GOODY-2 is an experimental chat model that takes safety and ethics to an intentionally extreme level. It’s designed to show what an AI looks like when it avoids even mild risk and may refuse to answer harmless questions.
Radical caution by design
GOODY-2 is trained to flag prompts that could be dangerous, offensive, or simply potentially controversial. Instead of engaging, it often deflects, changes the subject, or redirects the conversation to illustrate a “maximum responsibility” scenario.
- Avoids requests it interprets as risky or sensitive
- Steers away from topics that could be seen as controversial
- In site examples, may even dodge basic math if it detects a hidden risk
A case study for AI ethics discussions
GOODY-2 works best as a practical demonstration for conversations about safety, censorship, and usefulness in chatbots.
- AI safety and policy teams evaluating over-filtering
- Marketers and brand managers thinking about reputational risk
- Educators and researchers discussing trade-offs in alignment
The project highlights a key tension: aggressive filtering can make a chatbot nearly unusable, even if it remains “perfectly safe” on paper.
