real and the holding was almost right. Almost. The partners think I'm 3x more productive. By billable output, sure. But the cognitive load shifted from generative to forensic, and forensic is more tiring. I'd rather draft from scratch than audit ten pages of plausible-looking text for the one wrong subsection cite. My honest take: the tools made the easy part easier and the hard part more important. Nobody at my firm has adjusted hiring or training around that yet.

▲ 18·layla_haddad·2d

6 comments

0chinedu_eze·2d

Solo founder here, same pattern on a smaller scale. I can draft a launch post or a support reply in seconds, but verifying every link, version number, and pricing claim before it goes out eats more time than writing ever did.

0karim_r·2d

Same pattern in my classroom. I used to spend grading time rewriting feedback comments; now I spend it verifying that the lesson plans the tool generated actually cite real standards and not fabricated ones. Net time saved is maybe 20 percent, not the 70 the vendor promised.

0meiwong·1d

Same thing happened with student essays. I used to spend prep time finding sources to recommend, now I spend it verifying that the three citations in a six-paragraph essay actually exist and say what the student claims. Curious whether anyone in a non-academic setting has found a workflow that catches fabricated quotes faster than just opening each link.

0CamilaTorres·1d

Same pattern on our analytics team of 6. The model drafts a stakeholder memo with five plausible-looking metric references in 30 seconds, and I spend 40 minutes confirming each one actually matches what's in dbt. Net win is real but smaller than the demo suggests, and the failure mode is silent which is what worries me most.

0noah_anderson·1d

Same pattern here, except the citation checking expanded to include checking that the cited case actually says what the draft claims it says. About 40% of the time it doesn't, or the quote is from a dissent. Drafting time dropped maybe 60%, verification time roughly doubled, so net win is smaller than partners think.

0noah_anderson·1d

Same pattern in our lab. Three of us used to spend a week each on lit reviews; now drafting takes a day but verifying every citation still eats most of the week because the models hallucinate plausible DOIs and reassign findings to the wrong authors. Has anyone found a workflow that catches the subtle ones, like a real paper cited for a claim it doesn't actually make?