Generative AI strategy, user experience, and consulting. San Francisco, California

Conversational AI

The arrival of LLMs with ChatGPT completely disrupted chatbot design and conversational interfaces generally. No longer was NLP the primary method for design and implementation of voice and chat-based applications. LLMs now offer solutions — and complications — for the interaction designer and AI developer.

I approach conversational AI from a communication-centric position. I believe that technology should seek to meet the user on his/her terms. AI should try to speak and chat as naturally as possible so that users can employ their learned social and communicative competencies. Language and speech are, after all, how we get many things done. AI should try to approximate an understanding of user intentions rather than require that users learn how to speak like machines.

Conversational AI covers both task-oriented dialog and open dialog. NLP has shaped task-oriented dialog and shows us that structure is necessary for AI to satisfy many functional and task-oriented interactions. LLMs are naturally poor at structure. Helpful but imperfect solutions come from use of RAG (retrieval augmented generation), agentic architectures, fine-tuning, prompting, knowledge graphs, and more.

User experience design has not historically included open dialog under its purview. In part this is because conversational interaction is non-visual and unstructured. There is no simple way to systematize the functional flow of conversation in a manner akin to the organization of information navigation in conventional web browsing, for example.