How Colophon works
Colophon shows you and your teacher exactly how a piece of writing came together: your edits, your pastes, your AI use. Today, a course's case study is how you get it. If your class separately agrees, an anonymized version of the same data can also inform a published research report. That part is optional, and never assumed from enrollment alone.
FIRL (Functional Intelligence Research Lab) is an independent research lab building open standards for how AI use in writing gets recorded and disclosed. Colophon is its first tool, built to make that process visible without turning it into surveillance.
1. Bring Colophon to your course
A teacher or researcher applies via our case-study form with a course, section size, and semester. We review every application ourselves.
2. We set up your course
Once accepted, we activate your case study, which creates your course in Colophon and a join code for your students right away.
3. Students join, and separately choose about research
Each student joins with the code. Using Colophon in the course is one thing; agreeing to have your pseudonymized data included in the published research report is a separate, later choice, never implied by joining. See our consent page for the full breakdown.
4. You write, Colophon records the process
The Chrome extension records the writing process locally, exactly as described on our what gets recorded page: activity, not content. You can see it happening and turn it off any time. See what Colophon does not do.
5. You export and submit it yourself
You export your session as a TWFF file and submit it yourself while signed in. We check that the file is intact and hasn't been altered before it's recorded; a tampered or invalid file is rejected, not silently accepted.
6. If you opted into research, it's pseudonymized and aggregated
Names and identifying details are replaced with a stable, one-way pseudonym before anything leaves an individual record, never reversible, but consistent across the case study. The researcher dashboard only ever shows cohort-level patterns, never a named student's document.
7. We publish what we learn
Every case study that includes the research layer ends in a findings report you can read and share. A broader, anonymized dataset organized by research topic is the eventual goal; that release mechanism doesn't exist yet, and we'd rather say so than promise a date.
Read our full ethics and privacy commitments →See it with real sample data
The following is illustrative sample data generated for demonstration purposes. No real student data has been collected.
Interaction types tracked: typed-by-you, AI-accepted, pasted content, AI-paraphrased.
Typed by you made up 49% of tracked activity on average.
Class average,
weighted by activity
Consumer Behavior Research · 12 May 2025
Market Analysis Draft · 5 May 2025
Sustainable Packaging Proposal · 28 April 2025
Executive Summary · 14 April 2025
Implementation Plan · 7 April 2025
Competitor Benchmarking · 31 March 2025
Solution Concepts · 24 March 2025
Literature Review Draft · 10 March 2025
| Class average | |
|---|---|
| Typed by you | 49% |
| AI-generated | 27% |
| Pasted sources | 9% |
| AI-paraphrased | 15% |
| Consumer Behavior Research (12 May 2025) | |
| Typed by you | 0% |
| AI-generated | 53% |
| Pasted sources | 18% |
| AI-paraphrased | 29% |
| Market Analysis Draft (5 May 2025) | |
| Typed by you | 35% |
| AI-generated | 36% |
| Pasted sources | 10% |
| AI-paraphrased | 19% |
| Sustainable Packaging Proposal (28 April 2025) | |
| Typed by you | 84% |
| AI-generated | 8% |
| Pasted sources | 4% |
| AI-paraphrased | 4% |
| Executive Summary (14 April 2025) | |
| Typed by you | 1% |
| AI-generated | 51% |
| Pasted sources | 20% |
| AI-paraphrased | 28% |
| Implementation Plan (7 April 2025) | |
| Typed by you | 42% |
| AI-generated | 31% |
| Pasted sources | 11% |
| AI-paraphrased | 16% |
| Competitor Benchmarking (31 March 2025) | |
| Typed by you | 46% |
| AI-generated | 27% |
| Pasted sources | 12% |
| AI-paraphrased | 15% |
| Solution Concepts (24 March 2025) | |
| Typed by you | 55% |
| AI-generated | 25% |
| Pasted sources | 7% |
| AI-paraphrased | 13% |
| Literature Review Draft (10 March 2025) | |
| Typed by you | 92% |
| AI-generated | 5% |
| Pasted sources | 0% |
| AI-paraphrased | 3% |
How is this calculated?
