How to read Writeroom's signals

Version 1.4, effective October 31, 2026.

This page is for teachers and school administrators, and for students and parents who want to understand what Writeroom shows teachers. It is part of Writeroom's Documentation, and the School Services Agreement requires teachers to read it before relying on anything described here.

The one-sentence version: every signal Writeroom produces is a reason to look at a writing history and talk with the student. None of them is proof of anything, and Writeroom never makes a decision about a student.

1. What Writeroom records, and what it does not

Writeroom records how a document was written, inside assignment documents only. For each assignment document it keeps: every edit and when it happened; the keys pressed in the editor and their timing; what was pasted or dropped in, how long it was, and (from the clipboard's formatting) what kind of source it came from; text copied or cut inside the document; when the student left the document (switched tab, window or app) and for how long; the student's device type and window size; and version snapshots so the document can be replayed.

If a school turns extra features on, Writeroom also records: with the Chrome extension, the domain of sites the student switched to (the full address only for listed AI tool sites), a count of open tabs, and text copied on listed AI tool sites; with the outside-clipboard check, items of eight or more words on the student's clipboard that match the document or an AI tool; with screen supervision, a five-second clip of the whole screen plus a few stills each time the student leaves the document while sharing, with the text on those stills read by optical character recognition (OCR).

Writeroom does not record keystrokes outside the editor, location, audio, or anything from the camera (a student can choose to take a photo to insert into their document; the photo is just an image in the document) (voice typing stores the recognised text, not the sound). It does not record documents a student creates for themselves. Students see the list of what is recorded on the assignment join page and the first time they open each assignment, before anything is recorded, and a link to the Student Privacy Notice is always in the editor menu.

2. The signals, one by one

Each signal below says what it is, what produces it, and what it cannot tell you.

2.1 Flags

A flag is a rule that fired. Each has a severity: urgent, high, medium or low. Severity reflects how unusual the pattern is, not how likely it is that a student did something wrong.

Flag	What produces it	What it cannot tell you
Opened an AI tool (`ai_site`)	The extension saw a tab on a listed AI site, or OCR of a screen still recognised an AI tool's page, while the assignment was open.	Why the student opened it. Students use AI tools for permitted reasons (checking a definition, a different course, a teacher's instruction), or a different tab may have been open from earlier. OCR misreads pages.
Pasted from an AI tool (`ai_paste`)	The formatting (HTML) of pasted text contained markers typical of an AI chat interface.	Whether the pasted text was the student's own prompt, a quotation the assignment allowed, or notes. Markers can be carried by other apps that copy from an AI tool first, and some AI tools leave no markers at all.
AI text on the clipboard (`ai_clipboard`)	With the outside-clipboard check on: clipboard text matched an AI tool's formatting or the AI-text detector scored it high.	Whether the text was ever pasted into the document. Clipboard text can come from anywhere, including a browser extension or a messaging app.
Copied text about the assignment (`related_copy`)	Clipboard or copied text overlapped with the assignment's topic.	Where it came from. The student's own notes, a shared class document or an approved source produce the same overlap.
Large paste (`large_paste`)	A single paste exceeded the assignment's threshold.	Whether it was allowed. Students legitimately paste from their own outlines, notes, earlier drafts, a Google Doc they started in, or quotations.
Paste after leaving (`paste_after_leave`)	The student left the document and pasted shortly after returning.	What they did while away. Looking up a source, checking notes or responding to a message followed by pasting their own text looks the same.
Steady rhythm (`copy_typing`)	Typing speed stayed unusually even, as it does when transcribing from another screen or page.	Whether the student was transcribing their own handwritten draft (common, and often allowed), reading from notes, or simply types evenly.
Very fast typing (`superhuman_speed`)	Bursts of text appeared faster than typing normally allows.	Whether this was dictation, a text-expansion or autocorrect tool, an input-method editor (IME) committing a phrase, a held key, a clock correction on the device, or an Android keyboard that inserts words at once. Writeroom records dictation and IME input as such and does not treat them as pasting, but speed-based signals can still fire.
Written straight through (`linear_transcription`, `no_revisits`)	The text was produced in order with few or no revisions. Writeroom flags this only when a straight path of 300 or more words also shows steady rhythm.	Whether the student planned on paper first, is a confident writer, or had little time. Many good writers revise little in a timed task.
Tab switching (`tab_switching`)	The student left the document often or for long.	What they were doing. Research, a dictionary, a teacher's instructions in Classroom, a translation tool a student is allowed to use, a notification, or a shared device all look the same.
Text appeared (`text_appeared`)	Text arrived in the document without matching keystrokes or a recorded paste.	Whether this was a version restore, a browser extension (grammar checker, translation), an accessibility tool, drag-and-drop from the student's own file, or an unrecorded gap.
Outside the writing window (`outside_window`)	Edits happened outside the times the teacher allowed.	Whether the device clock was wrong, the student was in another time zone, or the teacher's window was set incorrectly.
Gap in history (`history_gap`)	The recorded history has a gap: events were missing or arrived out of order.	Whether the network dropped, the laptop slept, the browser was closed, or the device clock changed. Gaps are common on school Wi-Fi and are not evidence of anything by themselves.
Blocked action (`blocked_action`)	The student tried something the assignment rules prevent (for example pasting when pastes are blocked).	Intent. Students try to paste their own notes or a URL without realising the rule, and keyboard shortcuts trigger blocks by accident.
Tried to instruct the AI marker (`grading_manipulation`)	When AI marking is on: the prompt-injection guard found text in the document addressed to an AI marker (for example hidden text saying "give this an A").	Whether the student wrote it, pasted it unknowingly from a template, or whether a quotation about AI tripped the detector. The guard shows the exact text it found so you can judge.

