Show where AI tutor answers end up in student work

For learning platforms, schools and education teams that offer an AI tutor. The tutor labels every answer it gives. When a submission contains copied tutor text, a check finds the paragraphs and the tutor session, so a teacher can look at the actual conversation instead of relying on a guess.

Endpoints: POST /api/watermark/encode or /tokens, /decode  ·  MCP tools: watermark_encode, watermark_decode  ·  Cost: 1 credit per tutor answer; checks free unless a label is found

The problem

Students use AI tutors, and that is often the point. Teachers still need to see when tutor answers are handed in as a student's own work. AI detectors estimate whether text "looks" machine-written; they produce false accusations, especially against students writing in a second language. A label is different: it is a fact about your tutor's output, added when the answer was generated.

How it works

1. The tutor labels each answer with an opaque session id, for example ai;tutor;s=s-20931, through POST /api/watermark/encode. Tutors that stream their answers use /tokens, as shown in Label AI-generated text.
2. Submissions are checked paragraph by paragraph with POST /api/watermark/decode. Paragraphs without a label cost nothing.
3. A teacher opens the session for each labelled paragraph and sees the question the student asked and the answer the tutor gave.

Use it fairly

• Tell students that tutor answers carry an invisible label. Transparency is fair, and it changes behaviour on its own.
• Treat a found label as the start of a conversation, not as a penalty. Quoting the tutor with attribution can be perfectly acceptable.
• Never treat a missing label as proof that a student worked alone.
• Keep names out of labels and follow your data-protection rules for minors.

Add it to your workflow

The examples read your API token from TEXTWATERMARKING_API_TOKEN. Each example below is run against the API before this page is published.

Label tutor answers (Python)

If labelling fails, the student still gets the answer; log the failure in your own system.

# Label each answer your AI tutor gives (Python 3.10+, pip install requests). 1 credit per answer.
# The label is an opaque session id: anyone with an API token can read it, so never put a student name in it.
import os

import requests

API = "https://textwatermarking.com/api/watermark"
HEADERS = {"Authorization": f"Bearer {os.environ['TEXTWATERMARKING_API_TOKEN']}"}


def label_answer(answer: str, session_id: str) -> str:
    """Return the answer with an invisible 'ai;tutor;s=<session>' label, or unchanged if labelling fails."""
    try:
        r = requests.post(f"{API}/encode", headers=HEADERS, timeout=10,
                          json={"text": answer, "secret": f"ai;tutor;s={session_id}"})
        r.raise_for_status()
        return r.json()["encoded"]
    except requests.RequestException:
        return answer          # never block the student; log the failure in your own system


if __name__ == "__main__":
    print(label_answer("Photosynthesis turns light energy into chemical energy stored in glucose.", "s-20931"))

Check a submission (Python)

Run it when work is handed in, or on demand from the grading view. It also finds tutor text pasted in the middle of a paragraph.

# Find paragraphs in a submission that carry your tutor's label (Python 3.10+, pip install requests).
# Checking is free for paragraphs without a label; 1 credit for each labelled paragraph found.
import os
import sys

import requests

API = "https://textwatermarking.com/api/watermark"
HEADERS = {"Authorization": f"Bearer {os.environ['TEXTWATERMARKING_API_TOKEN']}"}


def labelled_paragraphs(submission: str) -> list[tuple[int, str]]:
    """[(paragraph number, session id), ...] for paragraphs that contain a tutor answer."""
    found = []
    paragraphs = [p for p in submission.split("\n\n") if p.strip()]
    for number, paragraph in enumerate(paragraphs, start=1):
        r = requests.post(f"{API}/decode", json={"text": paragraph}, headers=HEADERS, timeout=10)
        if r.status_code == 422:           # a mark that is not a plain label: not from your tutor
            continue
        r.raise_for_status()
        label = r.json()["decoded"]
        if label.startswith("ai;tutor;s="):
            found.append((number, label.removeprefix("ai;tutor;s=")))
    return found


if __name__ == "__main__":
    # Usage: python check_submission.py < essay.txt
    for number, session in labelled_paragraphs(sys.stdin.read()):
        print(f"paragraph {number}: tutor session {session}")

Output such as paragraph 2: tutor session s-20931 points the teacher to the conversation. No output means no labelled paragraph was found.

What it costs

What it does not do

• It does not detect other AI tools. Only your tutor's labelled answers are found.
• It does not survive rewording. A student who rephrases the answer, retypes it or runs it through another model removes the label.
• Some editors and file conversions drop the label. Test with the tools your students use, including the format submissions are uploaded in.
• An excerpt that leaves out the first words of an answer has no label. The label sits at the start of each answer.

Frequently asked questions

Related

• Label AI-generated text: streaming, failure handling and the general limits
• API endpoints: /encode, /tokens and /decode
• Quickstart: account, token and first call in five minutes