Which AI-text detectors work on Hungarian?

Submit a detector

1. Download HuDetect-1k and run your detector on every row, in order.
2. Save predictions.csv with a pred column (0 human, 1 AI) and a score column (AI probability).
3. Open a discussion in this Space with the file, the detector name, version and date. Results are verified and added to the table.

from sklearn.metrics import confusion_matrix, roc_auc_score
import pandas as pd

y = pd.read_csv("hf://datasets/mihalykiss/GenAIDetectHU/hudetect1k_v1.3.csv")["label"]
p = pd.read_csv("predictions.csv")
tn, fp, fn, tp = confusion_matrix(y, p["pred"]).ravel()
print("accuracy", (tp + tn) / len(y), "FPR", fp / (fp + tn), "FNR", fn / (fn + tp))
print("AUROC", roc_auc_score(y, p["score"]))

Why the results differ so much

GPTZero ranks texts well (AUROC 0.9923) but decides too cautiously on Hungarian, so it misses a fifth of AI texts. Originality.ai catches almost all AI texts but flags 32.6% of human ones.

ZeroGPT, isgen.ai and Smodin score an AUROC between 0.69 and 0.77, so their scores barely separate the two classes. No threshold fixes that; it usually points to little Hungarian training data. Preds 3, trained on Hungarian, reaches 0.9996.

For checking theses and essays, the false positive rate matters most. A practical guide in Hungarian: Szakdolgozat AI-ellenőrzése.