Test AI annotations on student problem sets
One month after the rollout of Dartmouth Claude to campus, the College lacks policies on faculty AI use. Task: One month after the rollout of Dartmouth Claude to campus, the College lacks policies on faculty AI use. On April 30, The Dartmouth learned of a public project posted by chemistry professor Paul Robustelli on the Dartmouth Claude enterprise portal. The project, titled “PS3 Grading,” contained problem sets by 11 different students from Robustelli’s CHEM 76: Physical Chemistry II class, as well as the students’ personal identifying information. It could be accessed by any member of the Dartmouth Claude enterprise group, which is open to all campus community members. The Claude project was taken down by Robustelli on May 2 following The Dartmouth’s request for comment. Files in the project contained CHEM 76 students’ first names, last names and problem set responses. Some students’ work was marked as “LATE” in file names. In a statement to The Dartmouth, Robustelli wrote that he has “never used Claude or any AI tool to grade student work” and instead uploaded student problem sets to his PS3 Grading project as part of “one limited test.” “I ran student assignments through Claude to see if its annotations might someday serve as a reference for manual grading,” Robustelli wrote. “It played no role in any student’s grade, in this course or any other.” Robustelli also used his Dartmouth Claude enterprise to create a separate public project, titled “Na Emission Lab Code,” which was also taken down. “I want to teach my students how to perform the analyses in a physical chemistry laboratory [and] build a didactic scaffold of python code … to enable the students to complete the lab assignment with Python instead of Excel,” he wrote in his project instructions to Claude. Robustelli added that he believed the projects were private. “As soon as this was brought to my attention, the issue was corrected immediately,” Robustelli wrote in his statement. The default visibility setting for new projects on Dartmouth Claude enterprise accounts is private. Users opt in to creating a public project by clicking an option labeled “Dartmouth: Everyone in your organization can view and use this project.” The College first announced its AI partnership with Anthropic in December. In a campus-wide email from the Information, Technology and Consulting office sent on March 30, the College announced that all campus community members would have free access to Claude for Education, which combines premium chat features with the ability to collaborate in projects across the Dartmouth enterprise group. College spokesperson Jana Barnello wrote in a statement to The Dartmouth that the College is “reinforcing guidance around project visibility settings” as part of its “ongoing support” to community members in the use of Dartmouth Claude. “Dartmouth offers training and support through ITC, DCAL [Dartmouth Center for the Advancement of Learning] and the Libraries to help faculty, staff and students use these tools effectively and securely,” Barnello wrote. Anthropic did not respond to requests for comment by the time of publication. CHEM 76 student Timothy Bonis ’26 said he felt Robustelli’s use of AI was “incongruous” with his “anti-AI posture.” “[Robustelli’s] word has been, as we agreed in the first class, that we will not use AI in the problem sets and we will not use AI in the labs,” he said. A copy of the CHEM 76 course syllabus obtained and reviewed by The Dartmouth includes an “expectation that problem set solutions will not be obtained from generative AI.” “All work that is submitted to be graded must be your own,” the syllabus states. Bonis added that he believes there should be “reasonably definitive standards” for faculty AI use within departments and that AI-generated course materials are “unlikely to be pedagogically effective.” Chemistry department chair Ekaterina Pletneva did not respond to requests for comment by the time of publication. There is no policy on faculty AI use listed on the department’s website as of the time of publication. “Faculty decide when and how to use AI in teaching and learning,” Barnello wrote. Another CHEM 76 student, who wished to remain anonymous, wrote in a statement that they do not believe Robustelli’s grading method is “unfair or inaccurate.” “He certainly goes through and checks manually as he gives us specific feedback and marks up our homework,” the student wrote.; tools: Claude. Human role: The original post does not say how a person stayed involved.; limitations: Prerequisites include access to a Claude enterprise account and strict adherence to data privacy regulations. Key constraints: Default project visibility is private, but users must actively avoid opting into public sharing. Operating cost includes potential reputational damage and administrative overhead for incident response. Human approvals are critical for any AI-assisted grading to ensure fairness and compliance with course syllabi..
Tools
Published on Agent Work Atlas:
Overview
One month after the rollout of Dartmouth Claude to campus, the College lacks policies on faculty AI use.
