Using Generative AI to provide feedback in Second Language Writing - Application insights by Prof. Shaofeng Li
In this post, I focus on using generative AI (GenAI) to provide feedback on second language learners’ written production, as providing feedback on students’ writing is one of the most commonly used functions of GenAI. This post is intended for teachers, students, researchers, and other individuals interested in benefiting from GenAI feedback. Some of the ideas I discuss here are derived from an article I authored on the application of GenAI in second language writing (Li, 2025; see the bibliographic information at the end of this post). In this post, I discuss principles for regulating GenAI feedback, which refers to users’ attempts to intervene in the feedback seeking process by providing criteria, supplying contextual information, assigning a role to GenAI, specifying feedback characteristics, vetting feedback quality, and exercising user agency.
Provide feedback criteria. GenAI feedback should follow criteria, which could be the rubric for a writing assignment in a language class or the scoring protocol of a writing test such as the writing sections of TOEFL or IELTS. The reason criteria must be provided is that GenAI needs to be informed of the expectations to be met for the writing task. Criteria are genre- and task-specific, and without knowledge of the criteria, it is impossible for GenAI to provide targeted feedback. It is also important to provide training on the criteria, especially those that are not straightforward or are subject to multiple interpretations. For example, one of the assessment criteria for the IELTS writing test is “Task Response”. Without elaboration and exemplification, GenAI will likely not know what it means or how to provide feedback on that aspect of writing, or it may interpret it in ways that deviate from what it is intended to mean by the test developer. Furthermore, in a language class, teachers must ensure that feedback criteria for writing assignments are consistent with assessment criteria or the criteria used by teachers to grade students’ essays. In other words, what is assessed must be aligned with what is taught.
Supply contextual information. Contextual information includes the writer’s background, the writing prompt, and the objectives of the writing task. The writer’s background may include information about the writer’s school level (high school), their second language proficiency such as levels in the Common European Framework of Reference for Languages (CEFR), the instructional setting, and any other information that may help GenAI provide customized feedback tailored to the learner or learner group. The writing prompt, which refers to the actual instructions for the writing assignment and may include information about the genre, audience, organization, word limit, and time limit, should be provided verbatim to GenAI. In addition, the objectives of the writing task, such as helping students learn how to write coherently or improving their accuracy in using the English subjunctive mood, should be communicated to GenAI.
Assign an appropriate role. Assigning a role means giving GenAI a certain identity, such as teacher, editor, or native speaker—typical roles GenAI is asked to perform in research on GenAI feedback. Role assignment should not be random and should be based on the objective and context of the writing task. If the essay is written by an ESL student, then the assumed feedback provider is typically a teacher, who provides feedback for the purpose of helping the student learn how to write and improve their language skills, instead of or in addition to making the essay perfect. The teacher may prioritize errors typical of writers at this level or errors associated with the curricular goals or objectives of the current instructional module. If the feedback provider is a journal editor (or copyeditor), the primary objective is to improve the quality of the text instead of helping the writer learn how to write. Thus, feedback for learning is distinct from feedback for textual improvement. If GenAI is given the role of a native speaker, its duties as a feedback provider are vague and often need to be specified by the feedback seeker.
Specify feedback characteristics. Feedback characteristics refer to the type, amount, language used (namely the writer’s first or second language), and layout (e.g., integrated in the text or presented as a list) of the expected feedback. According to research, these characteristics influence the effects of feedback on the development of students’ writing ability. For example, direct feedback, which provides the correct form to replace the error, has been found to be more effective than indirect feedback, which alerts the writer to the presence of an error without providing the correct form. Focused feedback that addresses certain aspects of writing is more effective than unfocused feedback that corrects all errors. Building on this point, I would like to argue that what feedback to elicit from GenAI is an empirical question, and teachers, students, and researchers should make informed decisions instead of decisions based on assumptions and experience.
Vet feedback quality. One important step in feedback regulation is to vet the quality of GenAI feedback, which can be defined in terms of whether the generated feedback is consistent with the requirements or instructions included in the prompt and whether it meets the criteria for high-quality feedback. In Li (2025), I discussed the criteria for feedback quality, including whether the feedback is accurate, accessible (easily understood), actionable (writers know what to do), clear (unambiguous), supportive (vs. critical or negative), necessary (as opposed to redundant), and informed by research. It is important to vet GenAI feedback because research has demonstrated that it can be inaccurate, inconsistent, and unsystematic.
Exercise user agency. One overarching, bedrock principle in feedback seeking is to exercise user agency, which is crucial in the era of GenAI. Exercising agency requires GenAI users to play an active role in the entire process of applying GenAI, using GenAI as a tool to meet their needs instead of allowing GenAI to dominate the process. Agency refers to a user’s endeavor to compare, select, decide, participate, evaluate, change, and innovate. Behaviors that reflect agency include, but are not limited to, comparing GenAI tools, making choices, conducting prompt engineering, specifying the parameters of the expected output (e.g., feedback), vetting the quality of the output, asking for justification, retaining writer voice and identity, improving GenAI’s functions, developing new applications, and developing GenAI literacy (basics of GenAI, strategies for effective use, limitations, and ethics). To be clear, exercising agency does not mean abandoning GenAI; rather, agency places humans at the center of technological advances, maximizing GenAI’s benefits and maintaining its subsidiary and instrumental status.
Reference
Li, S. (2025). Generative AI and second language writing. Digital Studies in Language and Literature, 2(1), 122-152. https://doi.org/10.1515/dsll-2025-0007
The information in this post is copyrighted.