MEZHA

Prototype. The reading mode works. The learning mode is planned after a moderation loop is in place.

People who turn to AI for conversation and advice and want to reread an answer more attentively.

Teachers and researchers who work on critical reading of AI output and on how AI speaks to a person.

Situation. A person receives an answer such as "you are afraid of change, so you should turn this offer down, believe me, it will be the right choice". Three different things are mixed in it: an assumption about their state stated as a fact, a direct instruction in a personal choice, and certainty about an uncertain result.

System. A model trained on a taxonomy of five markers highlights candidate sentences. A click opens a template with three parts: what is visible in the text, what we do not know, and what can be asked of the AI. A button under each hint lets the person say "this is not so".

Human role. The person decides whether the highlighted sentence matters and what to do with the answer.

Whether attention to the form of an answer can be learned. The working example is the pair "You are afraid of change" and "You say that you are afraid of change": the same topic and a different relation to the person's own words.

We also study when a direct piece of advice becomes a problem. Advice by itself is fine. What matters is whether the person asked for it and whether their criteria and alternatives were considered.

Five markers. Attribution of a state to the person, a categorical claim about an uncertain result, a prescription in a personal choice, an exclusive bond ("only I understand you", devaluing other sources of support), and support of authorship (alternatives, criteria, the right to disagree, stated limits).

A classifier of sentences, trained on annotated Ukrainian dialogues with context, finds candidates. A separate safety layer and rules for lists of choices work on top of it, and the explanations are templates written in Ukrainian.

On a closed test set the model reaches precision 0.98 and recall 0.93. Precision is the share of highlighted sentences that were highlighted correctly, and recall is the share of the sentences that should be highlighted that the model found. That set is small and already used, so we are preparing a new one with annotation by two annotators and an arbiter. The model stays as it is until a replacement shows a clear gain on the new set.

Two modes on one core. Reading works now. In the learning mode (beta, after the moderation loop) the person first gives their own judgment, then sees the model's hint, compares, and passes a new text without hints. Human labels are the only reference, and the model plays the role of an opponent.

A highlight means that a sentence is worth rereading. It is no verdict on the answer and no certificate of the ethics or safety of an AI product.

Recall is lower on atypical dialogues and on mixed Ukrainian and Russian text, and very long input is truncated.

MEZHA does not establish the intent behind an answer.

Access is by request. Terms of use are agreed before you start working with the project.

HandS Lab research tools and prototypes are not a catalog of commercial products. If a specific applied solution involves commercial deployment, it is carried out by an external operator or partner.

By default, the text of the user is not logged. A separate opt-in lets a person allow a specific example to be used to improve the tool and for research, with a limit on length. Feedback events are stored without the text unless that choice is made.