Generative AI for Grading & Feedback

For instructors who want students to be able to get additional, formative feedback during practice or on draft work, genAI automated feedback systems could potentially be another tool for students to use as part of their learning process. Although we lack sufficient experimental data to confirm its utility, some initial studies have integrated genAI as a digital tutor, to give feedback in class and outside of class (Hobert & Berens, 2024; Li. et al. 2023). These studies were not seeking to replace instructor feedback on major deliverables, however, nor did they include a control group without genAI to test whether student improvement was due to genAI specifically.

For standardized tests with objective, finite responses, genAI has been found to have high consistency with human scorers. However, it has low consistency with exams that feature more subjective or nuanced answers, such as essays, or high levels of complexity such as assessments with video components (Coskun & Alper, 2024).

GenAI has moderate consistency in scoring quantitative work, such as problem sets and coding, but with some important caveats. One analysis of GPT-4’s grading ability on handwritten physics problems found decent agreement compared to human graders (Kortemeyer, 2023). However, the authors note that even after running every solution through the genAI tool 15 times, it still gave incorrect or misleading statements as part of its feedback, awarded more points than the human graders, and was less accurate at the low end of the grading scale. Similarly, another study compared ChatGPT-3.5 to human graders on Python assignments (Jukiewicz, 2024). Like the previous study, they ran each set of assignments through the tool 15 times to gauge consistency. Once again the agreement with human graders was good, but the genAI tool consistently awarded fewer points compared to human graders. In another study, however, the authors found even lower agreement between ChatGPT-3.5 and Bard on Python problem sets compared to human graders, with an overall accuracy rate of 50% (Estévez-Ayres et al. 2024). Kiesler et al. 2023 also noted that ChatGPT’s feedback on programming assignments for an introductory computer science course contained misinformation 66% of the time, which would be particularly problematic for novice learners, since they would likely be led astray.

For writing-based assessments more generally, studies have consistently found large discrepancies (i.e., low reliability) between genAI and human scorers. In one study comparing human and genAI scoring of formative feedback on writing, genAI only scored equivalently well on quantitative criteria-based feedback (Steiss et al. 2024; see also Jauhiainen et al. 2024). However, human feedback was of higher quality because it was more accurate and actionable and better prioritized and balanced. These four features are evidence-based characteristics of effective feedback (Hattie & Timperley, 2007). Additionally, genAI feedback quality varied based on essay score, with greater leniency demonstrated on lower quality writing, and overly strict responses on papers that scored above average (Wetzler et al., 2024). Research has also shown that genAI has difficulty with certain types of cognition, such as analogies and abstractions (Mitchell, 2021), as well as nuances and subtle variations in the subject material (Lazarus et al. 2024). 

 The genAI tool used also matters: one study found that ChatGPT 3.5 and 4o graded the same essays differently, despite being fine-tuned on the same prompt and rubrics. Both models were significantly different from human scorers (Wetzler et al. 2023). Even when the same genAI tool (ChatGPT-4) was tested 10 times with the same data (assignment prompt, rubric, and student work), the same grade was assigned only 68% of time (Jauhiainen et al., 2024).

Another study compared how well ChatGPT, Claude, and Bard could accurately score and provide feedback on both undergraduate and graduate writing samples (Fuller & Bixby, 2024). The rubrics and writing samples were run through ChatGPT and Claude five times each to assess consistency in scoring and feedback. Bard was not able to complete the tasks: although it scored the samples initially, on the second iteration it responded that it was not able to be used for assessment purposes. It was therefore omitted from data collection. As with the other studies, both ChatGPT and Claude had significant discrepancies in their scoring of the same writing samples using the same rubrics through multiple iterations, and also differed widely from the human grader.

It is important to note that for all of the above studies, the authors/instructors spent considerable time training the genAI on the assignments and fine-tuning the prompt engineering in order to create the ideal setup to hopefully get the desired grading and feedback output. For instructors or TAs who are not customizing the genAI tool nor running multiple iterations of scoring for quality assurance, the grading and/or feedback quality is likely to be much poorer. See #2 below for more details about customizing a genAI tool.

In addition to considering the efficacy of genAI feedback, one should also keep in mind the recipients of that feedback. Although in some cases students report being open to genAI-assessment (Braun et al. 2023, Zhang et al. 2024), many studies show that students prefer human feedback or a combination of human and AI feedback; they do not trust or feel satisfied receiving AI feedback alone (Er et al., 2024; Tossell et al., 2024; Chan et al., 2024). We also do not have sufficient research about the impact of receiving genAI feedback on student motivation and engagement in their learning. However, we know that students are more likely to engage with feedback if they feel that it is helpful and fair (Jonsson 2013, Harks et al. 2013, Panadero et al. 2023). Since genAI feedback in its current state is prone to errors, misinformation, and the feedback quality is inferior to a human grader’s, it is crucial that instructors or TAs review and amend its output to ensure students get feedback that they can trust and can use. Additionally, it is important that instructors and TAs are transparent with students about both evaluation criteria as well as how (and by whom) it is assigned fairly and accurately to students’ work (see also #6 below about how to communicate your policy and practices to students).

