Early in my teaching career, a student told me something that changed how I would approach feedback forever. He said that whenever he got his papers back from me, he felt bad because they had so many marks on them.
I was devastated. I thought I was helping. I thought that if I did not point out every error, I was being negligent. I covered those papers in corrections and suggestions because I cared, because I wanted my students to improve, because I believed that more feedback would demonstrate to my students how much I care.
I was wrong. And I still think about the harm I caused more than 25 years later.
I bring this up now because AI feedback tools are making all kinds of mistakes when they provide feedback to learners. The mistakes happen instantly, and without the capacity to learn from the damage unless we set up systems and guardrails to prevent them from doing real harm.
What AI Feedback Gets Right
Before I get to the problems, let me acknowledge what AI feedback can genuinely offer.
Instant availability. Students can get feedback at 11 PM on a Sunday when that might be the only time they have to get work done. They can get it in the moment of writing, when the feedback is most useful, rather than days later (in some cases) when they have moved on mentally. Timely feedback is more effective than delayed feedback, and AI can provide something teachers structurally cannot.
Judgment-free iteration. Some students are afraid to ask teachers for feedback, especially multiple times. When I talk to people who mention that special ed. students should advocate for themselves if they are not receiving their accommodations I think of something I heard someone early in my career say, “Many students would rather go without than stand out.” I think about that a lot. In addition to that, they may worry about being a burden, about seeming like they missed something they were supposed to get, about bothering a teacher. There are no such concerns when going to AI for feedback. A student can submit a draft ten times, revising after each round of feedback, without any social cost. For students who are anxious about evaluation, this can be genuinely liberating. Sometimes I struggle to take feedback because what I create is personal and hearing critique can be tough to handle. I couldn’t care less what AI says about me personally and I am perfectly comfortable taking what it says and deciding if I think it’s worth considering in ways I cannot always do with humans.
Consistent technical feedback. AI can catch grammar errors, identify unclear sentences, flag structural issues, and note consistency problems. It does this without getting tired, without having a bad day, without losing focus on the fortieth paper being graded that day. For certain kinds of mechanical feedback, AI is reliable in ways humans sometimes are not.
These are real benefits. Teachers considering AI feedback tools should not dismiss them.
So let’s get to where I take issue with so much AI feedback I have seen.
The Problem of Overwhelming Feedback
What AI feedback typically does is give students everything, all at once, the way I used to.
Ask an AI to review a student essay and you will get feedback on thesis clarity, paragraph structure, transitions, evidence use, analysis depth, sentence variety, word choice, grammar, punctuation, citation format, and probably more. The AI does not know that this student has been working on paragraph structure for three weeks and finally got it right. It does not know that this student shuts down when overwhelmed. It does not know that pointing out fifteen things means the student will address none of them. Each of those categories also gets a full explanation that is a wall of text no learner is going to read.
As a young teacher, what I did not understand is that effective feedback is selective. Some times good teaching means choosing what to focus on based on where the student is in their learning, what they are ready to hear, and what will make the biggest difference right now. AI rarely gets all of that context when responding to a learner.
Research on feedback is clear on this point. Feedback works best when it is specific, focused, and actionable. Drowning students in corrections does not help them improve. It makes them feel hopeless. It teaches them that their writing is irredeemably flawed. It kills motivation. It can also give them those empty phrases like, “Good job!” which we also know hinders motivation.
AI feedback comes from its training about how writing should look, and it tries to cover everything. Research doesn’t enter into it. Just like with the biases it exhibits, a prompter needs to narrow the scope of the task and tell it how it should respond. If we leave that up to the AI tool itself, the assumptions the algorithms make when it comes to teaching/learning are often wrong.
If we’re suggesting to learners they should use AI for feedback, we should also be telling them how to give instructions to the machine.
The Missing Relationship
There is something else AI feedback cannot provide: the sense that someone who knows you has read your work and responded to it.
When I give feedback now, I try to document what a student did well, especially if it is something they have been struggling with. I try to focus on growth. I try to ask questions that show I am curious about their thinking, not just evaluating their output. I try to give them one or two things to work on, things I believe they are ready to tackle, things that will make a real difference.
This kind of feedback requires knowing the student. It requires remembering what they struggled with last time, what they care about, what shuts them down and what motivates them. It requires relationship.
