Tesler’s Law After Katsushika Hokusai
The wave has to land somewhere.
First, try thisThe wave represents the work that remains
Press and hold anywhere on the water. Try to make the big wave go away.
Tesler’s Law · LLMs and the workforce
Who gets the time saved, and who gets the extra work?
Larry Tesler described a core of complexity in an application that someone must handle: the user, the application developer, or the platform developer. This room borrows that idea to ask what happens when AI changes a job. Drafting may take less time, while checking and training need more attention. The wave stands for those remaining responsibilities. Tesler’s Law does not tell us how much labor a task will need or how many jobs there will be.
In Generative AI at Work, a study of 5,172 customer support agents found 15% more issues resolved per hour on average with AI assistance. Newer and less-skilled agents benefited most. The most experienced and highest-skilled agents saw small improvements in speed and small declines in quality. The researchers also found evidence of learning, so AI assistance need not prevent people from developing skills. Read the study.
METR’s early-2025 randomized trial followed 16 experienced open-source developers across 246 tasks in repositories they knew well. Tasks took 19% longer when AI tools were allowed, even though developers estimated afterward that AI had made them about 20% faster. That result applies to the tools and setting tested. Read the 2025 study. In a February 2026 follow-up, METR reported signs of faster work with later tools, but changes in which developers and tasks entered the study made the estimated gain unreliable.
The pay evidence is narrower. Humlum and Vestergaard linked surveys of Danish workers in AI-exposed occupations to administrative records. During the first two years after ChatGPT’s launch, they found no statistically significant average effect on earnings or recorded hours; their estimates ruled out effects larger than 2%. That does not mean every worker’s pay stayed the same, or tell us what will happen elsewhere. Read the paper.
There are concerns about entry-level hiring. In their August 2026 update using U.S. ADP payroll data through June 2026, Brynjolfsson, Chandar, and Chen found employment among 22–25-year-olds in AI-exposed occupations 19% below the level implied by growth among less-exposed peers. Reduced hiring accounted for most of the gap. They found no evidence of widespread displacement across the economy and describe the findings as early indicators. The comparison does not establish that AI caused the gap. Read the paper.
Using a chatbot has not excused mistakes in court. In Mata v. Avianca (2023), a New York federal court sanctioned two lawyers and their firm for submitting fabricated cases and standing by them after the court questioned them. In Moffatt v. Air Canada (2024), British Columbia’s Civil Resolution Tribunal ordered the airline to pay damages for misleading fare advice from its chatbot. These decisions concern specific duties and facts; they do not make every employee who signs off on AI output personally liable.
The argument of this room is that employers should budget for the checking and training their AI rollout needs, and involve workers in deciding how the job changes. Providers have work to do too, such as making sources easier to inspect. The ILO’s 2025 assessment expects transformation to be more common than full replacement of jobs exposed to generative AI. It measures potential exposure, so it cannot tell us how many jobs have actually been lost.
The five scenarios are invented examples. Each boat shows one possible response, and its label reflects how that response handles the problem. The wave’s size stays fixed for the interaction; actual workloads can grow or shrink.
Read more: Tesler’s original formulation · ILO: Generative AI and jobs · WGA: AI protections in the 2023 agreement
