Inspiration
We were inspired by a lot of AI-based games, such as Death by AI and With You 'Til The End where you have to use prompt engineering to manipulate the AI to bypass restrictions. We wanted to create a game where we push the limits of AI in a way that's entertaining.
Overall Idea
The game takes you through the DMV experience where you have to go from window to window, speaking with different DMV AI agents whilst trying to get your driver's license. It is infinite and the goal is to get through as many agents as possible to get a high score before the time runs out.
Technologies
We built this using the Unity game engine, Gemini as the backend AI model, and ElevenLabs text-to-speech service for the DMV agent voice. We used Google’s Gemini API to facilitate a conversation with the user by providing Gemini with a specific tuned personality, chat context, and important keywords needed to allow the user to advance. Then, we took the output of Gemini and used ElevenLabs text-to-speech service with our own customized AI voices for each DMV agent to allow the DMV agents to talk back to the player in near real-time. For the game assets and art, we used the pixel art tool Asesprite to design custom ducks, accessories, and UI components.
Challenges
We played with multiple AI models from many different AI providers, however, we had frequent issues with getting fast response times, which interrupted the flow of the game. We also noticed that some models were harder to get funny personalities than others. Another issue is that a particular AI personality decided to gain sentience and would bypass our debug commands. This issue was solved by playing around with the prompts for the personalities as well as changing some other hard-coded components.
Proudest Accomplishments
We are especially proud of the personalities and the safeguards we put in place to make our game challenging as well as fun. Each agent has its own unique personality with a voice to match, making them all a different challenge as well as interesting in their own ways. Since AI has a mind of its own, we put emphasis on putting restrictions in place so that the agents are not giving strange or overly lengthy responses as well as including tags to signify that the user can move on.
Lessons Learned
We learned many things, especially about AI. AI systems can be vulnerable to users intentionally giving misleading or malicious information, also known as prompt injection attacks. We had to think very carefully about how a player’s imputed text can affect our AI agents’ behavior. We chose to include filters in the prompt of the AI personalities to sanitize the responses. We also noticed that the AI would sometimes give unexpected outputs, so we had to include restraints and rules to make the responses received more reliable. Lastly, it became clear to us that different AI models can affect the game’s responsiveness and flow. Larger models slowed down the output time and therefore slowed down the game, so we had to find a balance between model complexity and performance.
What's Next for ChatDMV
Looking ahead, we want to add features that enhance both gameplay and replayability. One idea is to add a history button that lets players review their prompts and the responses from the AI so they can learn from their previous attempts and improve their strategies while playing. It is also possible that we could have the AIs learn from previous interactions to make the gameplay harder each time. Another idea is to include tracking for high scores, giving players a way to compete against themselves or see how far they can get. These additions would make the game more engaging and also create opportunities for new and evolving strategies.
Log in or sign up for Devpost to join the conversation.