Russian machine-text detectors disagree when only task wording changes

An arXiv study looks at automated detection of machine-generated text in Russian, a technology the authors say is applied where the verdict carries a cost for the author. Their starting point is that origin is a property of the production process, not an observable property of the text. The conditions under which a text is produced, they argue, push it in the same direction in which a generative model is presumed to push it. The study sets out to measure how much those conditions contribute while origin is held fixed.

The corpus has 1,882 documents. The machine part was generated by four generation channels, each backed by its own model family, across 45 tasks in three task modes. The human part is archival and was annotated for source regimentation level before the texts were read. The primary contrast sits entirely inside the machine corpus: both halves of a pair come from the same model for the same task and differ only in task wording.

Machine-likeness is measured four ways: an interpretable index with weights fixed before data collection, a trained classifier, a zero-shot metric on an open model, and a judge model.

The main result is disagreement between procedures. Changing the task mode shifted the scores of all four, but three gave a negative shift and the judge gave a positive one. None of the four intervals covers zero. The comparison plan and the specifications of procedures 2 to 4 were fixed after the result of procedure 1 and before procedures 2 to 4 ran. The disagreement survived correction of a corpus defect and of the nested validation.

Two further findings point the same way. A baseline that uses only the source identifier reaches a median AUROC of 1.0000 on the thirteen group splits where it is defined. And removing emphasis markers flipped a confident machine verdict to human without changing the full text by a single character.

The authors conclude that the scores are properties of a specific measurement system and corpus. They reflect data collection, repetition, formatting and preprocessing decisions as well as linguistic style.

Key facts

The corpus has 1,882 Russian documents: a machine part from four generation channels (each with its own model family) across 45 tasks in three task modes, plus an archival human part.
Within the machine corpus, changing only the task wording shifted the scores of all four measures: three negative, the judge model positive, and none of the four intervals covers zero.
A baseline using only the source identifier reaches a median AUROC of 1.0000 on the thirteen group splits where it is defined.
Removing emphasis markers flipped a confident machine verdict to human without changing the full text by a single character.
The authors say the scores reflect data collection, repetition, formatting and preprocessing decisions as well as linguistic style.

Why it matters

Detectors of machine-generated text are used where a verdict costs the author something. This study isolates one confound: it keeps the model and the task fixed, changes only the wording of the task, and still sees the scores move. Because the measures move in opposite directions, with three going negative and the judge model going positive, a single score does not behave as a stable stand-in for origin. The source-identifier baseline at a median AUROC of 1.0000 on thirteen group splits adds to the picture, since it shows how much of a detector's apparent success can come from properties tied to where a text came from.

Who it affects

Anyone who builds or relies on machine-generated text detection for Russian-language material, since the corpus and the experiments are in Russian. That includes people who act on a detector verdict about an author, the setting the abstract names as the motivation. Researchers who compare detectors across corpora are also affected, because the authors say the scores belong to a specific measurement system and corpus.

How to use it

The abstract states no practical recommendations or policy conclusions. What it does offer is a framing: a machine-likeness score is a property of a particular measurement system and corpus, shaped by data collection, repetition, formatting and preprocessing as well as linguistic style. Read any single detector output with that in mind. The design also has a reusable shape, a within-machine contrast where the same model and the same task differ only in wording, with several measurement procedures run side by side.

How solid is it

The design has some safeguards. The interpretable index used weights fixed before data collection, and the human part was annotated for source regimentation level before the texts were read. The comparison plan and the specifications of procedures 2 to 4 were fixed after procedure 1's result but before procedures 2 to 4 ran, so the plan was not fully set in advance. The disagreement survived correction of a corpus defect and of the nested validation. The text available here is the abstract of a preprint. The sizes of the shifts and the confidence intervals are not given, only their signs and the fact that none of the four intervals covers zero. The four generation channels and model families are not named, and neither are the three task modes.

Risks and caveats

The findings come from one Russian-language corpus of 1,882 documents, and the split between machine and human parts is not given. The abstract does not name the model families or say which detector produced the confident machine verdict that flipped when emphasis markers were removed. The study reports that procedures disagree; it does not say which one is right, and it does not claim to. The authors also say the scores depend on the specific measurement system and corpus, so the results should not be assumed to carry over to other detectors or other data.

“origin is a property of the production process, not an observable property of the text”

— Study abstract, arXiv 2611.05385