Workflow
One month after the rollout of Dartmouth Claude to campus, the College lacks policies on faculty AI use. On April 30, The Dartmouth learned of a public project posted by chemistry professor Paul Robustelli on the Dartmouth Claude enterprise portal. The project, titled “PS3 Grading,” contained problem sets by 11 different students from Robustelli’s CHEM 76: Physical Chemistry II class, as well as the students’ personal identifying information. It could be accessed by any member of the Dartmouth Claude enterprise group, which is open to all campus community members. The Claude project was taken down by Robustelli on May 2 following The Dartmouth’s request for comment. Files in the project contained CHEM 76 students’ first names, last names and problem set responses. Some students’ work was marked as “LATE” in file names. In a statement to The Dartmouth, Robustelli wrote that he has “never used Claude or any AI tool to grade student work” and instead uploaded student problem sets to his PS3 Grading project as part of “one limited test.” “I ran student assignments through Claude to see if its annotations might someday serve as a reference for manual grading,” Robustelli wrote. “It played no role in any student’s grade, in this course or any other.” Robustelli also used his Dartmouth Claude enterprise to create a separate public project, titled “Na Emission Lab Code,” which was also taken down. “I want to teach my students how to perform the analyses in a physical chemistry laboratory [and] build a didactic scaffold of python code … to enable the students to complete the lab assignment with Python instead of Excel,” he wrote in his project instructions to Claude. Robustelli added that he believed the projects were private. “As soon as this was brought to my attention, the issue was corrected immediately,” Robustelli wrote in his statement. The default visibility setting for new projects on Dartmouth Claude enterprise accounts is private. Users opt in to creating a public project by clicking an option labeled “Dartmouth: Everyone in your organization can view and use this project.” The College first announced its AI partnership with Anthropic in December. In a campus-wide email from the Information, Technology and Consulting office sent on March 30, the College announced that all campus community members would have free access to Claude for Education, which combines premium chat features with the ability to collaborate in projects across the Dartmouth enterprise group. College spokesperson Jana Barnello wrote in a statement to The Dartmouth that the College is “reinforcing guidance around project visibility settings” as part of its “ongoing support” to community members in the use of Dartmouth Claude. “Dartmouth offers training and support through ITC, DCAL [Dartmouth Center for the Advancement of Learning] and the Libraries to help faculty, staff and students use these tools effectively and securely,” Barnello wrote. Anthropic did not respond to requests for comment by the time of publication. CHEM 76 student Timothy Bonis ’26 said he felt Robustelli’s use of AI was “incongruous” with his “anti-AI posture.” “[Robustelli’s] word has been, as we agreed in the first class, that we will not use AI in the problem sets and we will not use AI in the labs,” he said. A copy of the CHEM 76 course syllabus obtained and reviewed by The Dartmouth includes an “expectation that problem set solutions will not be obtained from generative AI.” “All work that is submitted to be graded must be your own,” the syllabus states. Bonis added that he believes there should be “reasonably definitive standards” for faculty AI use within departments and that AI-generated course materials are “unlikely to be pedagogically effective.” Chemistry department chair Ekaterina Pletneva did not respond to requests for comment by the time of publication. There is no policy on faculty AI use listed on the department’s website as of the time of publication. “Faculty decide when and how to use AI in teaching and learning,” Barnello wrote. Another CHEM 76 student, who wished to remain anonymous, wrote in a statement that they do not believe Robustelli’s grading method is “unfair or inaccurate.” “He certainly goes through and checks manually as he gives us specific feedback and marks up our homework,” the student wrote.
Reported result
The actual outcome was a privacy incident where a project containing 11 students' personal identifying information and problem sets was accidentally made public on the campus enterprise portal. The project was taken down after media inquiry. No measured success metrics for the grading test were reported, and the incident highlighted the lack of clear faculty AI policies.
Human role
The original post does not say how a person stayed involved.
Limitations
Prerequisites include access to a Claude enterprise account and strict adherence to data privacy regulations. Key constraints: Default project visibility is private, but users must actively avoid opting into public sharing. Operating cost includes potential reputational damage and administrative overhead for incident response. Human approvals are critical for any AI-assisted grading to ensure fairness and compliance with course syllabi.
Try it
I am a chemistry professor at a university with a Claude enterprise account. I want to test if Claude's annotations can help me grade problem sets faster, but I must ensure student privacy and academic integrity. Here are my inputs: [replaceable: path to anonymized student problem set files], [replaceable: course grading rubric], [replaceable: target output folder]. Please execute these steps: 1. Create a new project in Claude with visibility set to 'Private' only. 2. Upload one sample anonymized problem set. 3. Prompt Claude to generate annotations based on the rubric, explicitly instructing it to provide feedback for my review, not a final grade. 4. Output the AI's annotations and a summary of potential privacy risks. 5. Stop and wait for my confirmation before processing any additional files. Do not share this project with anyone. If you cannot verify the privacy setting, abort the task. Measure success by: zero personal identifiers in the uploaded files, correct privacy setting, and annotations that are clearly marked as 'for instructor review only'.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.