Polish app aims to teach AI to tell Silesian dumplings from doughnuts

Global AI models can mistake kluski śląskie (Silesian dumplings) for French macarons, or for underbaked doughnuts, attribute Gdańsk's Neptune Fountain to Bologna, and mistake scenes from "Seksmisja" for Soviet cinema. The SpeakLeash Foundation set out to change that and launched a mobile app called Obywatel Bielik ("Bielik Citizen"), through which any smartphone user can submit a captioned photo to teach the Polish AI model to recognize domestic culture.

How the app works

The Obywatel Bielik interface is simple: users add a photo from their gallery or take one on the spot, then describe it in text or dictate a description by voice. AI helps correct language and punctuation before the material enters the training database. The creators split the content into eight thematic categories and added weekly goals, activity stats and leaderboards to encourage regular participation, including elements of team competition.

Every submission goes through two-stage verification, first automated screening, then manual review by volunteers. The creators say biometric data and facial recognition are deliberately excluded from the project, which they say sets Obywatel Bielik apart from commercial training datasets built without users' consent.

Where the problem comes from

Large vision models are trained mostly on data from English-speaking countries and Western Europe, which leaves them weak on the local cultural context of smaller language markets. Kluski śląskie, a dish typical of southern Poland, are sometimes recognized as French macarons or raw doughnuts, because the models previously had too few examples described in the right context. Similar mix-ups affect landmarks and places: Gdańsk's Neptune Fountain is sometimes confused with sculptures in Bologna, and AI can classify scenes from the cult Polish comedy "Seksmisja" ("Sexmission") as Soviet cinema.

This isn't purely an aesthetic curiosity. Vision models are increasingly built into shopping assistants, travel apps, retail image-recognition systems and educational tools. Misreading local context in these applications means worse recommendations, misleading descriptions, or a loss of user trust in Polish AI deployments built on foreign base models.

Who is behind the project

Obywatel Bielik extends the already well-known Bielik.AI language model project, developed by the community under the SpeakLeash Foundation in cooperation with the Cyfronet AGH Academic Computer Centre. Public and media institutions have also joined the data collection effort: the National Digital Archive and Polska Press Grupa have made part of their photo collections available as a starting point for further community contributions.

Releasing the mobile app is a turning point for the Obywatel Bielik project. It means literally anyone, a student, a senior, a parent, can take part. - Marcin Dąbrowski, co-creator of the project

The Bielik model is developed in a grassroots format based on volunteer contributions and open resources, which sets it apart from PLLuM, Poland's other major national language model project, which is institutionally funded and coordinated. Obywatel Bielik is meant to supply the visual data that has been missing so far, to expand the model from a purely text-based layer into a multimodal one that understands images alongside descriptions.

What this means for Poland's AI market

National language and vision models are gaining importance amid debate over technological sovereignty and Europe's dependence on suppliers from across the Atlantic or from China. The SpeakLeash project shows an alternative way to build such resources, not through multibillion infrastructure investments, but through mass community involvement, similar to Wikipedia or OpenStreetMap.

The initiative's success, however, depends on the scale of participation. Collecting a million captioned photos requires engaging a far broader group than the project's existing enthusiasts, and the quality of the descriptions and the regional diversity of the material will determine whether the model actually learns to distinguish local contexts, or simply repeats existing errors on a new dataset.

The creators have not yet given a specific date for when data from the app will feed into the next version of the Bielik.AI vision model, nor have they revealed when the project will reach its target of one million photos. The pace depends on how quickly the app spreads beyond the circle of people already involved in the project.