Sam Thayer, Director of the Office of Accessibility Resources

15 April 2025

Listen and subscribe to The Year of Curiosity: AI podcast on Apple Podcasts or Podbean.

Highlights from talk with Sam Thayer, Carleton College’s Director of the Office of Accessibility Resources and 2010 alum, focused on finding reliable AI solutions that we can trust.

Curiosity

George is curious about how listeners feel about this podcast, so he is urging folks to leave a review on their podcast platform of choice. He is also asking for suggestions for the show, including future guests.

Jennifer was inspired by a talk Laura Goering, Professor in Russian, gave at Carleton about the Five Stages of Grief: AI and Literary Translation to think again about how Large Language Models work. The next statistically probably word is sometimes sound and at other times makes no sense in a particular context. This reminded her of new research in neuroscience where our brains associate things together that are not really associated but somehow overlap. For example, when Jennifer talks to George while holding a warm drink, it will lead her to think of him as a person with a warm personality.

George connects this to our need to tell stories and use metaphors to connect something new and unknown to something we know. We use a lot of different metaphors when talking about AI – stochastic parrots, writing, thinking. And do we need to make these connections, for example, do we need to tell the story about AI work in relation to how human brains work? Kieran Barker, a current Carleton cognitive science major, started this conversation.

Sam Thayer, Director of Office of Accessibility Resources

Sam is curious about the order of an AI’s responses. How does it determine the order? Sam’s five-year-old asks a lot of questions, such as how big is a blue whale and what lives in the deepest part of the deep sea. Sam will ask Google for help, and she has noticed that the AI responses have been getting more robust, providing different perspectives with support. But how does Google decide which perspective get the first position? The team speculated that this may be a Google-specific strategy to be sensitive to different perspectives (even in situations where there is only one answer).

AI Origin Story

Sam’s AI origin story is from a couple of years ago when ChatGPT was becoming mainstream and widely available. Everyone was talking about ChatGPT, few people could access it because it was so bogged down. Sam had brunch with friends at a local restaurant, and everyone was very excited about their use of ChatGPT. For example, they were using it to plan vacations, including an itinerary or a meal plan. And then she saw more communications about how to use generative AI for work tasks.

Some Highlights

Sam, in the Office of Accessibility Resources, oversees the accommodation process for students with disabilities. Her office also works on different programming for professional development in the Carleton community about a wide range of topics covering disability and accessibility. The office collaborates with other departments to look at intersectional identities and how to support students in the classroom.

AI has impacted this kind of work, first, by overwhelming the news in Sam’s line of work. The listservs are inundated with AI-related topics, how AI tools are used to improve accessibility. Many companies reach out to try and sell these new tools.

One obvious space where AI has made a positive contribution is in how automated captioning has improved. Her office is getting fewer requests to help with captioning. The tools are making it easier to get immediate captions or to upload captions, which means more people can take advantage of this feature, not just students who have an explicit accommodation for a hearing-related disability.

Jennifer is glad that captions have significantly improved although she misses the time when the captions were occasionally producing amusing statements. For example, when her colleague Debbie Walser-Kunz was teaching mitosis, the caption would talk about my toesies.

Another feature that has helped with accommodations is the improvement of automatic voices. AI voices sound far less robotic, with a more natural rhythm and cadence, sometimes even with multiple voices. This makes listening to longer texts more enjoyable. George mentioned a colleague who is uploading a textbook to Notebook LM so that his students can use, among other things, the podcast option. Students who struggle with reading find this option incredibly useful; George is on the fence how he feels about listening as an equivalent to reading.

George, with many other faculty, is concerned about what generative AI may do to students’ abilities and willingness to read everything assigned to them. How can we encourage students to use generative AI in ways that makes the reading more equitable rather than avoiding reading materials? Sam is hoping that using Universal Design for Learning principles to empower all students to learn is one key strategy. She is wondering if faculty are seeing differences in understanding or whether they are considering assessing understanding differently. Is there a disparity or discrepancy between students who are listening to the podcast version of a reading, students who are listening to the reading version, and students who are reading without listening?

Reading scientific papers, students observed that they need to engage with these publications at different levels, culminating with reading as a scholar of a particular discipline. Teaching students how to read that way and engage with the text, drill down and read repeatedly is a key task of teaching them how to become a scholar of their discipline. Listening to a podcast is not the same but it may be a good starting point to decide whether you want to continue with this line of inquiry. Students are also not going to admit using AI with their readings unless a faculty member models that it is possible to do so – and when to do it.

Sam sees that especially for students who did not claim accommodations during high school, there is a potential stigma to start using them now. It is important to show these students that these tools are amazing and helpful. Occasionally, a student will pull back from using a tool and not self-advocate if a faculty member pushes back on providing accommodations. Lecture capture tends to fall into this category. A lot of faculty may have a no recording policy – for good reasons. However, if lecture recording is the approved accommodation, the students need to advocate for their rights and have at least a conversation with the instructor.

Carleton does not have an institutional license for note-taking software. At other institutions, Sam’s colleagues are moving away from peer note takers to note-taking software applications like Glean and Otter. Students will manage these applications themselves.

Even though we are offered many new tools or tools now enhanced with AI, Sam is recognizing that a lot of these tools are not as great as they say they are. Sam is willing to try out new things and is repeatedly disappointed. For example, she has been looking for a note-taking software because the current system of finding a peer note taker is time intensive. The peer needs to be in the class and needs to commit to volunteer – all of this takes a lot of time. The student needing the accommodation needs to reach out to the peer tutor if and when the notes are not sufficient. A reliable note-taking software would help greatly. However, one of Sam’s trials was very disappointing especially as it promised a lot of useful features and did not deliver. This just shows how important the trials are because the software will be essential to a particular group of students to be successful.

It feels like the current AI tools are essentially using the users as beta testers to improve the product. One wonders how these implementations at other schools are actually working out for the students, in a classroom with noisy backgrounds, vs the optimal environment the software may have been tested in. These recording solutions will also not capture all important content – they may only focus on voices near by and not able to record student contributions from the back of the classroom.

Sam would like to offload all the scheduling of kid things that fill up quickly – swim lessons, summer camps, before school care – but she would really need to trust the AI. She would not want to offload planning of vacations.

Tools/Recommendations

Return to AI at Carleton Blog