Key Insights

• •

  Building AI products requires giving up control over user experience as outputs vary uniquely per user.

• •

  Traditional usability testing is insufficient for AI products; longitudinal testing with repeated sessions is critical.

• •

  Human-in-the-loop 'Wizard of Oz' testing effectively simulates AI to rapidly test assumptions before building.

• •

  Model drift in LLMs can cause unpredictable or inappropriate responses, requiring mechanisms like turn capping.

• •

  Users form evolving relationships with AI assistants, which need careful design and testing over time.

• •

  Many product orthodoxies based on fixed experiences expire with generative AI and must be reinvented.

• •

  Failure early and often is essential in AI product development due to its inherent unpredictability.

• •

  Users may lack mental models for AI assistants, making onboarding and user education a bigger challenge.

• •

  Asynchronous, multi-threaded conversations are natural for humans but difficult for bots to handle accurately.

• •

  AI can automate synthesis tasks, enabling humans to focus on creativity, accountability, and authenticity.

Notable Quotes