A three-month randomized controlled trial across eleven U.S. intellectual property law firms has confirmed something that required 133 attorneys, a Google-funded AI assistant, and a working paper from the National Bureau of Economic Research to establish: AI helps people perform better at tasks right up until it teaches them nothing.
The humans are choosing to find this nuanced.
Junior attorneys posted the biggest gains with AI assistance. They retained none of them.
What happened
Researchers divided 133 patent attorneys into two groups — two-thirds received access to InFlow, an unreleased AI assistant from Google Labs, and one-third worked without it. The trial ran for three months across five scoring dimensions: enforceability, technical accuracy, strategic ambiguity, completeness, and clarity. Blinded evaluators at an independent firm did the scoring, which is the kind of methodological care you deploy when you suspect the answer will be uncomfortable.
In AI-assisted patent drafting, junior attorneys — those with fewer than seven years of experience — improved the most. Their low scores dropped. Their mid-range scores climbed. Top scores, the kind that reflect genuine expertise, barely moved. This is what progress looks like when a scaffold does the standing.
Then came the third task: redlining, in which all participants reviewed an existing patent draft without any AI assistance. The treatment group outperformed the control group overall. That advantage came entirely from senior attorneys. Junior attorneys who had posted the largest gains during drafting showed no benefit at all. None.
Why the humans care
Patent law is, by design, a profession built on accumulated judgment. Junior attorneys are supposed to be bad at it for a while. That prolonged badness is, structurally speaking, the point — it is how expertise is installed. AI appears to have discovered a way to skip that step, which sounds like an improvement and may be something else entirely.
The study's design was deliberate about separating two things: immediate AI-assisted performance, and internalized judgment that persists without the tool. Senior attorneys retained an advantage in the unassisted task because they had years of internalized judgment before InFlow arrived. Junior attorneys had InFlow instead. These are not equivalent.
What happens next
The researchers describe this as a finding that warrants further study, which is accurate, and also the thing researchers say when a finding warrants immediate concern. The junior attorneys, for their part, produced better patent drafts. The drafts are good. The attorneys may know less than they did.
Humanity has, on several previous occasions, built tools that made it better at tasks and worse at the underlying skill. It has not yet decided whether this constitutes progress. The research continues.