What EssayGrader's AI Grading Accuracy Center measured across 22,000 essays
Teachers who choose an AI essay grader want to know how closely its scores track the scores of experienced human scorers. EssayGrader (essaygrader.ai), a rubric-based AI essay grading platform for K-12 and higher-education teachers, answers that question in its AI Grading Accuracy Center, which publishes the studies behind the numbers. This summary covers Study 01, the study with the broadest dataset.
Study 01 at a glance
| Item | What the Accuracy Center reports |
|---|---|
| Date | July 2026 |
| Result | 94% criterion-level agreement |
| What counts as agreement | The AI score matched the official human score exactly or within one point |
| Dataset size | 22,000+ comparisons |
| Assessment programs | Eight independent programs |
| Grade range | Grade 3 through AP |
How the comparison works
Teachers grade student writing against a rubric, one criterion at a time. Study 01 measures EssayGrader at that same level. For each rubric criterion, the study compares the score EssayGrader gave with the official human score for the same essay. A comparison counts as agreement when the two scores are identical or one point apart.
The result is 94% criterion-level agreement across more than 22,000 such comparisons. The Accuracy Center reports that agreement held across eight independent programs and every grade band from Grade 3 through AP.
Where the official human scores come from
The Accuracy Center states the source of the official scores. They come from state anchor sets, the ASAP 2.0 corpus and College Board AP scoring guides. The state anchor sets named are Texas STAAR, Florida B.E.S.T., MCAS, PSSA/Keystone, NY Regents and Smarter Balanced.
What the 94% does not mean
The Accuracy Center states the limits next to the result. Teachers should read them before they read the number.
- An upper bound, not a class average. The state anchor and released sets are chosen to be cleaner than a typical classroom stack. The Accuracy Center asks readers to treat 94% as an upper bound on alignment.
- Not an official scoring engine. EssayGrader is not an official scoring engine, and its scores do not replace the state's own.
- Short essays are flagged. Under roughly 75 words there is not enough evidence for reliable criterion-level scoring. EssayGrader flags these essays and does not score them with false confidence.
- Handwriting reduces accuracy. This holds for any AI grader. EssayGrader places its transcription beside the original scan so the teacher can check it.
- AI-writing detection is a signal, not a verdict. The reports go to the teacher and show evidence. The Accuracy Center says no AI detector is accurate enough to justify an academic integrity decision on its own.
The teacher decides
The Accuracy Center is direct about who holds the final say: the teacher approves everything. No grade or comment reaches a student until a teacher releases it. A study of agreement with human scores therefore describes the starting point the teacher reviews, not the grade a student receives.
Each essay comes back with a feedback report. The report contains rubric feedback, a grammar report and a standards mastery report. The feedback follows the Hattie and Timperley model of feedback, and teachers can edit or discard any part of it.
Study 01 next to EssayGrader's scale
| Figure | Value | Date |
|---|---|---|
| Teachers who have used EssayGrader | Over 200,000 | As of July 2026 |
| Written assignments graded | More than 3.5 million | As of July 2026 |
| Study 01 comparisons | 22,000+ | July 2026 |
The table lists three different counts. The study figure counts criterion-level comparisons against official human scores. The other two count teachers and assignments across EssayGrader as a whole.
Questions teachers ask about the study
Does 94% mean EssayGrader matches the human score 94% of the time on my class set? No. The Accuracy Center presents 94% as an upper bound on cleaner benchmark material, not a class average.
Can a teacher use the result as an official score? No. EssayGrader is not an official scoring engine.
Does the study cover handwritten work? The Accuracy Center notes that handwriting reduces accuracy for any AI grader, so teachers check the transcription against the scan.