Fields used: fixture ai_rate, edits, pastes
Method: _demo_composition() splits ai_rate into AI-generated/paraphrased (65/35) and derives a paste share from pastes/edits; own-writing is the remainder. Class average: the same shares, weighted-averaged across documents by edit count.
Derived from fixture fields chosen for this demo, not measured.
2 of 9 documents (22%) were majority AI-accepted (≥50%).
| Literature Review Draft, Packaging Materials Report | 5% |
| Sustainable Packaging Proposal | 10% |
| Competitor Benchmarking, Solution Concepts | 25% |
| Implementation Plan | 30% |
| Market Analysis Draft | 35% |
| Executive Summary | 50% |
| Consumer Behavior Research | 55% |
How is this calculated?
Fields used: Same AI-generated share as the Composition tab
Method: Each document's AI share, bucketed to the nearest 5%. Documents sharing a bucket collapse into one dot sized by count, never stacked at different heights — height would otherwise imply a second variable that isn't there. Documents at or above 50% counted in the callout.
A proxy for reliance — see the Composition tab's note.
AI-usage rate rose from 8% in Week 1 to 83% by Week 9.
| Week 1 | 8% |
| Week 2 | 8% |
| Week 3 | 38% |
| Week 4 | 42% |
| Week 5 | 47% |
| Week 6 | 79% |
| Week 7 | 12% |
| Week 8 | 55% |
| Week 9 | 83% |
How is this calculated?
Fields used: fixture ai_rate, date
Method: Average ai_rate per ISO week of the fixture date field. Weeks are numbered sequentially from the first week with data, not by calendar week number.
A real case study plots the same shape from real accept/dismiss counts and session dates.
Most of the class (44%) used AI mainly for paraphrasing.
| Paraphrasing | 4 documents (44%) |
| Grammar | 3 documents (33%) |
| Brainstorming | 2 documents (22%) |
How is this calculated?
Fields used: new fixture primary_ai_use (categorical)
Method: Count of documents per primary AI-use category (paraphrase/grammar/brainstorm), authored directly per document — not derived from any other field, since nothing else in the fixture predicts which UI action was used.
Demo-only: no field currently tracks which UI action produced an AI interaction. The three categories match the sidepanel's real quick-action buttons (paraphrase/grammar/brainstorm); this breakdown will become available once that AI-action data is tracked.
On average, AI-assisted text scored 10 points higher in complexity than students' own writing.
Colophon doesn't analyze document text today.
| Consumer Behavior Research | own: 57 | AI-assisted: 78 | gap: +21 |
| Executive Summary | own: 50 | AI-assisted: 70 | gap: +20 |
| Market Analysis Draft | own: 44 | AI-assisted: 58 | gap: +14 |
| Implementation Plan | own: 49 | AI-assisted: 61 | gap: +12 |
| Competitor Benchmarking | own: 41 | AI-assisted: 51 | gap: +10 |
| Solution Concepts | own: 46 | AI-assisted: 56 | gap: +10 |
| Sustainable Packaging Proposal | own: 52 | AI-assisted: 55 | gap: +3 |
| Literature Review Draft | own: 48 | AI-assisted: 50 | gap: +2 |
| Packaging Materials Report | own: 55 | AI-assisted: 57 | gap: +2 |
How is this calculated?
Fields used: new fixture vocab_own, existing ai_rate
Method: vocab_own is an authored illustrative baseline per document; AI-assisted value = min(100, vocab_own + round(ai_rate * 0.25)); the gap between them is labeled directly on each row.
Demo-only. Colophon doesn't analyze document text/vocabulary today; this comparison is planned for a future version.
6/6
Used AI at least once
89%
Documents included a paste
No individual names or documents; real case studies export the same shape, pseudonymized. What does Colophon actually record? →
What the optional research layer studies
Provenance of Human-AI Knowledge Work
How people actually incorporate AI into writing, and what information is useful to authors versus educators.
Human-AI Collaboration
How AI should support writing without replacing authorship, and what interactions build confidence rather than dependency.
Computational Reproducibility of Writing
Whether the writing process can be made inspectable and reproducible, much like code is with version control, and what standards real interoperability needs.