Case Study
Consistent Writing Feedback at UC San Diego
How the Humanities Program uses ibl.ai inside Canvas to give roughly 1,500 students a quarter the same standard of analytical feedback — drawn from a pre-approved comment bank, with instructors auditing every suggestion.
Most educational AI helps students draft. This does the opposite: it helps instructors and TAs grade consistently. The agent reads student writing the way a TA would, then proposes two comments per section — drawn only from a bank the program approved in advance.
~1,500
Students served each quarter
Rubric-aligned
Feedback from an approved bank
Human final say
Every suggestion is audited
FERPA-aligned
Private cloud tenant, UCSD-owned
Dr. Kristina Markman
Associate Director, Humanities Program, UC San Diego
“Partnering with ibl.ai has been essential to bringing my vision for an AI writing agent to life in the Humanities Program at UC San Diego.
ibl.ai recognizes that a one-size-fits-all model is not conducive to the diverse needs of higher education. They provide truly personalized solutions by listening closely to instructors and designing tools that directly strengthen program-specific student learning outcomes.
From refining prompts to adapting AI outputs to my standards, ibl.ai’s collaborative approach, responsiveness, and willingness to meet as needed have made them true partners in innovation. Their ability to align technology with the Humanities Program’s pedagogical goals has been a driving force behind my project’s success.”
Same assignment, different section, different standard
The Humanities Program is the cornerstone of UCSD’s liberal arts and writing requirement, giving students an interdisciplinary introduction to the history of ideas through evidence-based, text-centered analysis. At that scale, across many sections and many instructors, the risk is not bad feedback — it is inconsistent feedback.
Grading consistency
A student's experience shouldn't depend on which section they were assigned.
Feedback quality at volume
Maintaining depth across hundreds of papers per quarter is a genuine workload problem.
Room for individual learning
Standardization must not flatten what makes a particular student's argument worth engaging.
Faculty time
Hours spent on mechanical scoring are hours not spent on ideas or one-on-one teaching.
A closed comment bank, not open generation
The agent does not invent feedback. It selects from comments the program wrote and approved, picking the two most relevant for each section of a student’s work and explaining why — so the instructor can choose rather than proofread.
Example approved comment
“A topic sentence should present an arguable claim, not just summarize an observation. Focus on stating what the paragraph will prove about the text.”
Instructors load the context
Prompts, rubrics, and exemplar papers are uploaded into a private faculty workspace inside Canvas — the agent works from the program's own materials.
The agent proposes, with justification
For each section of student writing it selects two comments from the approved bank and gives a short rationale for each, so the instructor can judge the fit quickly.
A human releases the feedback
Every suggestion is audited before it reaches a student. Final judgment stays with the instructional staff — the agent never has the last word.
Next to SpeedGrader, not in another tab
The agent integrates with Canvas over LTI 1.3, so there are no extra logins or browser tabs. An instructor toggles it on for a given assignment and it appears alongside SpeedGrader — a familiar workflow, which is most of what drives adoption.
Typical AI grading tool
- Generates novel feedback the program never approved
- Separate login and a second window
- Student work leaves the institution's environment
- Vendor holds the configuration and the data
ibl.ai at UCSD
- Selects only from the program's approved comment bank
- Embedded in Canvas via LTI 1.3 — no extra login
- Data stays in UCSD's existing learning environment
- UCSD owns the data and the model configuration
Student work is the most sensitive data on campus
Every interaction is logged for transparency, and UCSD keeps full ownership of both the data and the model configuration in its private cloud tenant — an approach that aligns with FERPA and keeps the exit door open.
Logged interactions
Every agent-student interaction is recorded, so the program can review what was said and why.
Private cloud tenant
Data and configuration live in UCSD's own tenant rather than a shared vendor environment.
Minimal lock-in
Owning the configuration means the program can move, extend, or replace the setup on its own terms.
The program set out to give instructors a common framework. Four things follow from it:
Assessment aligns with curriculum
Feedback is tied to the outcomes the program actually claims to teach, not to an individual grader's habits.
Feedback is democratized
Students encounter consistent expectations regardless of which instructor or section they are assigned.
Faculty focus shifts
Time moves from mechanical scoring toward thoughtful engagement with student writing and ideas.
The experience gets more personal, not less
Freed time becomes one-on-one tutoring and deeper engagement — the human part of writing instruction.
Bring consistent feedback to your writing program
Approved comment banks, rubric-aligned analysis, and human review at the end — embedded in the LMS you already use, on infrastructure you own.