November 2026. Every number on this page is measured on text where we know who wrote it, unless it says "target". Nothing here is live until it beats what is live now on the same tests.
What improved
Before and after numbers for the same test each time. The reducer's test paper is our own 6,400-word project report, run through every version.
Results in graphs
Reducer versions
| Version | What changed | Measured |
|---|---|---|
| v1 · 25 Oct | First release: rewrites the most flagged sentences in rounds, re-scores after each round. | Test paper 37.2% → 20.1% in 31 min |
| v2 · 26 Oct | Every change is judged on all five sentence windows it touches; a round that raises the score is undone; a final sentence-level pass on long sentences. | A customer paper 55.2% → 10.7%; no run ends higher than it started |
| v3 · 27 Oct | A second final pass chosen by the detector, kept only if the whole paper scores lower; targets set by starting score. | 6 of 7 reopened older runs dropped a further 4 to 9 points |
| v4 · 28 Oct | Neighbouring sentences rewritten together; scoring split across two servers; rounds chained back to back. | Test paper 37.2% → 10.2% in 32 min (v1: 20.1% in 31 min) |
| Next | Long flagged stretches of five to eight sentences in heavy papers; every change also scored by detectors it was not tuned on. | In testing, not live |
The two reducers
The same Rs 30 unlock gives you both. They differ in how much checking they do.
| Full reducer (website) | Quick reducer (inside Claude or ChatGPT) | |
|---|---|---|
| Where | Press Reduce on your report at paper-check.me | The reduce_ai tool over MCP, from your AI chat |
| How much it rewrites | The whole paper, in rounds, until the target or no more progress | Up to 40 flagged sentences per call, in groups of up to six neighbours |
| Checking | Every change re-scored on all five windows it touches; kept only if they go down, undone otherwise; then whole-paper passes | Each rewritten sentence comes back with its score before and after; you run the check again for the paper score |
| Time | Usually a few hours; result by email | Seconds |
| Measured | 50+ customer runs: average drop 18.6 points, none went up | Not yet measured on whole papers |
How the full reducer works
Detector versions
| Version | What changed | Measured |
|---|---|---|
| r5 · 16 Oct | Our own fine-tuned model replaces the open detectors for the headline score. | AI text caught on the TextSight test set: 19% |
| r5 · 17 Oct | Flag threshold re-chosen on held-out human writing. | TextSight caught 59%; every human group at or under 3% wrongly flagged |
| v9ft · 18 Oct | Fine-tuned further on Claude-written papers and more human academic text. | Claude papers caught 53% → 97%; human papers wrongly flagged 0.4% |
| v10 · 31 Oct | Trained on the RAID benchmark. | RAID 53% → 96%, but 12.6% of human papers wrongly flagged: not shipped |
| v11 · training | Three times more human academic writing than AI text, plus a round that retrains on the human papers it wrongly flags. | Ships only if it beats v9ft on every row below |
Where things stand today
The detector was built for one job: student and research papers. On that job it rarely accuses a human writer, and on ten papers with real Turnitin reports it landed within 5.4 points on average. Turnitin is not ground truth, so we treat that as a sanity check, not a goal.
Outside that job it is weaker. On RAID, a public benchmark that mixes news, Reddit posts, recipes, poetry and deliberately disguised AI text, it catches 53% of the AI writing at a 5% false-alarm rate, and 48% of the disguised part. That is the gap we are closing first.
Next for the detector: harder to fool
We already tried the obvious fix. Training on RAID took the benchmark score from 53% to 96%. It also started flagging about 1 in 8 genuine human papers. For a tool students use before they submit, that is worse, so it never went live.
The lesson: the new model saw a lot of disguised AI text and too little real academic writing by humans, so formal human prose started to look like AI to it. The next model fixes the balance.
| Live model | Tried, not shipped | Target for the next model | |
|---|---|---|---|
| Human papers wrongly flagged | 0.4% | 12.6% | under 1% |
| AI text caught on RAID (5% false alarms) | 53% | 96% | as high as possible without breaking the row above |
| Disguised AI text caught (RAID attacks) | 48% | 94% | same |
| AI-edited human text caught | about 50% | 58.5% | at least 50% |
How: about three times more human academic writing than AI text in training, a fresh test set of human papers that the model never sees, and a loop that finds the human papers it wrongly flags and trains on them again. It ships only if it beats the live model on every row.
Next for the reducer: the hardest papers
The reducer works well when a paper starts moderately. It struggles when most of a paper reads as AI.
| Starting score | Median drop | Median end score | Best measured |
|---|---|---|---|
| Under 35% | 11.8 pts | 13.3% | 29.7% to 11.4% |
| 35 to 55% | 16.7 pts | 27.3% | 49.5% to 13.4% |
| 55% and above | 20.1 pts | 51.0% | 89.3% to 20.1% |
Customer runs finished by 1 November 2026, our own test runs excluded. Every one went down; none went up. Heavy papers drop the most points but still end around 50%. What we are working on:
- Heavy papers. Long stretches of five to eight sentences in a row stay flagged even after rewriting. We are testing new approaches on our own papers before any customer sees them.
- Checking against more than one detector. Today the reducer is measured on our own score. Different checkers disagree, so every change is now also scored by detectors it was not tuned on, and a change only counts if those agree.
- Keeping what the paper says. Every rewrite is checked automatically: a dropped number, citation, equation or technical term fails the rewrite. Anything added has to be confirmed.
- Telling you what is left. When some parts still read as AI-written, the result email lists those stretches, because rewriting them in your own words lowers the score the most.
How we decide something is better
- Test on text where we know who wrote it, never only on agreement with another tool.
- Count false alarms on human writing first, per kind of writing (papers, encyclopedia text, non-native English, learner essays), and only then count how much AI is caught.
- Split test and training data by document, never by chunk, and pick thresholds on separate validation documents.
- Withdraw any number we cannot reproduce.
Limits
No detector is proof, ours included, and other checkers will give different numbers. The reducer lowers an estimate; it does not guarantee any score on any other tool. The numbers above are from small test sets (ten Turnitin papers, 50+ reducer runs) and will be updated as they grow.
How the system was built is on the engineering notes page, and the full system is drawn on the architecture page.