Head to head · Detector v10 · October 2026
ParaTrace against Pangram
21 results on the same public tests, with Pangram’s figures as Pangram published them and ours measured on the same texts.
ParaTrace catches 100 % of fully AI-written student essays, the same as Pangram, and misses fewer AI texts than both Pangram 4 and Pangram 3 in a test of writing by well-known authors. Across the 21 results measured on the same tests, it is within 5 points of Pangram 4 on 17 and of Pangram 3 on 18. The remaining gaps are on humanized text, above all under 50 words, and on very short texts, where it is just over 5 points behind Pangram 4.
Within 5 points
- Against Pangram 417 of 21 results
- Against Pangram 318 of 21 results
Results where ParaTrace is level, ahead, or no more than 5 points behind.
Strengths
Where ParaTrace matches or beats Pangram
Ahead on well-known authors
0.84 %
of AI texts missed in a test of writing by well-known authors, against 2.86 % for Pangram 4 and 5.05 % for Pangram 3.
Level on student essays
1.000
AUROC on all 4 student-essay tests, the same as Pangram 4, with 100 % of fully AI-written essays caught.
Ahead of Pangram 3 on short texts
94.66 %
of AI texts under 50 words caught at 1 % false alarms, against 88.87 % for Pangram 3.
The breakdown
All 21 results, test by test
| Result | Pangram 4 | ParaTrace | Gap |
|---|---|---|---|
| Fully AI-writtenAUROC | 1.000 | 1.000 | Level |
| Fully AI-writtenAI caught at 1 % false alarms | 100 % | 100 % | Level |
| AI-paraphrasedAUROC | 1.000 | 1.000 | Level |
| AI-paraphrasedAI caught at 1 % false alarms | 100 % | 98.9 % | −1.1 |
| Polished by AIAUROC | 1.000 | 1.000 | Level |
| Polished by AIAI caught at 1 % false alarms | 100 % | 98.9 % | −1.1 |
| OverallAUROC | 1.000 | 1.000 | Level |
| OverallAI caught at 1 % false alarms | 100 % | 99.4 % | −0.6 |
| Result | Pangram 4 | ParaTrace | Gap |
|---|---|---|---|
| All textsAUROC | 1.0000 | 0.9999 | −0.01 |
| All textsAI caught at 1 % false alarms | 100 % | 99.94 % | −0.06 |
| Result | Pangram 4 | ParaTrace | Gap |
|---|---|---|---|
| Full lengthAUROC | 1.0000 | 1.0000 | Level |
| Full lengthAI caught at 1 % false alarms | 100 % | 99.91 % | −0.09 |
| Under 50 wordsAUROC | 0.9999 | 0.9965 | −0.34 |
| Under 50 wordsAI caught at 1 % false alarms | 99.70 % | 94.66 % | More than 5 points behind, −5.04 |
| Humanized, full lengthAUROC | 0.9996 | 0.9959 | −0.37 |
| Humanized, full lengthAI caught at 1 % false alarms | 98.93 % | 92.03 % | More than 5 points behind, −6.90 |
| Humanized, under 50 wordsAUROC | 0.9810 | 0.8778 | More than 5 points behind, −10.32 |
| Humanized, under 50 wordsAI caught at 1 % false alarms | 73.32 % | 36.43 % | More than 5 points behind, −36.89 |
| Human writingHuman texts wrongly flagged | 0.00 % | 0.10 % | −0.10 |
| Result | Pangram 4 | ParaTrace | Gap |
|---|---|---|---|
| AI textsAI texts missed | 2.86 % | 0.84 % | ParaTrace ahead, +2.02 |
| Human textsHuman texts wrongly flagged | 0.00 % | 1.62 % | −1.62 |
The gap is in points, ParaTrace’s figure against Pangram’s, so a minus means ParaTrace is behind. ▲ marks where ParaTrace is ahead, and ▼ where it is more than 5 points behind. For AUROC, one point is 0.01.
How we compared
Like for like
- Only matching results.
- Every Pangram figure here is one Pangram published for a public test, in its Pangram 4 technical report of July 2026. We scored ParaTrace on the same texts and measured it the same way, and left out every result we could not match exactly.
- AI caught at 1 % false alarms.
- Each detector is set so that 1 in 100 of the test’s human texts is flagged, and we count the share of AI texts it then catches.
- AUROC.
- How well a detector ranks AI text above human text across every possible setting, from 0.5 for a coin toss to 1 for perfect.
- Missed and wrongly flagged.
- Each detector at its standard setting. For ParaTrace that is its Strict setting, applied to each text as a whole.
- Points.
- Gaps are in percentage points, and a minus always means ParaTrace is behind, even where a lower figure is better. “Level” means the same figure at the precision Pangram published.
- Versions.
- ParaTrace is detector v10, the model this site runs, scored in October 2026 with the service’s model and scoring code on our own hardware. Both Pangram versions are as Pangram reports them. For the test on well-known authors, Pangram 3 is version 3.3.2.
- About Pangram.
- Pangram is a product of Pangram Labs, Inc., which has not reviewed this comparison.
- More results.
- Our benchmark report covers many more models, evasion attacks and kinds of writing, and a field test on scientific papers repeats the check NeurIPS ran with Pangram.