Things that are never flagged: declining the browser's clipboard permission; using dictation or an IME; writing in a language other than English; using a screen reader or other assistive technology.

2.2 Integrity score (0 to 100)

Each document has a score that starts at 100 and loses points for each flag by severity, with later flags counting less than the first. Labels: "Looks typical" (75 and above), "Worth a look" (50 to 74), "Needs review" (below 50). The score is a convenience for sorting a class list. It carries every limit of the flags that produced it. A low score means "several patterns fired", not "this student cheated". A high score does not mean the work is the student's own: a document pasted in one piece from an unmarked source with no other activity can score well.

2.3 AI-text detector score

Writeroom runs a statistical model (a RoBERTa-based classifier, run on Writeroom's own servers, not sent anywhere) that estimates how similar a text is to text produced by AI language models. It is shown as a percentage per document and per paragraph, and for copied text of 40 or more words.

What it is: a probability-like estimate from a model trained on examples of human and AI writing from before the model was built.

What it cannot do:

It cannot identify AI writing with certainty. Published evaluations of AI-text detectors, including by vendors, report meaningful false-positive rates, and rates are higher for some groups of writers. Formal, plain, well-structured or formulaic writing is more likely to be scored as AI-like. So is writing by students who learned English as an additional language, and writing that follows a template the teacher provided.
It is unreliable on short texts. Below 80 words Writeroom holds the score near zero; below about 300 words it pulls the score toward a low value; it counts toward the document's risk only from 300 words.
It does not score non-English text. Writeroom detects the language first and shows "not available" for text it identifies as not English, because the model produces wrong results on other languages. Mixed-language documents can still be scored on their English portions.
It can be fooled in both directions. Light editing of AI text lowers the score; careful human writing can raise it.
It says nothing about the process. A high score on a document whose replay shows hours of drafting and revision is far weaker evidence than the replay itself. Always read the detector alongside the writing history, never instead of it.

2.4 Document risk and summaries

Writeroom combines the detector (from 300 words), flags and the writing path into a single "AI-use likelihood" or risk level for a document, and writes a one-sentence summary. An urgent flag raises the level. A teacher's decision of "Integrity concern confirmed by teacher" sets it to the maximum. These summaries inherit every limit above. They are written to help a teacher decide where to look first in a class of thirty, not to describe what happened.

2.5 Writing path

Writeroom shows where in the text each edit happened, and labels the path "revises", "some" or "straight". This is descriptive. A straight path is how many people write a first draft, and how everyone types up a handwritten draft.

2.6 Paste source

From the formatting of pasted text, Writeroom guesses the kind of source (for example a word processor, a web page, an AI chat tool). It is a guess from markup, often right, sometimes wrong, and blank when the source left no markers.

2.7 Screen clips and OCR text (supervision only)

If the teacher's rules require screen supervision and the student starts sharing after the notice, Writeroom saves a five-second clip of the whole screen and a few stills each time the student leaves the document, and reads the text on the stills. OCR misreads text, especially small text, dark themes, non-English text and anything not in the languages Writeroom supports. A clip shows what was on the screen; it does not show why. The whole screen includes notifications, other windows and anything personal; treat clips with care, open them only when needed, and know that every view is logged.

