The developer of the game UNSEEN encountered a fascinating issue during testing: the small AI model used to analyze photos had a tendency to hallucinate. When presented with two random photos, the model almost always invented a connection between them, displaying confidence that was not backed by facts. This forced the author to rethink the role of artificial intelligence in the game's core mechanics.
UNSEEN is a game that encourages users to go outside, notice real objects, and photograph them. The core philosophy is that the AI should only describe what it sees rather than decide whether an observation is valid. To achieve this, the author implemented an 'evidence gate,' which acts as a strict judge based on deterministic code rather than creative assumptions or AI-generated narratives.
Technically, the developer used the Qwen3-VL-2B model via Ollama for local image processing. This approach avoided cloud API costs and ensured data privacy. However, it turned out that the model often included phrases about its own uncertainty in its output, which the system mistakenly interpreted as signs of synthetic content, leading to the denial of rewards for genuine outdoor photographs.
The root cause was that the code analyzed all model output, including its own epistemic hedges. When the model mentioned the possibility that a photo might be a digital render, the system automatically blocked the reward. This was fixed by fine-tuning the classification logic to separate the AI's technical doubts from the actual assessment of physical objects present within the frame.
Furthermore, the developer discovered that the default sample photos used for demonstration were taken indoors. Since the game is designed for outdoor exploration, these photos were rejected by default. This created the false impression that the recognition system was broken, even though it was simply following its rules by filtering out content that did not meet the criteria for physical discovery rewards.
This experience highlights the importance of deterministic control in systems that rely on AI. Instead of allowing the model to decide the player's fate, the author used it as a tool for structuring data while leaving the final decision to hard-coded logic. This approach makes the game reliable and fair, turning the technology into a helpful assistant rather than a source of unpredictable hallucinations.
🔗 Original Source: Read Original Article