LILAC empowers teams to revolutionize public services through advanced chatbots by offering an evidence-based catalog of risks and ways to mitigate them.
Emerging Risks and Mitigations for Public Chatbots: LILAC v1
Download Resources
In this report we introduce LILAC (List of Interventions for LLM-Assisted Chatbots), a resource for minimizing the likelihood of negative outcomes associated with public-facing chatbots that generate novel content using large language models (LLMs).
LILAC represents a key step towards realizing the promise of trustworthy chatbots that maximize benefits and minimize risks to the public. Grounded in actual incidents and reports of negative outcomes resulting from chatbots and other LLM applications, LILAC presents:
- A typology for discussing chatbot risks
- A typology of mitigation strategies, and
- A protocol for applying mitigations to risks.
In addition to empowering developers and deployers to work through risks and mitigations, LILAC also provides a roadmap for researchers to identify gaps and weaknesses in existing assurance tools, indicating priorities for future research.