A co-design pilot with nineteen K-12 teachers found a pattern that many educators will recognize. Teachers valued the speed and usefulness of the AI's written feedback, especially for formative work, but they distrusted its numeric scoring and wanted human oversight over every grade. Students in the same study liked getting fast, revision-oriented comments but remained skeptical of grading handled entirely by a machine. The finding is a useful guide to where AI helps most and where it needs the firmest guardrails.

Teachers in related research described AI output as a helpful first draft of feedback, one that saved time and reduced the burden of writing initial comments. About 57 percent of surveyed teachers called the feedback clear and actionable, which is encouraging but also means a sizable minority found it vague or even misleading. Roughly a quarter of teachers called it vague or unhelpful and about one in five called it incorrect. Those numbers are the strongest argument for treating AI comments as drafts that a professional must read before students see them.

The score problem is different in kind. Teachers reported cases where the AI applied inconsistent point scales across submissions or deducted points for criteria unrelated to the assignment. A comment that is slightly off can be edited in seconds, but a score that is off can change a grade, and students notice. This is why teachers grow comfortable with comments sooner than with numbers.

Why comments are easier to review than scores

A draft comment is a piece of language you can accept, trim, or rewrite while you read the student's work. You see the reasoning on the page, so errors are visible, and the cost of fixing one is a few seconds. A score, by contrast, compresses many judgments into a single number, which hides the reasoning behind it. Unless the system also shows which rubric descriptor it matched and why, you have to rebuild that logic yourself to know whether the number is right.

• Use a teacher-written rubric with fixed point values so the scale cannot drift between papers.
• Require the tool to cite the rubric descriptor it used for each criterion score.
• Spot-check a sample of scores against your own before relying on the batch.
• Treat every AI comment as a draft and edit for accuracy and tone before release.
• Keep final grade entry in the teacher's hands, never automatic.

Setting the rubric scale so scores stay consistent

Consistency problems usually trace back to ambiguity in the scoring frame. If a rubric says a thesis can earn up to four points but never defines the difference between a three and a four, any scorer, human or machine, will fill the gap differently each time. Writing performance-level descriptors in observable terms, such as names a specific claim, previews at least two supporting points, and takes a position that could be argued against, narrows that gap considerably. A rubric tight enough to guide a new colleague is usually tight enough to guide an AI first pass.

Calibration is the other half. Before a full class set, run five or six papers you have already scored yourself and compare results criterion by criterion. Where the tool and your judgment diverge, the cause is almost always a descriptor that needs sharpening rather than a flaw in the technology. Fixing the descriptor once improves every later batch, and it leaves you with a rubric that colleagues can use as well.

Using AI comments for formative work first

Many teachers begin with low-stakes drafts, where the purpose of feedback is learning rather than a grade. Students receive comments quickly enough to revise while the assignment is still fresh, and the teacher reviews them in bulk instead of composing each from scratch. Because the stakes are lower, a minor imperfection in a comment costs little, and the teacher builds familiarity with the tool's tendencies. That experience informs later decisions about whether to use first-pass scoring on summative essays.

Formative use also tends to improve revision behavior. Students who receive criterion-specific comments within a day or two are more likely to act on them than students who wait two weeks for a stack of papers. The comments do not have to be perfect to be useful, as long as they name what is working and one concrete next step. Teachers can then spend saved time on conferences with the students who need the most help.

Where the teacher stays in the loop

The research on AI grading points toward a clear design principle: let software handle the mechanical first pass and keep professional judgment on the decisions that matter. That means the teacher reviews each score, adjusts what does not match their reading, and personalizes comments that feel generic. It also means students should be told how feedback is produced and who stands behind it. Transparency of that kind addresses much of the student skepticism the studies observed.

For school leaders choosing tools, the implication is to ask less about how impressive the scoring is and more about how easy it is to review and override. A tool that makes teacher edits fast and visible will be trusted and used. A tool that asks teachers to accept a number on faith will be abandoned after the first disputed grade. Trust, in the end, is a workflow property as much as an accuracy property.