Nominal News is an economics newsletter written by a PhD Economist that translates the latest economic research into clear, policy‑relevant insights on current issues. Join 4,000 readers to stay-up-to-date with Nominal News directly in your inbox:
AI continues to make news headlines – whether it is solving math problems or hacking companies. However, its benefit in the workplace is still up for debate. From a big-picture perspective, US productivity (how much output we produce per unit of input) measures provided by the Bureau of Labor Statistics (BLS) have not really increased1 in response to AI usage. But how about from a firm-perspective – a recent research paper looked at how AI use at work impacts employees – and the results have significant conclusions for how businesses and individuals should use AI.
An AI Experiment on Lawyers
Autor et al. (2026) ran a three month field experiment on 133 patent lawyers in the US. The lawyers were split into two groups – one set of lawyers (let’s call them the AI Group) was allowed to use a non-public AI tool as part of their daily work of drafting patent documents, while another group did not have access to the tool. Autor et al. then conducted two exams on these groups.
One test that was conducted twice – once after 10 days and once after 90 days – asked the lawyers in each group to draft a patent-related document. The AI Group could draft the document using the AI tool, while the non-AI Group could not use the tool. The second test, conducted after 90 days, had each group conduct a ‘redline’ exercise on a draft patent document.
In this redline exercise, lawyers in both groups were asked to review and comment on the draft document, without the use of AI. All the tests were evaluated by professional patent lawyers who provided a grade on the following 5 elements – “enforceability, accuracy, strategic ambiguity, completeness and clarity”. Each test was graded by two lawyers.
So what happened?
Better Work Product, But Mixed Experience Gain
The initial results were interesting:
On average, lawyers that had access to the AI tool drafted better documents; but,
On average, they did not get better at reviewing (redline exercise).
For the document drafting test, the output quality, on average, was at the 50th percentile for non-AI Group versus, approximately, the 65th percentile for the AI Group (so an average lawyer became an above average lawyer with the use of AI).
However, the average outcomes masked significant variance:
Junior lawyers (less than 7 years experience) benefited from the AI tool when drafting, but;
On the redline exercise, where AI was not allowed in both groups, junior lawyers in the AI group, showed no improvement in reviewing over junior lawyers in the non-AI group;
Senior lawyers (> 7 years experience) in the two groups showed limited difference in quality of drafting, but;
Senior lawyers in the AI group were better at redlining and outperformed their non-AI colleagues on the redline exercise.
Although AI allowed junior lawyers to create better drafts, it did not necessarily mean that they gained skills. On the other hand, senior lawyers, thanks to AI, appear to have gained experience.
More Equipped Individuals Gain More
Autor et al. investigated what was driving this dispersion. Two things stood out:
Junior performance in the AI group became more dispersed – among the junior lawyers who used AI, more junior lawyers got very low scores (bottom 20th percentile), but also more got above average scores (60th to 80th percentile) on the redlining exercise with fewer average scores than the non-AI group;
Senior lawyers in the AI group performed better on the redlining exercise by focusing on more substantive issues in the patent document review – this included focusing on maximizing a patent’s commercial scope, addressing legal liabilities and referencing legal principles as part of their review, instead of stylistic and administrative adjustments.
Senior lawyers appear to have used their prior experience to focus on more critical issues, instead of more trivial issues, which junior lawyers in general tend to focus on. Senior lawyers were also likely to fully re-write sections to improve the patent document strength, whereas junior lawyers only jotted down broad annotations. Overall, this suggested that prior experience and knowledge mattered on who benefited from AI.
My Takeaways – How to Implement AI at Work
The Autor et al. paper is a rare example of a study that explicitly aimed to measure the influence of AI in a non-routine job. Given the varied outcomes, AI can be a tool that, if used well, can enhance a business, but, if used poorly, can perhaps weaken knowledge acquisition.
Junior employees may create better quality outputs thanks to AI, but, as currently implemented, AI does not necessarily increase the skills and knowledge of all junior employees. In fact, AI might mask low performance or a poor employee-job match. Some junior lawyers drafted higher quality documents because of AI, but may actually have become worse lawyers due to its use (note: whether AI ‘caused’ these junior lawyers to perform worse is not demonstrated).
Seniors, who already have significant knowledge and experience may benefit from AI, as they are able to continue furthering their knowledge via the regular use of this tool. This shows that certain foundational knowledge is critical to get the most of AI. Naturally, senior employees are already pre-selected in a way – you likely wouldn’t become a senior employee unless you demonstrated a certain skill level in the first place.
So, for me, the takeaways from this paper are:
Allow senior employees to use AI for work;
Ask senior employees to share how they use AI with junior employees;
Perform knowledge/skill assessments for junior employees to incentivize the gaining of skills, rather than just using AI.
The last point is perhaps the most important – the Autor et al. study shows that there is a difference between doing and understanding a job. With the rise of AI use in the workplace, really understanding the job appears to be an important skill differentiator.
What You May Have Missed From Nominal News
On September 16 – the US Federal Reserve hiked interest rates. Find out why this happened, why the Federal Reserve is signalling further hikes and what I think will be the trajectory of inflation.
As this was a paid article – I am running a 40% off yearly subscriptions till the end of the week. If you’d like to see what is available at the paid tier, I have unlocked the below link:
A Hawkish Pivot - The Federal Reserve Raises Interest Rates and Signals Further Hikes
Although Kevin Warsh said little, the FOMC statement and dot plot suggest that the Federal Reserve is concerned about inflation and the threat of unanchored inflation expectations.
If you would like to support us in reaching our subscriber goal of 7,000 subscribers, please consider sharing this article and pressing the like❤️ button at top or bottom of this article!
Perhaps, surprisingly, there may have been a slow down in productivity, with productivity growth at 0.8% and 1.4% annualized for the first and second quarters of 2026, way below the long run average of 2.1% growth.
Anthropic actually had a similar study with junior software engineers: https://www.anthropic.com/research/AI-assistance-coding-skills
AI users produced the work a bit faster, but scored 17% lower on a test of what they had just learned. And a similar point to this other study you mention, where folks who used AI as a teacher vs just delegating did much better (though this is just the expectation)
At my work in software development, senior developers are an army with AI because of architectural knowledge and understanding. Not all Jr. Developers will become Sr. Ai cant help a Jr. Who just doesnt get it. This is why the study came out the way that it did. Pruning has occurred at the Sr. Level ... those that get it...Will do better with better tools. I think the study and my personal experience is AI is going to be transformative if used wisely.