AI transparency
Every game on playing.company is created with artificial intelligence. Here we explain which AI we use, how we label it and what AI is (still) not good at.
Which AI we use
| Purpose | AI system | Provider |
|---|---|---|
| Game description, texts, quiz questions, story, world projects (missions, economy, dialogue), changes by chat | The AI (language models) | External AI provider |
| Music, sound effects and voices (only if you order them) | ElevenLabs (models for music, effects, speech) | ElevenLabs |
| Video trailer (only if you order it) | Seedance (video model) | BytePlus, a company of the ByteDance group |
| Checking inputs for prohibited content | our own filter, supplemented by the instructions to the AI | us |
Depending on plan and task, different AI models are used – on the Free plan, for example, a particularly fast, smaller model. Which models built a game is shown in the game information and in the machine-readable labelling.
The AI takes on these tasks:
- From your selection (genre, scope, style, game language, extras, idea) it designs a game description: rules, levels, difficulty, names, texts, quiz questions or story scenes – in 20 languages.
- In world projects it designs economy, world, missions and dialogue as data.
- It carries out change requests from the chat and estimates the price beforehand.
- It helps check inputs for prohibited content.
The AI writes no executable code. Our own game engine (“runtime”) plays the description and draws the graphics from fixed style templates. The graphics are not created by an image model. In 3D worlds, characters, vehicles and buildings come from our own library; the AI only decides what stands where and what happens.
How we label AI content
Since 2 August 2026 the transparency duties of the EU AI Act apply (Art. 50 Regulation (EU) 2024/1689, amended by Regulation (EU) 2026/1744). We provide AI systems that generate text, audio and video, and therefore label every game and every medium on several levels. We do not use a transitional period: we label from day one.
| Level | How | Where |
|---|---|---|
| Visible | Notice AI-generated and the sentence “AI-generated: this game was created with AI.” | In the player, on the share page, in the arcade and in the export |
| In the game itself | “ai” block in the game description: AI-generated, provider, models, creation time | In every saved and exported game |
| Machine-readable (game) | Meta tags (ai-generated, ai-provider, ai-models, generator) and schema.org data with the IPTC source type trainedAlgorithmicMedia | In the HTML of the player, share page and HTML5 export |
| Audio and video: visible | Notices such as “This music was created with AI.”, “The voices are AI-generated and synthetic – no real person is speaking here.” and the label “AI video” on the trailer | In the game player, in the studio and in the game settings |
| Audio and video: in the file | Metadata in every media file (ID3 for MP3, RIFF-INFO for WAV, comment for MP4, EXIF for the preview image) and, where the provider supplies it, a C2PA provenance record | In every AI media file |
| Audio and video: provenance document | Machine-readable record with checksum, provider, model and time, digitally signed (Ed25519); the public key is available for verification | Available per file, plus the HTTP header “X-AI-Generated” |
| In the chat | “You are chatting with an AI, not a human. Answers may contain mistakes.” | When changing by chat |
We follow the EU Commission’s code of practice on marking AI-generated content (published 10.06.2026) and the Commission’s guidelines on Art. 50 (published 20.07.2026): several complementary labelling layers rather than a single one. Where media can only be heard (audio only), we plan an additional audible notice. An invisible watermark and a self-written C2PA manifest are the next steps we are looking at.
Voices, videos and real people
- Voices are synthetic. We only use standard voices and create no voice clones. No real person is speaking, and we don’t imitate the voices of celebrities or other real people.
- Trailers show invented games. Characters and places are not real. We don’t imitate real people (no “deepfakes”) and ask the video provider to avoid real people and protected characters; our description for the video contains only the genre, style and mood of the game.
- 3D worlds: Our engine assembles the 3D world from an AI-generated description; the 3D models come from our own library.
Even so, AI videos and music can unintentionally resemble existing works or styles. If you notice something, use the “Report” button on the game.
Limits and mistakes
AI can be wrong. Typical weaknesses:
- Facts: quiz questions or explanations can be wrong or outdated. Check facts before using a quiz, e.g. in class.
- Languages: quality can vary across the 20 game languages; rare languages and right-to-left scripts are more error-prone.
- Game balance: levels can turn out too easy or too hard. You can adjust that by chat.
- Texts: wording can be inappropriate, unintentionally similar to known works, or misleading.
- Same ideas: similar inputs can lead to similar games – including those of other users.
- Media: music, voices and videos are unpredictable; pronunciation, emphasis and picture content can differ.
We check every game description technically (format, limits, required fields) and filter known brands and characters. We cannot guarantee that every result is error-free and legally unobjectionable. If you notice something, use the “Report” button on the game.
Your inputs and AI training
We do not train our own models on your inputs or games. In its commercial terms, our AI provider commits not to train models on customer content from the API. According to our AI provider, API inputs and outputs are, depending on the model and configuration, either not stored or stored for at most 30 days; if there is suspicion of violations of the usage policies, the provider may store data longer (up to two years).
We send our AI provider only what is needed for the game: your selection, title, idea and change requests. We insert the recipient’s first name and the greeting only after the AI has produced the game, on our own server; we don’t send your email address or payment data. If you order media, only cleaned texts or a scene description and a preview image go to ElevenLabs and BytePlus. Details are in the privacy policy.
Human oversight and moderation
Automatic checks run before and after every build. They can stop an input, for example when a protected name is detected – you can then rephrase. Notices via the “Report” button and decisions on suspending games or accounts are always reviewed by a human on our team.
More in the usage policy and under report content.