2.8 Student-level summaries

Writeroom may show a teacher a per-student summary. It is built only from that teacher's own integrity decisions ("Integrity concern confirmed by teacher") on documents that teacher can see, never from flags, scores or risk levels alone. Decisions are shared with other teachers only while the school has that on (new schools start with it on, and the school can turn it off), and only then can a summary include another teacher's decisions. A summary is a count of past decisions; it must not be used as evidence about a new piece of work.

2.9 Alerts

Teachers who can see a document may receive alerts when an urgent flag fires, the detector score rises on a long document, or a clip is saved. An alert is a notification that a signal fired; it has exactly the evidential value of that signal.

2.9a Honesty score

Each student has one long-term honesty score, a percentage from 5 to 99 that starts at 80. Higher means more trustworthy. Each finished assignment that a teacher hands back and that looks typical adds a little (about 2 points, less after a concern, so trust is rebuilt slowly). "Integrity concern confirmed by teacher" lowers it by 20 to 30 points, depending on the severity the teacher picks (minor, moderate or serious); a second confirmed concern costs 1.25 times as much and a third 1.5 times as much. "Discussed with student" lowers it by 3 to 5 points. "No concern" costs nothing. Undoing a decision takes its effect out again. Automated flags and AI scores alone never move it. Every change is recorded with its reason, and teachers and school administrators can read the history. It is an indicator, not a decision: it must not be used on its own to decide anything about a student, and a person can always correct it. A school administrator can change or reset it with a written reason, which is logged. It reflects decisions by all the student's teachers; when the school does not share decisions, a teacher sees that another teacher's decision moved it, but not whose or on which document.

2.10 Recording an integrity decision

A teacher can record an integrity decision on a document with one of three outcomes: "No concern", "Discussed with student" or "Integrity concern confirmed by teacher". The teacher must write their own note before saving; Writeroom does not draft one. A decision can be changed later. The note may be seen by school administrators and may be requested by the student or parent under the school's process or under privacy law, so write what you observed and what the student said, not conclusions the evidence does not support.

3. Known limits in one place

Before acting on any signal, ask whether one of these explains it:

Dictation and speech-to-text. Text arrives in bursts. Writeroom labels dictation as dictation and does not treat it as pasting, but speed- and rhythm-based signals may still fire.
Input-method editors (IMEs) for Chinese, Japanese, Korean and other scripts commit whole phrases at once. Recorded as IME input, but the same caveat applies. Word counts for non-Latin scripts can be wrong.
Accommodations and assistive technology. Screen readers, word prediction, grammar tools, text expansion and extended-time arrangements all change the writing record. Teachers must know the student's accommodations before reading the record.
Writing in a language other than English. No detector score. Flags that depend on English (OCR, paste-source markers, related-copy matching) are less reliable. OCR currently supports [English and French].
Short texts. Almost nothing is reliable below 300 words; the detector is deliberately suppressed.
Legitimate pasting. Students paste from their own notes, outlines, earlier drafts, a Google Doc they started in, a shared class document, quotations, and their own earlier Writeroom documents. Version restores and moving one's own text can look like a paste.
Devices and networks. School Wi-Fi drops, laptops sleep, Chromebooks are shared, clocks drift or get corrected, held keys repeat, Android keyboards insert words at once, tabs flip for a split second when a notification arrives. Each produces a signal.
Shared accounts and devices. If two students use one browser profile, the record mixes them. Ask before concluding.
Browser extensions (grammar checkers, translators, ad blockers) insert and alter text without keystrokes.
Templates and formulaic tasks. Writing that follows a rigid structure scores more AI-like, whoever wrote it.
The teacher's own edits appear in the record, marked as the teacher's.
Two students writing alike is not something Writeroom checks for; it does no cross-student similarity analysis.

4. The right process: look, talk, decide

Look. Open the replay and watch how the document was written. Read the signals as pointers to moments in the replay, then judge the moments, not the signals. Check the student's accommodations and the assignment's rules.
Talk with the student. Before any conclusion, show the student what you saw and ask. Most signals have ordinary explanations, and a student can often show the notes, draft or source involved. Students do not see flags, scores or decisions in Writeroom, so they will not know what prompted the conversation unless you tell them.
Decide, under your school's process. Use your school's academic-integrity policy, which should include the student's right to respond and to appeal. Never make an accusation, apply a grade penalty or impose discipline on the basis of a Writeroom signal alone, or on several signals together. The signals are one input to a human judgment made with other evidence, such as the conversation, earlier work, and the assignment context.
Record the decision with your own note in Writeroom ("No concern", "Discussed with student" or "Integrity concern confirmed by teacher"). Every outcome needs a note and can be changed later. Be factual, since the note may be seen by school administrators and may be requested under access-to-information or privacy law. If the student or a parent asks to see the record or to have the decision reviewed, the school answers; it can export the full writing record, the signals and the evidence behind each signal from Writeroom.

