Understanding the results

What maid measures

maid looks for patterns associated with human-written and AI-generated text in its training data. It cannot establish who wrote a passage, or how it was written.

The score and the decision boundary

The AI score runs from 0 to 100. Higher scores are on the AI side of the model’s decision boundary; lower scores are on the human side. We adjust the scale so that the boundary is always 50, even when different models use different thresholds.

The two sides are scaled separately and the displayed score is rounded to a whole number. A displayed 50 can fall on either side of the actual boundary. A 0 or 100 is not certainty.

A score of 80 does not mean “80% AI,” nor does it mean 80% of the words were generated. It is a threshold-adjusted score, not a probability.

Probability and calibration

Inference details shows the underlying AI probability and the actual decision threshold. Calibration adjusts model outputs using labeled examples so that probabilities better match observed frequencies on similar data. For example, among comparable passages assigned a probability near 0.8, good calibration would mean roughly eight in ten are AI-labeled.

That relationship may not hold for your text. A different topic, language, writing style, or generator can change the model’s reliability. The displayed score is not directly comparable evidence across models.

Mistakes in both directions

A false positive is human-written text classified as AI-generated. A false negative is AI-generated text classified as human-written. Changing the decision threshold trades one kind of error against the other; it does not eliminate either.

Edited, mixed-authorship, and unfamiliar text can be especially difficult to interpret. Treat the result as one signal, not proof of authorship.

How much text is analyzed

Each model has a token limit. Tokens are pieces of text, not necessarily whole words. If your passage exceeds the limit, only the beginning is analyzed. The result reports how many tokens were used.

Your text stays local

Text analysis runs in your browser. We do not send your submitted text or results to our servers. Model files are downloaded and may be cached on your device.

The beta is free and open to everyone, with no account required. Optional sign-in and account records are handled by Supabase.