Initially grade a subsample of student work manually and do rubric norming (see #4 below for more about norming). Compare these norms against genAI outputs to assess the accuracy of the tool’s grading as well as the appropriateness and effectiveness of its feedback.

Check over AI-generated outputs and incorporate human edits/revisions. Do NOT assume that quality grading occurred or haphazardly share unreviewed genAI feedback with students. 

Vague or overly broad prompts often yield inconsistent results. Avoid subjective language. Instead, align the prompt with precise explanations of performance expectations for the assignment (e.g., a well-structured rubric). In the prompt for the genAI tool, it can help to include examples of student submissions of varying quality along with the feedback you would provide each of them.

 GenAI models typically perform better on widely discussed topics due to more extensive training data. Assignments on highly specialized or very recent subjects may result in less accurate or reliable outputs.

GenAI tools are primarily text-based, and thus their ability to provide accurate feedback diminishes significantly if submissions include visual elements such as diagrams and images.

1. Use a rubric. Randomly select 5-7 assignments to check your rubric and make sure it works. If necessary, adjust your rubric, put those 5-7 assignments back in the pile, and re-grade all of the assignments.
2. Work with the other TAs. Make sure your grading is consistent among everyone (this is usually called “grade norming”). Working together can help with motivation and provide a second set of eyes if you get stuck. You can also divide up labor among TAs to maximize grading/feedback consistency within an assessment item. For example, on an exam or homework assignment, assign TAs to grade all students for a subset of the questions/items, rather than assign TAs to grade all questions/items for a subset of the students.
3. Leverage educational technology. Depending on your course context, educational technology like Canvas or Gradescope may make your grading easier and more consistent.
4. Develop a “key” or "common comments” document (e.g., AWK = awkward phrase, “TS” = topic sentence missing or needs revising, etc.). This will save you time so you do not have to write out the same feedback every time.

• Grade unbiased by student names. Place Post-It notes over students’ names to ensure fair, unbiased grading, and then shuffle the assignments.
• Prevent grading drift. Go back and compare the first five assignments to the last five. Have your standards changed?
• Set a timer. This will ensure that you spend the same amount of time on each student, and will also help prevent burnout.
• Grade/provide feedback question-by-question. Grading one question at a time (e.g., question #1 for everyone, then question #2…) rather than student-by-student will help you stay in the same mental space. Gradescope can help with this!
• Provide specific group-level feedback. This strategy is an efficient way to provide feedback on common errors, and can be done via Canvas, email, or even verbally at the beginning of class.

1. Individuals must be logged in as instructed via CMU authenticated mechanisms. 
2. Not all tools listed on Computing Services site are FERPA compliant and are typically indicated as such (see CMU’s Google Workspace for Education webpage here showing “Core” vs “Additional” services). If you are uncertain about whether or not a tool you are using or want to use is meeting these privacy and legal requirements, don’t hesitate to contact us.

1. Reach out to us to get the tool vetted.
2. If the tool cannot be made FERPA compliant and you have received guidance from University Contracts on how to proceed (i.e., make sure you are using it responsibly with respect to FERPA compliance and data security), then you should plan out the details of how you work with the tool to ensure responsible use. 
3. Once your use case and process is well thought out, next consider what data you are entering into any publicly available/consumer tool. Generally speaking you should only be entering data you would share publicly into these systems. Become familiar with the classifications of data you are handling and using with tools that are not CMU licensed. 

Tools you use and the content you create must be digitally accessible. University licensed tools have been vetted for digital accessibility, but your content also needs to be accessible. There is guidance on how to make your content/feedback digitally accessible and we can also consult with you on this aspect.

Do you have proper copyright permissions to enter data into these tools? Consider what will happen with that data – in CMU licensed environments, the models will not train the consumer models and privacy is managed. In public/commercial genAI tool environments, the content you enter will be used for training their models and CMU data privacy requirements will not be met.

To facilitate the X, the TAs will be using a customized ChatGPT bot to assist in grading and providing feedback on homeworks and draft assignments. This bot has been carefully trained on the assignment types, rubrics, and specific kinds of feedback that we require; additionally, TAs will review its output to ensure that the grading and feedback are accurate. All final deliverables will be graded by the TAs, without the use of GenAI. If you have any questions or concerns about this process, please do not hesitate to reach out to myself or one of the TAs. 

All assignments in this class will be graded by the instructor and TAs, without the use of GenAI. If you wish to receive additional or more frequent feedback, you are welcome to use the tool yourself (and see Student Use of GenAI policy above). If you have any questions or concerns about this process, please do not hesitate to reach out to myself or one of the TAs.

To best support your own learning, you should complete all graded assignments in this course yourself, without any use of generative artificial intelligence (AI). Please refrain from using AI tools to generate any content (text, video, audio, images, code, etc.) for any assignment or classroom exercise. Passing off any AI-generated content as your own (e.g., cutting and pasting content into written assignments, or paraphrasing AI content) constitutes a violation of CMU’s academic integrity policy. If you have any questions about using generative AI in this course please email or talk to me.

Similarly, all assignments in this class will be graded by the instructor or TAs, without the use of AI. If you have any questions about the grading and feedback process, please do not hesitate to reach out to myself or one of the TAs.