AI grading is close to teacher grading on some tasks and noticeably off on others. The studies below show where it lands, where it drifts, and the simple review routine that makes it safe to use.
The quick answer: partly
Studies from 2024 and 2025 find AI grades broadly similar to teacher grades on some tasks, with predictable differences. In one study ChatGPT was more lenient than instructors on weaker essays and stricter on stronger ones (Wetzler et al.). How close AI gets depends on the subject, the task and how you set it up:
- Multiple choice and short answers with clear criteria: the most reliable use
- Rubric-based essay grading: often close, but expect differences at the top and bottom of the class
- Open-ended, holistic writing: the least consistent
- English learners and non-standard English: the riskiest case, so review every score
Bottom line
- AI is most accurate on structured tasks with clear right answers.
- On essays it tends to pull scores toward the middle: too kind to weak work, too tough on strong work.
- Use it for a first pass and feedback drafts, then review the outliers yourself.
- Keep your own judgment primary on anything high-stakes.
What the studies found
These are the peer-reviewed studies most often cited on AI versus human grading, with what each one actually measured.
Flodén (2025), British Educational Research Journal
Grading exams using large language models compared ChatGPT’s grades with teachers’ grades on university exams. About 70% of ChatGPT’s grades fell within 10% of the teacher’s grade. ChatGPT avoided extreme scores and did worse on questions tied to specific lecture content.
Wetzler et al. (2024), Teaching of Psychology
Grading the graders compared ChatGPT 3.5 and 4o with eight instructors on 186 psychology essays scored with the same detailed rubric. Both AI versions were more lenient than humans at lower performance levels and stricter at higher levels, a pattern the authors call proportional bias. GPT-4o was closer to the instructors than GPT-3.5.
Morjaria et al. (2024), International Medical Education
Examining the efficacy of ChatGPT in marking short-answer assessments found Spearman correlations of about 0.6 to 0.7 between ChatGPT and human assessors. ChatGPT’s score differed from the human score on most answers, but by more than one point on only 20 to 38% of them. The authors call it “a viable, though imperfect, assistant to human assessment.”
Deepshikha (2025), Discover Artificial Intelligence
A comprehensive review of AI-powered grading pulls together recent research and concludes that AI does well on objective, criteria-based evaluation and struggles more with nuanced, subjective work, so human oversight stays essential.
What universities recommend
Ohio State’s Arts and Sciences distance education office summarized this research in 2025 and concluded that “AI and auto-grading tools should be seen as supportive technologies and not replacements.” MIT Sloan’s teaching technology team advises instructors to “regularly audit AI systems for accuracy, fairness, and potential biases,” anonymize student work, and tell students when AI is part of grading.
Almost all of these studies used college students. K-12 research on large language model grading is still thin, so treat the numbers as a guide, not a guarantee for your 7th graders.
Where AI grading does well
No fatigue
Paper 90 gets the same attention as paper 1. Mood, hunger and the clock don’t creep in.
Speed at scale
Seconds per paper, compared with the minutes a careful human read takes.
Matching explicit criteria
Good at checking work against criteria you spell out, like “cites two pieces of evidence.”
Detailed feedback drafts
Can comment on several parts of a piece at once, which gives you a starting point for feedback.
Practice and drafts
The research is most supportive of AI for formative work, where a rough score is fine.
Where it falls short
Creativity and originality
May mark down an unconventional approach that you would reward.
Cultural context
May miss references and ways of arguing from students’ own backgrounds.
Complex arguments
Less reliable on sophisticated rhetoric and nuance.
Pull toward the middle
Too kind to weak essays and too harsh on strong ones (Wetzler et al., 2024).
Run-to-run variation
The same essay can get a different score if you grade it again.
How AI grading works
Knowing how the two main kinds of AI grading work explains both their strengths and their blind spots.
Automated essay scoring: the older approach
Traditional automated essay scoring, like ETS’s e-rater, scores measurable features of the text: length, vocabulary, sentence variety, organization signals and grammar. It doesn’t read for meaning. It is reliable on surface features and weak on content depth and the quality of an argument. That is why testing programs pair it with people: on the GRE, each essay gets a human score and an e-rater score, and a second human reader steps in when the two disagree.
