> ## Documentation Index
> Fetch the complete documentation index at: https://docs.gradingpal.com/llms.txt
> Use this file to discover all available pages before exploring further.

# How AI grading works

> What the AI reads, what it compares against, when it declines to score, and what it will not do.

export const updated_0 = "19 September 2026"

export const page_0 = "/teachers/trust/how-ai-grading-works"

This page is for the cautious teacher and the administrator. It says what happens to a piece of student work between upload and your review.

## What it reads

Three things: the student's work, your instructions, and either your rubric or your answer key. If you added handout materials, it reads those for context. It does not read other students' work while grading one student, and it never sees the rest of the class.

## The two paths

|                         | Open response                                                                                                     | Question based                                                                                                                        |
| ----------------------- | ----------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------- |
| The AI judges against   | Criteria you wrote or approved, one level at a time                                                               | An answer key you approved, one question at a time                                                                                    |
| What it produces        | A level per criterion, the reasoning behind it, quoted evidence from the work, feedback, and comments on the text | A recorded answer per question, a score computed against the key, and feedback with strengths and improvements                        |
| Where judgment comes in | Choosing the level and writing the feedback                                                                       | Only on rubric-graded questions. Multiple choice, true/false, matching and ordering are computed from the recorded answer, not judged |

## When it declines to score

If a criterion needs evidence the AI can't see, it doesn't guess. The card reads **Needs your review** with the reason, and offers **Grade it myself** or **Add context & regrade**. A note marked **For your attention** means it graded but wants you to check something. Both are visible only to you.

## Where you stay in the loop

Every submission lands in **Needs review**. You can change any level or points, with a reason; rewrite any feedback; and ask the assistant for changes it proposes and you approve. Nothing reaches a student until you share.

## What it never does

* It does not decide when students see anything. You do, by sharing.
* It does not show students its analysis. The reasoning, the evidence quotes and your override reasons are stripped before a page is sent to a student. [Who can see what](/teachers/trust/who-can-see-what)
* It does not detect plagiarism or AI-written work, and GradingPal declines instructions that ask it to. Those checks are unreliable and misfire on students writing in a second language.
* It does not curve, rank, or hold a floor across the class. Each student is graded alone.
* Grammar findings never change a score.

## What to read next

<CardGroup cols={2}>
  <Card title="Who can see what" icon="eye" href="/teachers/trust/who-can-see-what">Teacher, co-teacher, student, school.</Card>
  <Card title="Ask GradingPal to make changes" icon="wand-magic-sparkles" href="/teachers/review/ask-gradingpal">Propose, review, apply.</Card>
</CardGroup>

<hr />

<p style={{ fontSize: "0.875rem", opacity: 0.75 }}>Last updated {updated_0}</p>

<p>
  <strong>Was this page helpful?</strong>{" "}
  <a href={"mailto:support@gradingpal.com?subject=" + encodeURIComponent("Help page feedback (helped): " + page_0)}>Yes, this helped</a>

  {" · "}

  <a href={"mailto:support@gradingpal.com?subject=" + encodeURIComponent("Help page feedback (problem): " + page_0) + "&body=" + encodeURIComponent("Which of these fits? Couldn't find the step / Didn't match what I see / Missing information / Out of date / Other\n\nTell us more:\n")}>No, something's wrong</a>
</p>

<p style={{ fontSize: "0.875rem" }}>
  <a href={"mailto:support@gradingpal.com?subject=" + encodeURIComponent("Problem with help page: " + page_0)}>Report a problem with this page</a>
</p>
