Organizations: School of Artificial Intelligence, Beijing Normal University, Beijing 100875, China · Institute of Artificial Intelligence and Future Networks, Beijing Normal University, Zhuhai 519087, China
Web platforms screen posts for text written by large language models, and every false flag hits a human writer. Detectors must therefore run at a very low false positive rate (FPR). Zero-shot detectors such as Binoculars score text with frozen language models, but invisible edits such as homoglyphs shatter tokenization and push machine text into the human range. Unicode normalization undoes these edits, yet the normalized score still misses half of the machine text at a strict FPR, under attacks that leave nothing to clean. We find that with the Falcon-7B scoring pair, heads on token statistics recall less than the normalized score at this FPR, and several flag too many human posts in some sections. We propose NormDetect, which anchors a learned per-section head to the normalized zero-shot score. A blend weight chosen on training folds under a human FPR cap combines the two scores, and one threshold is calibrated on human text only. On eight sections of the RAID benchmark with eleven attacks, NormDetect raises recall at a 0.1% calibrated FPR by 5.0 points over normalized Binoculars. Its human FPR bound stays at most 1% in every section, and no other detector within this bound recalls as much.
Figures & tables
Figure 1. A machine-written review with look-alike letters at a 0.1% calibrated FPR. Raw and normalized Binoculars score it below τ , and NormDetect flags it. A machine-written review with look-alike letters goes through three detectors drawn as score axes with a threshold. Scoring the raw text and scoring the normalized text leave it below the threshold, and NormDetect flags it.
Figure 2. Overview of NormDetect. A per-section head is anchored to the normalized Binoculars score z , with κ∗ set on training folds and τ on human calibration text. Four-panel pipeline: text normalization and two frozen scoring models, an anchored blend of a per-section head with the normalized Binoculars score, an offline choice of the blend weight, which keeps the best weight whose human false positive rate stays at most 0.5% on all five training folds, and a calibrated threshold.
Figure 3. Measured motivation on RAID with the Falcon-7B pair at a 0.1% calibrated FPR. Three bar charts. Recall of five zero-shot detectors on unedited, homoglyph and zero-width documents; machine documents missed by three detectors split by attack type; recall of four learned heads against the normalized Binoculars score.
Figure 4. Recall per section, recall gain per scoring pair, and the two scores of each document at a 0.1% calibrated FPR. Three panels. Recall in each of the eight RAID sections for MCP, normalized Binoculars, the normalized clipped LRT and NormDetect; recall gain of NormDetect over normalized Binoculars with 95% intervals for the Falcon-7B, Gemma-4-12B and Gemma-4-31B pairs and for Gemma-4-31B at blend weight 0.2; scatter of the normalized Binoculars score against the blend score with both calibrated thresholds dashed.
Method
TPR
Edited
Human FP
Max bound
Falcon-7B pair
Likelihood ( Gehrmann et al., 2019 )
14.500
22.45
8/7521
1.05 ✗
Log-Rank ( Gehrmann et al., 2019 )
13.000
12.24
5/7521
0.77
LRR ( Su et al., 2023 )
8.833
9.18
5/7521
1.05 ✗
Lastde ( Xu et al., 2025 )
12.333
13.27
3/7521
0.91
Telescope ( Nassif and Cooper, 2026 )
0.167
0.00
11/7521
1.31 ✗
Table 1. Recall (%) at a 0.1% calibrated FPR on RAID per scoring pair. Single-model baselines take the better of the original and normalized text.
Detector
Time (s)
GiB
Detector
Time (s)
GiB
Likelihood
0.06
13.2
Binoculars + norm.
0.35
26.2
Log-Rank
0.05
13.4
Clipped LRT + norm.
0.35
26.2
LRR
0.05
13.4
MCP + norm.
0.35
26.2
Lastde
0.06
13.2
RADAR
0.02
2.0
Telescope
0.06
13.3
DivEye
0.06
13.2
DNA-GPT
0.94
8.9
BiScope
0.35
26.4
Table 2. Time and peak GPU memory per document of every detector in Table 1 with the Falcon-7B pair on one A800 GPU.
Figure 5. Recall per attack slice at a 0.1% calibrated FPR with two scoring pairs. Two grouped bar charts of recall on twelve attack slices for Binoculars, normalized Binoculars, the normalized clipped LRT and NormDetect, one for the Falcon-7B pair and one for the Gemma-4-12B pair.
Figure 6. Upper bound of the human FPR per guard section at a 0.1% calibrated FPR. Grouped bar chart of the one-sided 95% upper bound of the human FPR in each of the eight guard sections and in the pooled guard set for normalized Binoculars, the normalized clipped LRT, MCP, Fast-DetectGPT and NormDetect, with a dashed line at 1%. Only the Fast-DetectGPT bars on poetry and Reddit cross the line.