Large language models: the newer approach
Models like ChatGPT, Claude and Gemini work with meaning, not just surface features, so they can comment on whether an argument holds up. That is why they can give useful feedback. It is also why their scores can drift: they are making a judgment, and judgments vary.
Do rubrics make AI more accurate?
You’ll often read that AI is “much more accurate with a rubric.” The evidence is more mixed than that. In Morjaria et al. (2024), ChatGPT’s scores matched the human markers best when it had no rubric at all. Wetzler et al. (2024) gave the AI a detailed rubric and still found it scored differently from instructors.
A rubric is still worth it, for a different reason: it makes the AI’s feedback specific and checkable. When every score is tied to a criterion you wrote, you can see exactly where you disagree and fix it in seconds.
Give the AI your rubric so its feedback lines up with what you taught, but don’t assume the rubric alone makes the scores right. Your review is what makes them right.
What this means in your classroom
When to use AI grading
- Feedback on drafts and revisions
- Big stacks (30 or more papers on one assignment)
- Short-answer and structured responses
- A first-pass score before your own review
- Tasks with explicit criteria
When to be careful
- High-stakes summative grades
- Creative or very personal writing
- English learners, unless you review each score
- Your highest and lowest performers, where the bias toward the middle is strongest
- Assignments with no clear criteria
The calibrate and verify routine
This is the approach most consistent with the research:
Run the AI first pass on the whole set
Let it score everything and draft feedback.
Check the outliers
Read the highest and lowest AI scores yourself. That is where the studies found the biggest gaps.
Spot-check the middle
Read 5 to 10% of the mid-range papers to make sure the AI is on track.
Adjust where you disagree
Change any score that doesn’t match your judgment. You know the student; the AI doesn’t.
Edit the feedback
Treat AI comments as a first draft. Add the line only you could write.
Curious how much time this saves? See how much time teachers spend grading, or follow our step-by-step guide to grading essays with AI. Comparing tools? Start with our ranking of AI grading tools.
The research consensus
- Close to human grading on structured assessments
- Useful for formative feedback and drafts
- Not suitable as your only grader
- Needs teacher review and adjustment
- Not for high-stakes grades without a human check
Frequently asked questions
How accurate is AI at grading essays?
Close, but not interchangeable with a teacher. In Flodén (2025), 70% of ChatGPT’s exam grades landed within 10% of the teacher’s grade. In Morjaria et al. (2024), ChatGPT’s short-answer marks correlated with expert markers at about r = 0.6 to 0.7, which the authors called a viable, though imperfect, assistant. AI is closest on structured tasks and least reliable on holistic essay judgment.
Does AI grade differently than teachers?
Yes. Wetzler et al. (2024) found ChatGPT was more lenient than instructors on weaker essays and stricter on stronger ones, even with the same rubric. AI also grades the text as written: it does not know a student’s growth, effort or circumstances the way you do.
Can AI grading be biased?
It can be. Language models learn from mostly standard, Western academic English, so researchers and universities recommend auditing AI grades for fairness, especially for English learners and students whose writing follows other rhetorical traditions. We have not found a strong K-12 study that measures this directly, so review those students’ scores yourself.
Should teachers use AI to grade student work?
For low-stakes feedback and big stacks of drafts, AI can save real time. The research supports using it as a first pass that you review, not as the final grade, and keeping your own judgment primary on summative or high-stakes work.
What makes an AI grading tool trustworthy?
Look for three things: it grades against criteria you can see (your rubric), it shows its reasoning for each score, and it lets you change any score before students see it. Being able to check and override the AI matters more than any accuracy claim on a website.
Is AI better at grading math or writing?
Structured work. Problems with clear right and wrong answers and short answers with objective criteria are where AI agrees most with human graders. Complex analytical writing is where it agrees least. The more objective the task, the closer AI gets.
Is AI grading more accurate for elementary or high school writing?
No study we found tests this directly. Our expectation, based on the research above: AI should be closest on shorter elementary pieces with simple rubric criteria and needs the most review on AP, IB or college-prep writing with nuanced arguments. Treat AP-level AI scores as a rough draft.
Is AI grading used in high-stakes testing?
Yes, with a human in the loop. On the GRE, ETS has each essay scored by a trained human reader and by its e-rater engine. If the two agree, the scores are averaged; if they disagree, a second human reader scores the essay and the two human scores are used. That human check is the model to copy in a classroom.
