AplikaceAplikace
Nastavení

This is an old revision of the document!


AI Detector

AI Detector is a web-based tool designed to distinguish between human-written and AI-generated texts. It supports English and Czech and can also estimate which large language model (LLM) a text most closely resembles. The interface is available in both English and Czech.

The classifier was trained on the AI Brown and AI Koditex corpora, which contain texts generated by various language models alongside their human-written counterparts. It identifies characteristic patterns in vocabulary, phrasing and grammatical structure to distinguish human from AI-generated writing.

How to use the application

Paste the text you wish to analyse into the input field, or upload a plain text file (.txt) using the Upload .txt button.

The input must contain at least 30 tokens (words and punctuation marks). However, longer texts generally produce more reliable results, and at least 200 tokens are recommended.

The application detects the input language automatically (Language: auto), but you can also select English or Czech manually.

Two analysis modes are available:

  • UDPipe (accurate) – uses morphological and syntactic analysis provided by UDPipe. The text is sent to the LINDAT service for linguistic annotation. This is the more accurate option.
  • Local (fast) – uses surface-level text features without external linguistic processing. It is faster but slightly less accurate.

Click Classify to start the analysis.

Understanding the results

The application divides the input into chunks of approximately 200 tokens and classifies each chunk separately.

The overall verdict is displayed as one of three labels:

  • HUMAN – none or almost none of the chunks were classified as AI-generated.
  • AI – nearly all chunks were classified as AI-generated.
  • MIXED – the text contains both human-like and AI-like chunks.

Two indicators are displayed below the verdict:

  • mean P(AI) – the average AI score across all analysed chunks.
  • chunks flagged AI – the percentage of chunks classified as AI-generated.

The Attribution of AI-flagged chunks section shows which language models the detected AI-like passages most closely resemble.

The Per-chunk detail table provides individual scores, classifications and model attributions for each chunk. Clicking a text preview reveals the corresponding passage.

Adjusting the sensitivity

The Prior P(AI) slider allows you to adjust the classifier's sensitivity according to how likely you believe the text is to have been generated by AI.

  • 0.50 – the default, neutral setting.
  • Lower values – more conservative detection, reducing the risk of incorrectly flagging human writing as AI-generated.
  • Higher values – more sensitive detection, reducing the risk of missing AI-generated passages at the cost of more false positives.

For example, if you are analysing a collection of texts that you have good reason to believe were written by humans, lowering the prior makes the classifier more cautious about assigning an AI label.

Changes to the prior are reflected immediately in the results, without having to analyse the text again.

How accurate is the classifier?

The classifier was evaluated on held-out test documents that were not used during training. The results for both languages and processing modes are shown below:

Language Mode AI detection recall AI detection precision* Balanced accuracy
English UDPipe 94.0% 93.6% 93.8%
English Local 93.1% 93.2% 93.1%
Czech UDPipe 92.4% 94.7% 93.6%
Czech Local 90.6% 94.7% 92.8%

Recall measures how many AI-generated texts the classifier successfully detects. For example, a recall of 94% means that 94 out of 100 genuinely AI-generated texts are correctly identified, while 6 are missed.

Precision measures how often an AI label is correct. For example, a precision of 94% means that, out of 100 texts flagged as AI-generated, 94 really are AI-generated and 6 are human-written.

Balanced accuracy measures overall performance, giving equal importance to correctly identifying human-written and AI-generated texts.

*The precision values in the table assume an equal proportion of human-written and AI-generated texts (50% each). In real-world datasets where AI-generated texts are less common, a larger proportion of positive detections may be false alarms.*

The results are based on individual 200-token chunks. Longer documents may produce more reliable overall verdicts. In the UDPipe mode, attribution to a specific language model or the human class achieves 67.0% accuracy for English and 68.3% for Czech.

These figures were obtained on test data drawn from the same sources as the training corpora. Performance may differ for other genres, domains and language models.

Limitations

The classifier provides an estimate, not proof of authorship. Its results should not be used as conclusive evidence that a particular text was written by AI.

Some human-written texts, particularly translations, may be incorrectly classified as AI-generated. Detection is also less reliable for very short texts and for newer language models that were not represented in the training data.

The classifier was trained on specific types of written text, so its accuracy may vary across genres and domains.

Attribution to a particular language model indicates stylistic similarity, not verified provenance. A text resembling the output of a specific model was not necessarily produced by that model.

How to cite AI Detector

Milička, J.: AI Detector. Faculty of Arts, Charles University. Prague, 2026. Available at: https://www.korpus.cz/ai-detector/.