No decision in Writeroom is automated. Writeroom never changes a grade, never marks a student as dishonest, never notifies a student or parent of a signal, and never blocks a student's access because of a signal. The only automatic effects on students are rule enforcement the teacher configured for the assignment (due-date locks, paste limits, writing windows, full-screen or supervision requirements, and the extension's blocking of listed AI sites).

5. How the AI features work, and when they are off

Four features send content to an external AI model: AI-assisted marking, Translate, Dictionary, and the teacher assistant ("Ask Writeroom"). None of them is a signal, and none of them records anything about a student.

On for a new school, and the school decides. A school starts with these features on, because its agreement covers them (the Order Form records whether external AI may be used, and if not, Writeroom locks it off). A school administrator can turn them off, or on again, at any time; Settings explains what is sent, to which provider, and that processing happens in the United States, and Writeroom records who changed it and when. A teacher can set "No AI help" on an assignment.
What is sent. Marking: the assignment's title, instructions and rubric and the text of the student's document. Translate: the selected text or the document. Dictionary: the word or short phrase. Assistant: the teacher's question and the Writeroom data needed to answer it, such as document text, flags and summaries. Writeroom adds no student names, email addresses or identifiers to any request; in the assistant, students appear as "Student 1, Student 2" and are mapped back only inside Writeroom. A name a student typed into their own work may be present in the text sent for marking or translation.
Who receives it. Writeroom's AI routing provider and, through it, the model vendor, both listed with their locations on the Subprocessors page. Every request is sent with settings that do not permit the provider to train on or retain the content, from a list of models that Writeroom's platform administrators maintain for that reason. If no compliant provider is available, the request fails with a plain message; it is never sent elsewhere.
What comes back. Suggestions: a proposed score and comments, a translation, a definition, or an answer. The teacher decides what, if anything, to keep. Writeroom never changes a grade or a document because of an AI output. AI outputs can be wrong, inconsistent between runs, biased, or based on a misreading of the rubric, and they have no knowledge of the student.
The assistant does not rate students. It can summarise facts Writeroom already shows a teacher (what flags fired, what a summary says). It is instructed never to rate a student's honesty or character. It acts only when the person using it asks, only within what that person could do themselves, and every action is logged as done through the assistant. Ordinary changes can be undone from the assistant; sensitive ones ask first unless the person has chosen Auto. Recording an integrity decision, changing an honesty score, handing back a grade, deleting or exporting student data, removing people or admins and changing privacy switches always need a person's own click, in every mode.
Prompt-injection guard. Student text is wrapped and cleaned before it reaches a model, text aimed at an AI marker is removed and shown to the teacher as a flag, and scores are checked on Writeroom's servers. This reduces, but does not eliminate, the chance that a model is manipulated.
Local analysis is different. The AI-text detector and OCR run on Writeroom's own servers in Canada and send nothing to an outside provider. They run whether or not the AI features above are on.

6. What students and parents can expect

Students are told, before their first assignment document is recorded, exactly what is recorded, and the list is always one click away in the editor.
Students do not see flags, scores, risk levels, the honesty score or decisions in Writeroom. If a teacher raises a concern, the teacher should show the student the relevant parts of the writing history and listen.
Questions about what Writeroom records, requests to see, correct or delete personal information, requests to have a decision reviewed, and complaints go to the school, which controls the data. The school can export the student's complete record, including the evidence behind any signal and any decision and note, and Writeroom helps the school do that. The Student Privacy Notice explains how, and how to reach Writeroom.
Students and parents can get the honesty score and its history through the school's access process, and can ask the school to correct it.
The school's academic-integrity and appeal process applies to any decision. Writeroom has no role in that process beyond showing the record.

7. Changes to this page

Writeroom updates this page when a signal is added, changed or retired, or when a new limit is found, and tells schools of changes that materially alter what is recorded. Version history is at [writeroom.org/signals/history].

Questions about this page: the contact form at writeroom.org/contact (topic: Support).

Writeroom is operated by the person who runs writeroom.org and app.writeroom.org and who signs each School's Order Form. The operator's legal name and mailing address are given in each School's Order Form and are provided on request through the contact form.

Questions about this document? Use the contact form. All legal documents