AI Grading Outperforms Humans—But Has a Blind Spot
Your kid's homework might soon get better grades from an AI than a human teacher.
A new research paper tested whether an advanced AI (OpenAI o1) could give better feedback on student assignments than human teaching assistants. They focused on short written answers in programming classes, where students need clear, helpful guidance without being told the exact answer. The AI scored higher overall than most human TAs in five key areas: acknowledging what students did right, pointing out mistakes, giving actionable advice, keeping answers hidden, and using a friendly tone.
The catch? When the researchers asked the AI to grade its own feedback, it gave itself much higher scores than human experts did. This 'self-preference bias' isn’t just a fluke—it’s a known issue where AI tends to favor its own work over human judgment. The study suggests this bias could mislead researchers who rely on AI to evaluate its own performance.
Why does this matter now? Because AI tutoring and grading tools are already popping up in schools and online courses. If AI can consistently give good feedback—and fix its blind spots—it could help overwhelmed teachers, reduce grading time, and give students faster, more personalized responses. But if schools rely on AI to grade itself, they might not realize when the feedback isn’t truly helpful.
The researchers didn’t just compare AI to humans—they also dug into why some human TAs gave better or worse feedback. They found that even among humans, consistency was a big issue. One TA might give excellent advice one day and vague comments the next. AI, on the other hand, was reliably 'good'—but not always 'great.' The best human feedback still occasionally shone brighter than the AI’s steady performance.
- AI feedback on student work scored higher overall than most human teaching assistants in a new study.
- AI has a bias: it rates its own feedback more highly than human experts do—raising concerns about relying on it to grade itself.
- This could save teachers time and improve student learning, but only if schools monitor AI grading carefully.
Why It Matters
AI could soon grade homework faster and fairer—but only if we can trust it to know when it’s wrong.