AI has no relationship with students unless there is some ongoing collection of submissions and progress. It cannot say “I noticed you finally nailed your thesis statements, that is real progress” because it does not know that thesis statements were ever a problem. It cannot calibrate its feedback to what a particular student needs because it meets every student as a stranger, every time.
Feedback from someone who knows and cares about you lands differently than feedback from a machine. It carries weight. It matters. Students know when an adult has invested attention in their work, and that investment is part of what makes feedback effective.
The Sycophancy Problem, Again (Still)
I have written about sycophancy before, but it shows up particularly clearly in feedback contexts. AI tools are trained to be encouraging. They tend to lead with praise, soften criticism, and frame everything positively.
For some students this is exactly the wrong thing. Some students need to be pushed harder. Some students need honest assessment of where they are falling short. Some students have been told their whole lives that their work is great when it is actually mediocre, and what they need is someone willing to say that clearly and kindly with sufficient support.
Without proper guidance, AI cannot make these judgments. It may give everyone the same encouraging tone, regardless of whether encouragement is what they need. I would have hated that when I was a student. AI can validate work that should be challenged. It can tell students their essays are “well-structured” when the structure needs significant work. It prioritizes making students feel good over helping them get better.
Good feedback sometimes feels uncomfortable. That discomfort, delivered by someone who cares about you and believes in your capacity to improve, is part of how learning happens. AI cannot deliver that kind of productive discomfort because it has no relationship to fall back on, no trust to draw from.
How to Use AI Feedback Well
I am not arguing that teachers should never use AI feedback tools. I am arguing that we need to use them thoughtfully, with awareness of their limitations.
Constrain what AI focuses on. If you are using AI feedback tools in your teaching, set parameters. Ask the AI to focus only on paragraph structure, or only on evidence use, or only on one specific skill you have been teaching. Prevent the overwhelming flood of undifferentiated feedback by limiting the scope. Create a document that provides learners with a template they can use when they start their prompt. Here are some things I like to use:
- Use a tone that demonstrates care, support and empathy for the learner. Never give the learner responses that will do the assignment for them. You can give examples for what that principle might look like in a different context.
- Feedback must be 10 sentences or fewer.
- Two to three strengths These are things the learner did well. Why those features are effective, based on the assignment and rubric criteria. Support your response briefly.
- Two important areas for revision based on priority, if the assignment did not meet requirements. Support your reasoning, based on the assignment and rubric. Offer a suggestion for revision.
- If the teacher is the one using it then I would also suggest to the LLM that it should, “Use a tone that demonstrates care, support and empathy for the learner. Never give the learner responses that will do the assignment for them. You can give examples for what that principle might look like in a different context.“
Use AI for technical review. AI is reasonably good at catching grammar errors, identifying unclear sentences, and flagging mechanical issues. It is much less good at evaluating argument quality, creativity, voice, or growth over time. Match the tool to the task.
Teach students to filter AI feedback. If students are going to use AI for feedback (many will regardless of what we tell them) teach them to be critical consumers. Not all feedback is equally important. Students need to learn to identify which suggestions matter for their specific goals and which they can ignore. This is actually a valuable skill in itself.
Preserve human feedback for what matters most. Given limited time, teachers should invest their feedback energy where it will have the most impact: responding to ideas, noticing growth, asking questions that push thinking, providing the relationship-grounded assessment that only humans can give. Let AI handle the mechanical stuff if needed, but do not outsource the human parts.
Model good feedback practices with AI. Show students what research-based feedback looks like. Compare overwhelming AI feedback with focused human feedback and discuss why the approaches are different. Help students understand that more is not always better, and that good feedback is an act of selection and care.
What I Wish I’d Known
Not every mark on a paper helps. Some weigh a student down. Feedback is communication between people, shaped by relationship, trust, and understanding of where someone is in their journey.
It falls to us to know better. To use these tools where they help and protect our students where they do not. To remember that feedback exists to help a specific person, in a specific moment, take a specific next step.
That student who felt bad when he got his papers back from me taught me more about teaching than any professional development I have ever attended. I hope I can pass that lesson on to students, other teachers, and maybe even to the people building these AI tools.
More feedback is not better feedback. And better feedback requires knowing who you are talking to.