Junior patent attorneys drafted stronger claims with AI help but lost the advantage once assistance was removed, while senior colleagues retained lasting gains. A three-month randomized trial with 133 attorneys across eleven US intellectual property firms found that access to an AI assistant raised drafting scores at 10 and 90 days, yet mattered for business because immediate productivity did not convert into independent judgment for less experienced staff.
How the three-month trial with InFlow worked
The working paper comes from the National Bureau of Economic Research, a nonprofit US economics research organization, and describes a randomized controlled trial in professional practice. Two-thirds of participants received access to "InFlow," an unreleased AI assistant from Google Labs, in a Google-funded study, while one-third worked without AI. The design separated assisted performance from internalized skill by testing drafting with AI and then requiring unaided evaluation afterward.
Participants completed two patent claim drafting tasks in which only the treatment group could use InFlow, alongside a third redlining task performed by everyone without AI. In redlining, attorneys had to critically review an existing patent draft, work the authors describe as routine and normally done without tools. All results were scored by blinded patent attorneys at an independent firm across five categories: enforceability, technical accuracy, strategic ambiguity, completeness, and clarity.
In drafting, the AI group outperformed the control group at both measurement points and across all five scoring dimensions. The improvement came mainly from fewer low ratings and more mid-range scores, with top scores barely moving. Gains were much larger for juniors with fewer than seven years of experience than for seniors, a pattern the researchers link to earlier findings in writing, customer service, and software development where AI narrows gaps between experience levels.
What the result means for firms using AI
Redlining reversed the picture. The treatment group scored better overall, but the entire advantage came from senior attorneys, while juniors showed no benefit despite their strong drafting gains with AI. Junior scores polarized instead: fewer mid-range results gave way to more poor and more good outcomes, with no growth at the top. The authors interpret this as AI accelerating skill formation for some juniors while replacing practice for others, with the largest immediate beneficiaries retaining the least durable ability.
Text analysis points to a difference in working method rather than effort alone. Senior attorneys more often restructured the flawed draft and justified edits with legal principles, while juniors frequently identified serious flaws without fixing them and left notes resembling instructions to an AI assistant. Since this pattern also appeared in the control group, the authors treat it as a baseline junior deficit that AI use did not correct, suggesting that prior expertise helps turn assisted practice into lasting capability.
The findings carry limits that buyers of legal AI should weigh before changing training or staffing. The authors cite small sample size linked to high billable rates of top attorneys, participating firms that do not represent all professional services, no baseline skills measurement, and a three-month window that is short against years needed for patent expertise. Similar signals appear elsewhere: an Anthropic study with 17 percent lower follow-up scores for younger developers learning with AI, MIT Media Lab work on outsourced thinking in writing, a 2023 BCG study where weaker consultants gained most from GPT-4, and Anthropic reporting junior engineering roles falling from 21 percent to 14 percent since early 2025.
The marker to watch is whether firms preserve unaided judgment tasks such as live redlining, client meetings, and courtroom exchanges where real-time assistance is unavailable. If junior hiring shrinks or training shifts fully to AI-assisted drafting, the profession may gain output while spreading weaker independent review across firms. Durable advantage will belong to organizations that measure both assisted drafts and solo revision quality.
