Lawyers given an AI drafting assistant produced better patent drafts, but gains in later unaided judgment were concentrated among experienced lawyers, according to an October experiment.
The working paper, by MIT economist David Autor and six Google researchers, follows 133 patent lawyers at 11 US firms over three months. Ninety-one completed the full protocol. The paper is dated October 7 and is not a peer-reviewed journal publication.
Independent attorneys, unaware of treatment assignments, scored benchmark work. AI access raised drafting quality by 0.34 standard deviations after 10 days and 0.38 after 90 days. Those figures describe movements on a scoring distribution, not percentage improvements.
On a later task reviewing an application without AI, gains were concentrated among lawyers with at least seven years’ experience. Junior lawyers showed no average gain, with more dispersed results.
A better document and a better-trained employee are separate outcomes
An assisted task measures what a worker and a tool can produce together. An unassisted task asks what the worker can do when the tool is unavailable.
For a company purchasing AI software, both can affect the business case. Faster or better drafts may reduce current production costs. The ability to detect errors, challenge an output and explain a decision affects how work is supervised and what employees can take responsibility for later.
Patent drafting makes that question concrete. A patent application describes an invention and contains claims defining the protection sought. A polished sentence can still create a problem if its technical meaning is wrong or its scope does not match the intended invention.
Reviewing a draft therefore requires more than judging whether it reads well. The reviewer needs to assess the substance, identify omissions and understand the consequences of revisions.
The experiment tested a custom tool, not every available AI product. Google funded the work, and participating firms had patent-drafting relationships with the company. Three months cannot establish long-term career effects.
Responsibility for a filing stays with the people submitting it
The US Patent and Trademark Office’s guidance on AI-based tools, issued in 2024, explains that existing professional obligations apply when practitioners use AI in work before the office.
Those obligations include checking the accuracy of material submitted and protecting confidential information. AI-generated text is not an independent assurance that a filing is correct.
The guidance also discusses risks from sensitive information entering third-party tools, including how a provider may handle or retain that information. A firm’s assessment needs to cover the service and its arrangements, alongside the quality of the document it produces.
For employers, a software demonstration that produces fluent text leaves several operating questions unanswered. Who verifies the work? How much review is required? Can the employee explain why an edit is right? Does the process preserve the training needed to perform those checks?
Training needs its own measurement
A manager evaluating an AI rollout can record completed assignments, turnaround times and corrections. Those measures describe current output. Separate assessments are needed to establish whether staff are acquiring the judgment required for more demanding work.
For example, an employee could explain the reasoning behind a draft or review a prepared example before consulting an assistant. These are possible assessment approaches, not interventions established by this experiment to produce better learning.
Our earlier reporting on robot adoption and management practices examined a different technology but a related operating issue: buying equipment and reorganizing work are separate decisions. An AI subscription likewise leaves decisions about review, responsibility and training with the employer.
For firms moving more drafting work into AI-assisted processes, current document quality and the development of future reviewers should appear as separate measures. A gain in one does not establish a gain in the other.
