Stanford CS547 HCI Seminar | Spring 2026 | Just-in-Time Objectives for Specialized AI Interactions

Current AI systems often provide generic outputs because they are designed as rigid "sporks" that cannot adapt to changing user contexts. A "Just-in-Time" (JIT) framework improves this by automatically observing a user's current actions, inferring their goals, and dynamically shaping AI tools to match those needs.

• The "Spork" limitation — Current AI tools combine many features into one interface, causing generic results that fail to adapt to distinct user goals [01:00].

• Just-in-Time objectives — Instead of relying on manual prompting, systems observe interaction traces like screenshots or web layouts to infer intent and tailor performance in the moment [07:13].

• Custom UI generation — The Poppins system builds temporary software tools on the fly based on visible work, including:

  • Presentation planners for slide outlines [16:35].
  • Interactive diagram editors for research notes [17:20].

• Steerable topic modeling — The Loom tool lets users guide text analysis using clear concepts instead of abstract keywords, improving accuracy for researchers [21:53].

• Social science integration — Researchers can apply academic frameworks, such as those for reducing partisan hostility, to adjust social media algorithms and improve outcomes [31:02].

• User-owned AI — Shifting control allows individuals to manage the underlying assumptions of their AI, moving toward a "Mary Poppins bag" model where the correct tool surfaces automatically [35:55].

• How can AI systems protect user privacy while observing interactions?

• What role does social science play in the design of future AI?