The Feedback Problem: Why AI Can't Replace the Human Touch (But Might Help Anyway)

Image credit: Kathryn Conrad / https://betterimagesofai.org / https://creativecommons.org/licenses/by/4.0/

I've been doom-scrolling through LinkedIn a little too much lately (but, to be fair, who doesn't?), and there's one panic that keeps popping up with alarming frequency: "What's the point if students submit AI work and teachers give AI feedback?"

Fair question. Terrifying question. But also, maybe, the wrong question.

What even is feedback?

Let me start with something that might sound obvious but apparently isn't: feedback should be relational. If you're just marking right and wrong like some sort of pedagogical traffic light system, nobody's learning anything meaningful. But equally, those lengthy comments about "areas for improvement" that we've all written at 11pm on a Sunday? They're not necessarily better.

Here's what actually works: recognition. This article mentions a brilliant piece of research where students received feedback with one simple addition: "I'm giving you these comments because I have very high expectations, and I know you can reach them." That's it. That tiny relational nudge, that signal of high expectations? It improved student achievement. Even more remarkably, it had an even higher impact on Black students who'd previously been less engaged with learning.

What does this tell us? That feedback is personal. It's about understanding a student's entire learning journey, not just the work that's directly in front of you.

(And no, before you panic, I'm not advocating for a return to the marking insanity of my NQT year. Perhaps you know the drill: marking on every page, compulsory DIRT questions that led to Q&A activities for the sake of Q&A activities, and my personal favourite—writing "VFG" everywhere just to prove you'd actually had a conversation with a student. We've mercifully moved on from that particular circle of hell.)

All of this brings us to our current predicament. AI doesn't recognise. It analyses.

Sure, AI can technically identify what a student should work on next based on the specific piece they've submitted. It can spot grammar issues, suggest structural improvements, even identify gaps in argumentation. But it can't recognise that Sarah always struggles with conclusions because she's terrified of being wrong. It doesn't know that Ahmed's writing has massively improved since September but he still needs pushing on analytical depth. It has no idea that Charlie is usually brilliant but this piece is rushed because their parents are going through a divorce.

(Unless, of course, you feed it all that information. Which of course you absolutely, definitely, categorically should not do. Ever.)

The Workload Temptation

But here's where it gets properly complicated. Picture this: it's Sunday night, you've got 90 'Why did William win the Battle of Hastings' paragraph responses to mark by Tuesday, your own kids need help with their homework, and the AI feedback tool is right there, generating perfectly serviceable comments that would genuinely help students improve.

The temptation is real. And honestly? Sometimes the AI feedback might even be better than what you'd produce at midnight, running on fumes and your fourth coffee.

It's not your fault. It's just another "wicked problem"; one without a neat solution, where every answer creates new question.

And there's an even thornier issue lurking here. If students are submitting AI-generated work and receiving AI-generated feedback, how do you actually know what they can do? How do you recognise their growth, identify their struggles, understand their thinking? You know, the whole point of teaching?

So what do we do? As always, I'm looking for approaches that hit at least one of my holy trinity: workload, wellbeing, or attainment. If EdTech doesn't improve one of those, bin it.

So here's what I'm thinking: some human-AI feedback workflows that actually keep the human where it matters. How about trying a couple of these and letting me know your thoughts?

1. AI First Draft, Human Recognition

Let AI generate the technical feedback on student work first: grammar, structure, evidence use. Then you add the recognition layer by adjusting the feedback accordingly:

"I'm giving you this feedback because I noticed you've been working on analytical depth since our last conversation. Look how far you've come with your evidence selection! Now let's push that analysis even further..."

The AI handles the time-consuming technical stuff, you handle the relationship and growth trajectory. Students get both, you don't die of exhaustion.

2. Comparative Analysis with AI

Depending on the age of your students (this one might be better with 18+ student in FE, for example), you could get them to first submit their own work to multiple AI tools for feedback, then write a reflection on which feedback they found most useful and why. You mark the reflection, not the original piece. This way you're assessing their metacognitive skills and their ability to evaluate feedback quality—much more valuable than marking the essay itself. Bonus: They're learning to critically evaluate AI suggestions, which is exactly the skill they need moving forward.

3. Prompt Libraries as a Rubric

Create assessment-specific prompt libraries that students must use to self-assess before submission. They submit their work WITH their AI-generated self-assessment and a paragraph explaining what they agree/disagree with. You're marking their evaluative judgment and self-awareness, not just the work. Plus, you can quickly scan to see if they've identified the same issues you would have.

4. Viva Voce Plus AI Prep

Another one for older students or perhaps much smaller classes; viva-style assessment is a bit crazy with an entire cohort of Year 7s! In this format, students use AI to prepare for verbal assessments. They can practice with AI, generate potential questions, refine arguments. But the actual assessment? That's a conversation with you. No AI present, just human connection. You get to actually hear their thinking, they can't hide behind generated text, and the feedback happens in real-time. The relationship stays intact.

Making It Work in Practice

The key is being explicit about which model you're using when. Students need to know:

• When they're building independence vs. distributed intelligence

• What type of feedback they'll receive and why (remember, why is the key part to anything we do as teachers!)

• How their growth is being tracked beyond individual pieces of work

• Honestly, that you're a human looking at their human work! Don't let them think that you're offloading your marking and feedback to an AI tool. Make it clear who is doing what!

  
  

And critically, you need to protect the recognition element. That's the bit AI can't do - knowing that Sarah's conclusion is actually massive progress, or that Ahmed needs pushing because he's capable of more, or that Charlie needs checking in with because something's off.

The uncomfortable truth is that we're trying to solve a fundamentally human problem with technological tools. Feedback works because it's relational, because it recognises the whole student, because it signals that someone who knows them believes they can do better. AI can augment that recognition, support it, maybe even free up time for more of it. But it can't replace it.