Three months of AI assistance improved patent lawyers' AI-assisted drafting quality for everyone, but only senior lawyers showed durable skill gains in unassisted work; junior lawyers showed none on average, with their scores bifurcating into more poor and more good outcomes.
Pre-registered three-month RCT with 133 practicing patent lawyers (91 completers) at eleven US IP law firms, randomized 2:1 to a custom Google-built AI drafting assistant or control, stratified by firm and experience. Blinded expert patent attorneys scored drafting tasks at 10 and 90 days and an unassisted redlining task at 90 days on a five-domain rubric; LLM-based ratings and permutation tests were used for robustness.
AI-assisted drafting gains of 0.34 SD at 10 days (p=0.03) and 0.38 SD at 90 days (p=0.01), larger for juniors; unassisted redlining gain of 0.32 SD overall (p=0.04), concentrated entirely among seniors (0.45 SD, p=0.02), with no junior gain. Time savings were modest: 10 minutes on a 112-minute baseline task at 10 days (p=0.05), not significant at 90 days. Findings hold across rubric subscales, raters, seniority thresholds (3-9 years), and non-adherence exclusions.
Authors concede: no baseline skill assessment (firm recruitment constraints), so the skill-acquisition interpretation rests on assumed treatment-control balance at baseline; 91 of 133 completed the protocol; sample is limited to Google-affiliated, sophisticated patent firms; 15 of 91 submissions showed possible AI non-adherence; three months is short and the junior score dispersion may not persist; the mechanism for the junior bifurcation is unknown. Unstated: heavy Google funding and authorship of the tool, single-task single-domain external validity, and the possibility that the junior 'no gain' result is a low-power null.
A skeptic would note this is a vendor-funded study of a custom in-house tool at firms dependent on Google's patent business, that the headline 'expertise erosion' finding is really a null plus a dispersion shift in a small junior subsample, that blinded human raters and LLM raters disagree (LLM ratings did detect junior redlining gains), that non-adherence detection via stylized 'tells' is itself unreliable, and that one unassisted redlining task is a thin proxy for 'professional judgment.'
If replicated, the near-term economics point to AI compressing junior-level output quality gaps (faster, better drafts from juniors) while weakening the traditional apprenticeship pipeline that produces tomorrow's seniors, which threatens law-firm and other professional-services labor models more through training collapse than through immediate headcount cuts. For the next 12 to 24 months this is a directional signal, not a staffing-plan-grade result.
Labor & Knowledge Work: Junior knowledge workers get the largest immediate productivity lift but appear to accumulate the least durable skill, directly bearing on entry-level hiring, training investment, and career-ladder design.
Market & Valuation: If AI compresses junior-senior output gaps in high-billing professions, time-based revenue models in law and consulting face margin pressure even before headcount changes.
Technical Capability: The result shows that realized capability gains from AI depend on user expertise, conditioning claims that model quality alone determines real-world value.