[sic] Detector

Sic reports the text exactly as written.

Sic 1 · sic-2026a

[sic] One paper, four ways

Fig. 1. Sic 1 reading four versions of one published paper.^† Every sentence is scored from 0 to 1, and those above the selected threshold are marked. Recorded output. Score your own text →

1What it does

Sic scores every sentence of a scientific text for machine authorship.

Each score comes with its position in the text, so a result can be traced to the exact sentence. Sic reports two thresholds: a strict one that almost never marks human writing, and a sensitive one that catches more machine text at the cost of more false alarms. Every response states how often each threshold is wrong.

Sic provides evidence, not verdicts. It does not identify an author, and a score should never be the sole basis for a decision about a person.

2Compared with commercial detectors

We ran Sic 1 and the two most widely used commercial detectors on the same texts. Sic 1 is equally conservative with human writing, performs alike on text from GPT and Claude models, and is better at locating machine-written passages inside an otherwise human document. The commercial detectors flag more text from open-weight models.

Human texts

How many texts written by people does each detector flag? Lower is better.

Sic 10 / 118 A0 / 118 B0 / 38

Level.

Inside a document

In human documents with machine passages spliced in, how well are the machine's words picked out? AUROC per word.

Sic 10.92 A0.79 B0.65

Sic 1 leads.

GPT and Claude

How many whole texts written by GPT and Claude models does each flag?

Sic 129 / 37 A28 / 37 B28 / 37

Level.

Open-weight models

How many whole texts written by seven open-weight model families does each flag?

Sic 172 / 100 A92 / 100 B19 / 20

Sic 1 trails.

Fig. 2. Sic 1 and two commercial detectors (A and B) on identical texts, each at its default setting. B was run on a subset of the texts. Tables are in Research, §1.5.

Locating machine-written text is what Sic is built for. Inside mixed documents it recovered 57% of the machine-written words at its sensitive threshold without marking a single human-written word. The commercial detectors recovered about 40%.

3Error rates

A text counts as flagged when any one of its sentences scores above the threshold.

Table 1. Share of texts flagged by Sic 1.
Texts	Strict threshold	Sensitive threshold
Written by people
Research papers	0.05% 4 / 7,950	0.35% 28 / 7,950
Commentary and opinion	1.0% 6 / 599	3.7% 22 / 599
Real-world test set	0% 0 / 150	4.7% 7 / 150
Written or edited by models
Written by GPT or Claude models	86% 278 / 325	92% 299 / 325
Written by open-weight models	72% 72 / 100	81% 81 / 100
Human text edited by GPT or Claude models	59% 47 / 79	76% 60 / 79

All texts are held out from training. Methods are in Research, report 1.

Fig. 3. Sic 1's score for each of the 554 texts in the real-world test set, one dot per text, placed at its highest sentence score (log-odds axis; scores beyond the axis are stacked at the ends). Red dots are above the strict threshold, pale red above the sensitive one.

4Limits

Granularity: A sentence's score reflects the passage around it. Sic locates machine-written paragraphs; a single machine-written sentence in a human paragraph, or a grammar pass, is usually not detected.
Scope: Sic is calibrated for English scientific prose. Scores on other kinds of writing are not meaningful.
Commentary: Opinion pieces and editorials are flagged more often than research papers (Table 1).
Low scores: A low score does not establish that a person wrote the text.

5Use it

Listing 1. shell

curl https://api.etal-labs.com/v1/detect \
  -H "Authorization: Bearer $ETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "sic", "text": "Paste the manuscript text here."}'

The response lists each sentence with its score, the two thresholds, and the error rates in Table 1. See the API reference, or try it in the playground.

Table 2. The Sic family.
Model	API string	Status
Sic Minor	n/a	forthcoming
Sic 1	`sic-2026a`, alias `sic`	Available. Up to 10,000 words per call, about 2 s per 1,000 words.
Sic Major	n/a	forthcoming

Get an API key →Responsible use