When AI hands lawyers a quick summary, what actually eats their time: checking if it's true, or tracing where it came from?

This explores where lawyers' verification time goes when GenAI is part of the workflow. The short answer from the collection: the costliest step isn't judging whether a fact is true. It's finding out where the claim came from in the first place. Interviews with 18 lawyers found that GenAI summaries look efficient but often take longer than doing the work by hand, because lawyers have to trace each statement back to an unclear source before they can even start checking it Does GenAI actually save lawyers time on fact verification?. The culprit is opacity rather than error rate. Even a mostly correct summary has to be reconstructed line by line, because the lawyer stays professionally accountable for every line. The collection has only this one lawyer-specific study, so it can't give a finer breakdown of which tasks (case law, contracts, deposition records) cost the most.

The same pattern shows up well outside law, and this explains why the lawyers' experience isn't a quirk. Across the research lifecycle, AI produces plausible output faster than anyone can confirm it. The failures are mostly fabricated content and retrieval misses, not misunderstanding Can AI verify research outputs as fast as it generates them?. Formal mathematics is the extreme case. Even when proof-checking is completely free and automatic, the bottleneck moves to a harder question: does the formal statement actually mean what we think it means? Only experts can answer that, at a ratio of hundreds of checked proofs per statement that still needs a human to audit it Does free proof checking actually reduce verification burden?. Lawyers face the legal version of this. Confirming that a citation exists is cheap. Confirming that it says what the summary claims it says, in context, is expensive.

Two findings explain why the upstream work is so easy to skip and so costly when skipped. Deloitte refunded the Australian government after a report reached the client with invented court quotes and fake references, despite a nominal review process Can AI quality control catch fabricated citations in professional reports?. Fake references also work on machines: LLM judges give higher scores to answers that include authoritative-looking citations, whatever the actual quality Can LLM judges be tricked without accessing their internals?. A confident citation is exactly what makes a reviewer relax. And when professionals do push back, the model may not cooperate. A BCG study of consultants found that challenging GPT-4's output made it argue harder rather than admit its limits Does validating AI output make models more defensive?. Verification can turn into a debate instead of a lookup.

The collection hints at what would actually shrink the upstream burden: build checking into how the answer is produced, instead of auditing a finished product. Checking each intermediate step during long reasoning tasks raised success from 32% to 87%, because most failures happen in the process, not in the final answer Where do reasoning agents actually fail during long traces?. Research agents work better when they audit a draft answer one requirement at a time and go back for each unresolved one Should research agents verify answers before searching longer?. For a lawyer, this suggests the real time-saver isn't a better summary. It's a tool that shows its sources and reasoning steps, so the work of tracing claims back is already done.