AWS Public Sector Blog
Transforming classroom conversations: Cornell University’s AI-powered Socratic Chat on AWS
In his small graduate seminars, Cornell University Associate Professor of Earth and Atmospheric Sciences, Toby Ault, could probe a student’s understanding: “You mentioned greenhouse gases. What kind? What can you tell me about CO2 throughout Earth’s history?” With each exchange, he’d identify exactly where that student’s knowledge started to falter.
But in his more than 300-person climate and energy lecture? Impossible.
Working with Amazon Web Services (AWS), Cornell built Socratic Chat, an artificial intelligence (AI) assistant that delivers personalized Socratic teaching at scale. By fall 2024, over 500 students across multiple courses were using the tool, spending 10-15 minutes per assignment, working through progressively harder questions based on Bloom’s Taxonomy—freeing Ault and other professors to focus on in-depth instruction.
A six-year journey from campus workshop to working solution
The seeds for Socratic Chat were planted in 2018 at an AWS workshop on Cornell’s campus. AWS showcased their capabilities to Cornell IT staff, including Marty Sullivan, cloud/DevOps engineer. Ault attended as one of the few faculty members present, and one demonstration caught his attention: a project using logic trees to guide incoming students through their first week on campus.
“I thought the whole ecosystem of AWS was fantastic,” Ault recalled. Could they build something similar to guide students through hierarchies of knowledge? Sullivan and Ault briefly explored the concept together, but the technology wasn’t sophisticated enough. They shelved the idea.
Years later, the 2020 pandemic revealed what technology could enable. During that spring, while teaching remotely, Ault posed questions to all students through Zoom’s chat feature. Teaching assistants monitored responses in real time, messaging him when patterns emerged. “The TAs would text me and say, ‘It looks like from that last question, people are confused about XYZ,’” Ault explained.
When everyone returned to campus in the fall of 2022, Ault wanted to recreate that engagement. Late in 2022, he uploaded final project instructions to ChatGPT. The large language model (LLM) returned quality work. “That was the click for me,” Ault said. “How am I going to adapt to this technology to promote learning in my classroom?”
In 2023, following renewed interest, Sullivan and Ault partnered with their AWS account manager from the AWS Higher Education team to submit a proposal to Cornell’s AI innovation competition.
Collaborating with AWS to build a proof of concept
In spring 2024, Cornell worked with the AWS Professional Services team to build a proof of concept. Later, AWS Generative AI Innovation Center guided a second phase. Before coding began, Sullivan and Ault spent several sessions defining requirements.
“Toby and I sat down and hashed out what he wanted from his perspective in the form of user stories and acceptance criteria,” Sullivan explained. “That really helped me as an engineer to know what I’m building.”
These planning sessions benefited from their unique working relationship. Not only is Sullivan a Cornell IT engineer, but he’s also a PhD student in Ault’s atmospheric science program. This enables the two to collaborate in ways that bring together institutional resources with deep pedagogical understanding.
The AWS team designed the proof of concept using Amazon Q Business, creating a structured approach where instructors preapproved questions at different complexity levels based on Bloom’s Taxonomy. The proof of concept gained support from Cornell’s Center for Teaching Innovation, the Museum of the Earth, and other partners.
With AWS’s architectural guidance and Cornell’s existing AWS agreement enabling rapid provisioning, Sullivan and a small team built the production version in under eight months, from project start to completion. The system utilizes Claude by Anthropic in Amazon Bedrock for all LLM functionality, combined with open-source frameworks such as Streamlit, Chainlit, and FastAPI.
“The majority of my time was doing the actual integration into Canvas, the learning management system we use,” Sullivan said. “If we want other faculty to use this, it needs to be in the tool that they’re already using.”
The team also knew ease of use would determine adoption. “Another faculty member said, ‘If this takes me longer than 15 minutes, I’m not going to do it,’” Sullivan recalled. He designed the instructor interface to meet that threshold: Instructors upload lecture content and approve topics, then the system automatically generates the conversational flow.
Delivering personalized instruction that mirrors the professor’s teaching style
Socratic Chat structures learning around Bloom’s Taxonomy, which categorizes understanding from basic recall to advanced creation. As noted, instructors upload content—Ault uses lecture transcripts recorded via Zoom, for example—and approve topics at different complexity levels.
Students then access Socratic Chat through Canvas assignments during or after class. They start at the “understand” level. If they struggle, the assistant rephrases questions or offers hints. Correct answers move them to the next level: apply, analyze, evaluate, then create. A progress bar, added after feedback from Cornell’s Center for Teaching Innovation, shows advancement.
“The progress bar was a great addition,” Sullivan said. “If you don’t have some sort of goal that students are working toward, they won’t know when they’re going to be done.” The feature had an immediate positive impact, with students comparing it to gamification.
Additionally, the assistant mirrors Ault’s teaching style because it’s trained on his lecture transcripts. “I like that it expresses the questions the way that I would want them to be expressed, and it offers helpful suggestions,” Ault said. Students engage more readily when the AI’s style feels familiar, maintaining the connection to their instructor even during independent work.
Reaching hundreds of students with adaptive instruction
Cornell deployed Socratic Chat across five courses in fall 2024, reaching over 500 students. Students typically spent 10-15 minutes engaging per assignment, motivated by the conversational flow and the desire to “level up.” Some spent much longer. “Some students spent an enormous amount of time on it because they did not want to be defeated by the robot,” Ault noted.
When Ault tested the tool on his lecture about Carl Sagan’s “Pale Blue Dot,” it prompted him to the “create” level—the highest in Bloom’s Taxonomy—asking him to design an advertising campaign centered on sustainability. “That was the moment we discovered that we had developed something pretty powerful,” Ault said. “It forced me to think more deeply about how to use my own lecture.”
The tool also automatically generates quiz questions. This gives Ault visibility into where students struggle and helps identify topics needing more attention.
Making time for the human connections that drive student success
For Ault, Socratic Chat doesn’t replace teaching. It enhances it. “There’s an enormous amount of prejudice against AI in the classroom right now,” he acknowledged. “But every choice we’ve made is designed around promoting learning by humans.”
By handling structured learning conversations at scale, the tool gives Ault more time for what he believes matters most: face-to-face interactions that deepen students’ learning and success. “It buys me more time to do the in-person lecturing, discussion in class, after class, and office hours,” Ault said. “It frees up the human interaction time.”
To address integrity, Ault randomly asks students to recall and elaborate on chat responses in person. “If they’ve really done it, they’ll have no trouble recalling and reflecting on what they meant,” he explained.
Scaling Socratic teaching across Cornell and beyond
After gathering feedback throughout fall 2024, Cornell is expanding Socratic Chat to additional courses and sharing the tool through collaborations with the Museum of the Earth and the Paleontological Research Institution. Sullivan continues to refine the platform, planning to update it to the latest Claude models through Amazon Bedrock.
For institutions facing similar challenges, Cornell’s approach demonstrates that when every design choice focuses on promoting human learning rather than replacing human teaching, AI becomes a powerful ally in scaling personalized instruction without sacrificing the connections that inspire learning.
Learn how AWS helps institutions build, deploy, and scale AI solutions for education. Contact AWS today.