AI Feedback: Formative Assessment Meets the Bias Test

Formative assessment — the kind that helps you learn rather than sanctions you — lives on one thing: useful feedback, at the right moment. That is precisely what AI can provide at scale: immediate, detailed, personalised feedback on a draft, a line of reasoning, an exercise. But recent research reveals human subtleties that no generation speed can solve.

Feedback that genuinely improves work

Efficacy evidence is arriving. In a randomised trial in a large university course, students receiving AI-mediated feedback produced significantly higher-quality revisions, with a growing effect as teaching assistants adopted the AI's suggestions ^[1]. Theoretically, careful prompt design lets LLMs deliver varied scaffolds — metacognitive, motivational — that support self-regulated learning ^[2].

Source bias: who gives the feedback matters

More troubling: students often misjudge feedback correctness. A study of nearly a thousand feedback pieces rated by hundreds of students shows they struggle to tell correct from erroneous feedback, and readily assume the best came from a human ^[4]. In other words, you cannot rely on the learner to filter the AI's mistakes.

Use matters more than the tool

The effect depends heavily on how feedback is used. A physics study shows compulsory hints help low performers but harm high achievers' self-regulation, while on-demand help aids the more advanced but can erode struggling learners' autonomy ^[5]. A review of educators confirms that speed and personalization gains are offset by risks of bias, privacy and weakened teacher-student relationships ^[6].

• Strong: immediate feedback, self-regulation scaffolding, reduced teacher workload.
• Weak: perception biased by source, unreliable learner judgement of quality.
• To calibrate: compulsory vs on-demand hints, depending on student level.

Good feedback does not only say what is wrong; it makes the learner able to see it themselves next time.

Learnya synthesis

The emerging consensus is clear: AI works better as a supplement to the teacher than as a replacement, with hybrid models most effective ^[7]. Neuro-symbolic approaches also make mastery tracking more interpretable by injecting explicit pedagogical rules ^[8]. Fast, AI can be; trustworthy, it becomes so through design, human verification and controlled data governance.

Making feedback actionable and credible

Feedback only has value if it is understood and acted upon. Yet research reveals a double human obstacle: learners misjudge the correctness of feedback ^[4] and perceive it differently depending on whether it is attributed to a human or an AI ^[3]. In other words, excellent AI-generated feedback can be ignored simply because it carries the 'AI' label. Transparency about origin and human support therefore become levers of adoption as much as of quality.

The effect also depends finely on dosage and student level. Imposed hints can help the weakest but harm the self-regulation of the more advanced, while on-demand help produces the opposite ^[5]. There is no universal setting: a good system adjusts the amount and timing of support, and gradually lets the learner take back control. The ultimate goal of formative feedback is to make itself unnecessary.

This is why the consensus leans toward hybrid models, where AI amplifies the teacher's work rather than replacing it ^[7]. The teacher brings judgement, relationship and knowledge of context; AI brings speed, availability and coverage. More interpretable approaches to mastery tracking ^[8] further strengthen trust by making the reasons for a recommendation legible. Fast and credible: feedback becomes so through design, verification and controlled data governance.

Feedback, the quiet engine of learning

Among all AI uses in education, formative feedback is perhaps the most promising because the most aligned with what makes people learn: useful feedback, frequent, at the right moment. AI removes the main historical obstacle — the cost in time — by making near-immediate feedback possible for every learner, at every step.

But recent findings on perception biases and learners' unreliable judgement remind us that good feedback is not just about producing it. It must be made credible, actionable and suited to each level, and kept under a teacher's eye. The future is not AI replacing human feedback, but an alliance where each brings its best — speed from the machine, meaning and relationship from the human.