GENPTCHA

Train your intuition against a rogue AI that patches its own flaws daily based on human feedback.

Comment

Inspiration

We are living in the "Post-Truth Era." With the rise of Generative AI, distinguishing between human creativity and a machine-generated content has become a survival skill.

I was inspired by community discussions comparing AI-generated text with the works of great writers. People were trying to pinpoint exactly what was missing in the machine's output, and what makes human writing feel real. I decided to gamify this process.

Think of it as RLHF (Reinforcement Learning from Human Feedback) made into a game. As players provide feedback and opinions for the generated-content, the AI evolves. This competition helps humans sharpen their intuition for art, while also teaching the AI to better understand the human spirit.

For years, we proved we weren't robots by clicking traffic lights in a CAPTCHA box. But in the age of Gemini, that is no longer enough. GENPTCHA asks the ultimate question: When AI can write like a poet, how do we prove our humanity now?

What it does

GENPTCHA is a cyberpunk-themed Adversarial Turing Test Game.

  • The Duel: Users are presented with two text segments: one written by a human author, the other generated by Gemini 3.
  • The Detection: Users vote on which is human, They highlight specific sentences that gave the AI away and tag the AI's weaknesses (e.g., Soulless).
  • The Evolution: Every 24 hours, the system aggregates global user feedback. GENPTCHA analyzes its Battle Report, reflects on the detection rate, its weaknesses, and generates a System Log and New Duel Cases designed specifically to counter the previous detection strategies.

How we built it

  • GenAI API: We use the gemini-3-pro-preview API for three critical reasons: 1) its advanced reasoning capability allows the AI to analyze game stats and evolve strategies; 2) its high factual accuracy when combined with Google Search ensures sourced literary excerpts are authentic, not hallucinations; 3) its native multimodality paves the way for expanding into image and video deepfake detection in the future.

  • The Backend: We wrote a Python script to grab the latest statistics from database, and trigger Gemini's reflection process to analyze and generate new game content.

  • The Frontend: We built the frontend with Flutter to create a high-fidelity Cyberpunk Terminal experience, and used custom animations to make the web feel immersive.

  • The Database: Firestore stores users' feedback, syncs user votes in real-time.

Challenges we ran into

  • Context Management: To ensure the AI evolves based on the full history, we engineered a structured prompt that summarizes long-term stats while encouraging the model to prioritize recent user feedback for immediate adaptation.

  • JSON Mode vs. Search Tools: We found that strict JSON output is not compatible with the Google Search tool. To fix this, we split the pipeline into two API calls, one for real human text search using Google Search tool, and the second strictly for creative generation. This also effectively prevents the AI from fabricating fake citations when generating text."

Accomplishments that we're proud of

  • Gemini 3 Integration: We successfully moved beyond simple text generation to Data-Driven Reasoning, where the LLM acts as an analyst before acting as a writer.

  • The UI/UX: We achieved a high-fidelity Hacker Terminal with Flutter!

What we learned

  • Prompt Engineering is Always Important: Getting Gemini to return valid, strict JSON while maintaining a rogue AI persona required precise instructional tuning.

What's next for GENPTCHA

  • True Autonomy: We plan to integrate GitHub Actions to fully automate the daily evolution loop.

  • Global Localization: We plan to expand the this game beyond English, starting with Chinese.

  • Multimodal Battles (Image & Video): Leveraging Gemini 3's native multimodality, we will evolve the duel from text to visual media. A planned feature for the Image Duel is inspired by classic "Spot the Difference" games, users have to identify unreasonable pixel artifacts (e.g., distorted hands, unnatural lighting) in AI-generated images, training their eyes against Deepfakes.

Share this project:

Updates