Obywatel Bielik: Photo-Collection App for Polish AI Launches on Smartphones

A photo of a roadside shrine, a plate of pierogi, or a local market can now feed directly into the dataset training Poland's artificial intelligence. On July 14, 2026, the SpeakLeash Foundation launched Obywatel Bielik (Citizen Bielik), a mobile app for Android and iOS that lets any user upload photos and descriptions to help build the first Polish vision model.

The mechanics are simple. Users photograph a place, landmark, regional dish, object, or custom, add a short description, and submit it through the app. The description can be typed or dictated, with the text version of Bielik.AI helping to clean up the wording and punctuation. Photos are sorted into one of eight thematic categories, and the app adds gamification elements: weekly challenges, statistics, and participant rankings.

Why Ordinary Photos Matter

The project's creators stress that it's precisely the ordinariness of the material that counts. Photos are meant to be taken on a phone during everyday activities rather than as artistic shots, since the model needs to learn to recognize Poland as its residents actually see it, not a curated version drawn from professional photo shoots.

All it takes is a walk outside to photograph a roadside shrine, a regional dish, a landmark, or a place - Marcin Dąbrowski, head of the Obywatel Bielik project

Marcin Dąbrowski, who leads the project, and Sebastian Kondracki, the originator and co-creator of Bielik.AI, argue that existing global vision models struggle with local cultural context. The project's promotional materials cite specific examples of mistakes: Silesian dumplings (kluski śląskie) identified as French macarons, Gdańsk's Neptune Fountain confused with a monument in Bologna, and a scene from the Polish comedy "Seksmisja" mistaken for a frame from the Soviet film "Solaris."

A multimodal model will only be as good as the diversity of the data it's trained on - Sebastian Kondracki, co-creator of Bielik.AI

Who Is Behind the Project

Obywatel Bielik builds on the work of the SpeakLeash Foundation, the organization behind the development of the text-based Bielik model. The Cyfronet AGH Academic Computing Center, one of Poland's largest computing hubs, has joined to help build the infrastructure and process the data. Partners in the collection effort include the National Digital Archive (Narodowe Archiwum Cyfrowe), which operates under the patronage of Poland's Ministry of Digital Affairs, and Polska Press Grupa, meant to give the project access to additional visual resources and promotional channels.

The model the team is building is meant to retain an open architecture, as earlier versions of Bielik did. That means public institutions and companies will be able to run it on their own servers without having to send data to foreign cloud providers.

Practical Applications

The creators point to specific areas where a vision model that understands Polish context would have an edge over global solutions: product categorization in Polish e-commerce, document analysis and text recognition (OCR) in public administration, and tools that make visual content more accessible to blind and visually impaired users, in line with EU digital accessibility requirements.

The project fits into a broader push to build national data resources for local AI models, alongside earlier efforts to collect text for Bielik and PLLuM. Unlike text, visual data reflecting Polish cultural context has so far been practically absent from open training datasets, which explains the decision to involve a broad base of citizens rather than buying ready-made photo databases.

What's Next

The team has not yet given a firm release date for the multimodal model itself, pointing to the data collection and labeling stage as a necessary precondition for further work. The pace of photo submissions following the app's launch will determine whether the target of one million photos can be reached this year or whether the process will take longer. For Polish companies and institutions already experimenting with the text version of Bielik, a version that understands images would open the door to building tools that classify documents, photos, and visual material without relying on models trained mainly on data from Western Europe and the US.