Halfway through a debate round in Washington, Aria-Vue Daugherty became aware that her vision wasn’t working properly. She had been struck earlier that morning by a black SUV. Nevertheless, she completed the round, went on her own to the George Washington University Hospital emergency room, and spent seven and a half hours there.
By Monday, she was back in Cambridge, sitting in lecture halls that were difficult for her to comprehend and heavily relying on Harvard’s Peer Notetaker program, a tiny but unyielding organization.
| Topic Snapshot | Details |
|---|---|
| Subject | Harvard’s Peer Notetaker Program under the Disability Access Office |
| Institution | Harvard University, Cambridge, Massachusetts |
| Stipend Per Course | $600 |
| Notetakers Hired (2025–26) | 84 students |
| Students Receiving Accommodations | Around 50 each year |
| Common AI Alternatives | Claude, NotebookLM, Otter.ai |
| Reported AI Hallucination Rate | Up to 79% in newer systems |
| Key Concern | Accuracy, context, and human interpretation |
| Notable Voice | Aria-Vue L. Daugherty ’29, who relied on the program after a concussion |
| Counterview | Lia T. Zheng ’27, president of Harvard’s undergraduate AI society |
It’s the kind of service that doesn’t often make news. Each year, about fifty students are eligible to receive the $600 stipend that the Disability Access Office pays to students who take notes in eligible courses. Eighty-four graduate and undergraduate students were hired for the position this year. For a quiet program, use quiet numbers. However, since every student on campus now has access to a free AI tool that says it can accomplish the same task more quickly, more affordably, and without requiring a stipend, the discussion surrounding it is suddenly less quiet.
Some students believe that AI may already be adequate. The head of Harvard’s undergraduate AI society, Lia Zheng, is quite direct about it: an AI tool can probably match a human notetaker in large lecture courses where the material is standardized and the instructor is essentially acting out a script. Perhaps even surpass one. The math is difficult to dispute. Approximately 150 words are spoken per minute by a professor. About thirty are written by a student. AI never gets tired, never skips class, and never forgets to email the file.
However, the picture becomes more hazy when you speak with the students who genuinely rely on the program. When it comes to AI, Daugherty acknowledges that she is somewhat of an extremist. She fears hallucinations, which are caused by large language models’ peculiar propensity to create things that seem plausible but are not. According to a study conducted last year, hallucinations were present in 79% of the outputs produced by more recent AI systems. When you have a concussion and are studying for a midterm, that is not the type of error rate you want.
Due to rheumatoid arthritis, Eman Seyal has been utilizing the program since his freshman year and has experimented with tools such as Otter.ai. Although the transcripts were technically correct, she thought they were essentially worthless. They failed to understand how her lecturer presented a concept, were unsure of which equation was important, and were unsure of what to do when a graph appeared on the board. According to her, peer notes are focused on the actual class. You get a wall of text from AI. Hearing every word and knowing which words mattered are two different things.
More of the story is revealed in the smaller moments. A student named Michael Isayan was pulled aside by Daugherty following their writing seminar a few weeks into the course. He didn’t know if anyone was actually reading the notes he was taking for the DAO. She informed him that she was. A transcription tool cannot create an exchange like that, where a stranger thanks another stranger for the small invisible labor of paying attention in a room.
Of course, AI might be able to bridge that gap. Otter’s auto-summaries are getting better every semester, and tools like NotebookLM can already transform a recording into a study guide. It is difficult to ignore the economic reasoning: $600 per course, 84 students, multiplied by the number of Ivy League schools offering comparable programs. That math is being done somewhere by an administrator.
However, for the time being at least, the program endures because it provides an alternative to AI. Mostly trust. the understanding that an actual person sat in the same room, heard the same joke, saw the same diagram appear on the board, and determined what was important. It’s genuinely unclear if that will be sufficient to keep peer notetakers employed in five years. As we watch this develop, it’s difficult not to wonder if we’re witnessing the final years of a small, specific type of campus job or the start of an obstinate defense of why some things still require a person in the chair.
