Accuracy

Every number on this page comes straight from the measurement file of the Engine that judges your Scans. We never type them in by hand, and they change only when we measure again.

How we measure

We test on our own evaluation set: about 5,000 human texts and 2,200 AI texts per language. The human texts were all written before ChatGPT existed: news, encyclopedia articles and research abstracts, including English written by authors in Spain and Latin America. The AI texts come from twelve current models made by nine companies, in five genres, and some of them were paraphrased by a second model to hide their origin.

Half of the set sets the thresholds. The other half, which is never used for anything else, gives the rates below. Each rate shows how many texts it rests on and a 95% confidence interval.

The thresholds are cautious on purpose. Likely AI starts at the score that only 1 in 100 human texts reaches in our tests. Likely human ends at the score that only 1 in 100 AI texts falls below. Anything in between is Unclear, and that is an honest answer.

Engine `2026.11-oculus-v0`

English: Passes the launch gate

Likely AI above	0.941
Likely human below	0.123
Human texts called Likely AI	1.2 % (0.9–1.7) · 2490 texts
Raw AI texts called Likely AI	89.2 % (87.1–91.0) · 961 texts
Paraphrased AI texts called Likely AI	90.1 % (83.5–94.2) · 121 texts

Human texts called Likely AI, by source

English research abstracts by authors in Spain and Latin America, 2010–2021	5.1 % (3.6–7.2) · 585 texts
Research papers by non-native English authors, 2015–2021 (TextSight benchmark)	0.0 % (0.0–4.2) · 88 texts
English Wikipedia revisions before 2022 (TextSight benchmark)	0.0 % (0.0–4.9) · 75 texts
English Wikinews, 2004–2021	0.0 % (0.0–0.5) · 838 texts
English Wikipedia, March 2022	0.0 % (0.0–0.4) · 904 texts

Raw AI texts called Likely AI, by model

`anthropic/claude-haiku-5.5`	93.1 % (86.4–96.6) · 101 texts
`anthropic/claude-sonnet-5.5`	71.4 % (52.9–84.8) · 28 texts
`deepseek/deepseek-v4.1-flash-1017`	86.8 % (79.9–91.6) · 129 texts
`google/gemini-3.7-flash`	92.5 % (83.7–96.8) · 67 texts
`google/gemini-4-argon`	89.3 % (72.8–96.3) · 28 texts
`meta-llama/llama-4-maverick`	92.1 % (84.6–96.1) · 89 texts
`mistralai/mistral-small-2603`	100.0 % (96.9–100.0) · 119 texts
`openai/gpt-6-luna`	85.7 % (78.3–90.9) · 119 texts
`openai/gpt-6.1-terra`	66.7 % (47.8–81.4) · 27 texts
`qwen/qwen3.8-flash`	100.0 % (96.3–100.0) · 101 texts
`x-ai/grok-4.3`	89.2 % (75.3–95.7) · 37 texts
`z-ai/glm-5.3`	76.7 % (68.3–83.5) · 116 texts

Spanish: Passes the launch gate

Likely AI above	0.132
Likely human below	0.088
Human texts called Likely AI	1.4 % (1.0–2.0) · 2434 texts
Raw AI texts called Likely AI	99.5 % (98.8–99.8) · 1005 texts
Paraphrased AI texts called Likely AI	99.2 % (95.7–99.9) · 129 texts

Human texts called Likely AI, by source

Spanish Wikinews, 2004–2021

1.4 % (1.0–2.0) · 2434 texts