AI detectors and writing records answer different questions.
A reference page for the limits of detector scores and the limits of writing-process records.
An AI detector reads finished writing and estimates the probability that a machine produced it. Receipts captures observable editing events such as typing, pauses, rewrites, deletions and pastes, then replays them. A detector classifies the finished text. A writing record presents events for a person to inspect and question.
What the published research says about detectors
- A Stanford-affiliated study published in the journal Patterns (2023) tested seven widely used detectors and found they wrongly flagged 61% of TOEFL essays written by non-native English speakers, while performing near perfectly on essays by native speakers. The bias is structural: detectors reward predictable phrasing, and careful second-language writing is often predictable.
- Turnitin has publicly stated it targets a false-positive rate under 1% per document for its AI writing detection, and cautions against relying on scores below certain thresholds.
- Even a 1% false-positive rate across Vanderbilt University's roughly 75,000 annual papers could wrongly flag around 750 papers a year. Vanderbilt disabled its detector in 2023; NTU has announced it will switch its own off at the end of 2026.
Sources: Liang et al., "GPT detectors are biased against non-native English writers," Patterns (2023); Turnitin's published false-positive statements; Vanderbilt University's public guidance on disabling AI detection; NTU's announcement, reported by CNA, 14 August 2026.
Detector scores and writing records
| Comparison point | AI detector (the common products) | Writing record (Receipts) |
|---|---|---|
| Input | The finished text only | Observable editing events recorded while the document is written |
| Output | A probability score | A replay plus facts about how it was written, no score, no definitive verdict |
| Who runs it | The school, after suspicion | The student, before any accusation exists |
| Can it be questioned? | Limited; the scoring model is not open to inspection | Yes; anyone can watch the replay and judge |
| Failure mode | False accusation of an innocent student | An incomplete or ambiguous record stays neutral; a determined person can fake a plausible student-reported record |
| Non-native writers | Elevated false-flag rates documented across the seven detectors tested in Patterns (2023) | No inference is made from vocabulary, grammar, writing style or English fluency |
A Receipts record does carry a colour level. It is a reading order over the rows a teacher can open and read for themselves, saying which record to look at first and which line to open in it. It is not a score, and it is not a number for how the text was made.
What a writing record cannot establish: it cannot prove a student did not read AI output on another screen and retype it. Receipts shows the supplied process without turning smooth, unusual or missing behavior into a cheating inference. The facts may support the student; otherwise they stay neutral or unmeasured. Receipts produces no definitive verdict.
If you are comparing approaches
Detector scores classify finished text. Writing records show how the text was written, for a reviewer to inspect. Some detector companies now also sell writing-process products for institutions. Receipts records event-level activity inside its editor, while a separate Google Docs import can provide periodic snapshot context. The student holds the record, the core workflow stays free for every student, and Receipts servers do not hold the writing.
How accurate are the detectors?
Detector vendors publish low false-positive rates for their own products. The Patterns (2023) study tested seven detectors on 91 essays by non-native English speakers and found a 61% average false-positive rate. That result documents a risk in the tested products, not the rate for every detector. Some detector companies also offer a Google Docs replay built on Google's revision history, which records periodic checkpoints rather than individual input events. Receipts records observable events inside its own editor and produces no score.
What about a grammar company's authorship report?
The closest established product is a grammar company's authorship report, and it is more than a label. It works in Google Docs and Microsoft Word, offers source analytics, a color-coded report, writing time, session counts, text insight cards and a full assignment replay, and integrates with Canvas and Blackboard. That product is stronger today at native editor and institution-wide coverage. Receipts is built for a narrower job: student-owned evidence for a real dispute, with neutral facts, exact jumps to recorded moments, explicit blind spots and no authorship verdict. A direct AI paste may be categorized more specifically by that tool when it can see the source; Receipts records the paste without guessing which tool produced the words. Retyping or device-level automation can look like keyboard input to either product, so neither independently proves who supplied the words. Grammarly's own documentation says attribution may not capture every case accurately.
AI humanizers and writing records
AI humanizer tools paraphrase generated text to lower a detector score. They change the finished sentences, including their rhythm and word choice, after generation.
A writing record does not judge whether finished sentences sound human. If a humanized draft is pasted into a document Receipts is recording, the record shows the block insertion. If it is retyped or entered through device-level automation, the result can look different and may imitate ordinary timing or revision. That is why Receipts makes the facts about how it was written visible but produces no AI-authorship score.
A determined person can fake a plausible student-reported record. Browser-identified page scripting is rejected and unrecorded text is removed before export, but device-level input can still appear keyboard-like. The record is evidence for a teacher to weigh, not a verdict.
How schools can use both
A school may use a detector as a screening signal and a writing record as separate context for human review. Neither should be treated as proof on its own.
See the workflow
Open a made-up record, read the recorded facts, then inspect a moment in the replay. No account is needed.
Open a sample record Read the teacher guide
Receipts. Show how it was written. Operated by Receiptsproof Pte. Ltd. Home · Accused of AI? · For educators · For schools © 2026 Receiptsproof Pte. Ltd.