Figure 7. Recall per length third and per decoding strategy of the development set at a 0.1% calibrated FPR. Left, recall of four detectors on the three length thirds of the development set. Right, recall of the same detectors under greedy decoding and under sampling.
Figure 8. Machine-generated documents that NormDetect flags and normalized Binoculars misses at a 0.1% calibrated FPR. Shading marks words or letters the attack changed, and a caret marks deleted words. Six text cards in two rows. The first row holds Reddit, reviews and recipes posts with a misspelling, a deleted article and changed numbers. The second row holds a poem with homoglyphs, a paraphrased abstract and a Wikipedia article with British spellings. Edited words or letters are shaded. Each card lists the score of normalized Binoculars below its threshold and the score of NormDetect above its threshold.
Figure 9. Recall and largest per-section human FPR bound of NormDetect and three variants, recall against the blend weight, and gain against the size of the training pool. Four panels. Recall of NormDetect and of three variants with one head for all sections, head features from the original text, and no normalization, with paired differences over the bars. The largest per-section human FPR bound of the same four, with a dashed line at 1% that only the variant without normalization crosses. Recall of NormDetect against the blend weight, from the anchor alone to the head alone. Recall gain over normalized Binoculars against the number of labelled training documents.
Figure 10. Human FPR bound when each RAID attack edits the same 1,000 human documents, at the thresholds of Table 1 . Grouped bars for six detectors over twelve groups, the unedited documents and the eleven RAID attacks, with a dashed line at a 1 percent bound. On the unedited documents and on the ten edits other than paraphrase, NormDetect stays below the line. On paraphrase every detector is above the line, and Fast-DetectGPT reaches 15.8 percent.
Figure 11. Recall at four calibrated FPR targets and the gain of NormDetect on three evaluation splits. Left, recall against the calibrated FPR target for three detectors. Right, recall gain of NormDetect over normalized Binoculars with 95% intervals on three evaluation splits, at the calibrated threshold and at equal human false positives.
Figure 12. Recall against the human FPR when the threshold is swept over the guard documents, with two scoring pairs. Two line charts of recall against the human FPR on the guard set, on a log axis from about 0.05% to 1%, for seven detectors with the Falcon-7B pair and with the Gemma-4-12B pair. A light vertical line marks 0.1%.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Detector
Abs.
Books
News
Poetry
Recipes
Reddit
Reviews
Wiki
All
n
1000
1000
1000
1000
1000
1000
521
1000
7521
Binoculars
0 (0.30)
0 (0.30)
0 (0.30)
3 (0.77)
0 (0.30)
3 (0.77)
1 (0.91)
0 (0.30)
7 (0.17)
Binoculars + norm.
0 (0.30)
0 (0.30)
0 (0.30)
3 (0.77)
0 (0.30)
3 (0.77)
1 (0.91)
0 (0.30)
7 (0.17)
Clipped LRT + norm.
0 (0.30)
0 (0.30)
0 (0.30)
4 (0.91)
0 (0.30)
4 (0.91)
0 (0.57)
0 (0.30)
8 (0.19)
MCP + norm.
0 (0.30)
0 (0.30)
0 (0.30)
2 (0.63)
0 (0.30)
3 (0.77)
0 (0.57)
0 (0.30)
5 (0.14)
Fast-DetectGPT
2 (0.63)
3 (0.77)
1 (0.47)
8 (1.44) ∗
2 (0.63)
5 (1.05) ∗
0 (0.57)
0 (0.30)
21 (0.40)
Appendix
Table 3. Human false positives per guard section at a 0.1% calibrated FPR, with one-sided 95% upper bounds (%) in brackets.
Method
TPR
Gain [95% CI]
Human FP
Binoculars + norm.
7.667
–
5/7521
NormDetect, κ=0.2
9.500
+1.83 [0.67, 3.03]
6/7521
NormDetect, κ=0.4
14.333
+6.67 [4.66, 8.84]
9/7521 ✗
NormDetect, κ=1.0
20.667
+13.00 [9.95, 16.06]
10/7521 ✗
Appendix
Table 4. Recall (%) at a 0.1% calibrated FPR on RAID with the Gemma-4-31B pair, by blend weight.
Falcon-7B
Gemma-4-12B
Method
Original
Normalized
Original
Normalized
Likelihood
14.500 ✗
13.167 ✗
4.500
4.167
Log-Rank
11.667
13.000
3.333
3.333
LRR
7.667 ✗
8.833 ✗
3.000
2.167
Telescope
0.167 ✗
0.167 ✗
0.167 ✗
0.333 ✗
Lastde
9.333
12.333
1.000
1.167
Appendix
Table 5. Recall (%) of single-model baselines at a 0.1% calibrated FPR on the original and the normalized text.