Adaptive Intelligence
How can the use of LLMs be optimized through user behavior?
Most users default to starting a new chat rather than revisiting previous conversations. At Google scale, this creates billions of similar conversations with overlapping intent — repeated AI generation without added user value.
This project explores how interaction design in Gemini can reduce that waste by surfacing relevant past conversations before generating new responses.
Users restart instead of rediscovering
Many users default to starting a new chat rather than revisiting previous conversations. This creates many similar conversations with overlapping intent. At Google scale, small inefficiencies repeat billions of times.
Repeated AI generation without added user value is a sustainability and efficiency problem hiding in plain sight. Improving how users rediscover existing conversations aligns directly with Google's strength and mission.
Re-discovery before generation
A search-first layer inside Gemini that routes simple, factual, and repeatable queries to keyword-based retrieval before invoking full chat intelligence. Before generating a new response, Gemini checks previous chats for similar prompts and surfaces them to the user.
A standard Gemini query costs 0.24 Wh. Indexed keyword matching over existing chats costs 0.001–0.01 kWh, orders of magnitude less. By routing familiar queries to retrieval first, the intervention reduces redundant generation and surfaces answers the user already has.
The energy cost of repeated queries
A single standard Gemini query consumes 0.24 Wh — roughly 10x the energy of a Google Search. Across 2B+ daily queries, that adds up quickly. 100 Gemini queries consume approximately 24 Wh: enough to drive 450 feet, toast 2 slices of bread, or stream Netflix for 15 minutes.
A large portion of Gemini usage comes from repeated, everyday tasks. Redundant prompts asking the same or similar questions multiple times consume unnecessary compute and energy. Repeated tasks are an opportunity for smarter reuse — and a well-placed design intervention can shift that behavior at scale.
Energy equivalence100 Gemini queries ≈ ~24 Wh of electricity.
From research to intervention
Key InsightsThree behaviors surfaced consistently across user interviews:
Repeat
Users return for the same tasks daily.
Restart
Instead of continuing past conversations, users start new prompts.
Forget
Past answers are rarely revisited even when still relevant.
Next Steps
Test with real usage data to better measure the impact of search-first retrieval on redundant query reduction.
Expand keyword-matching to semantic similarity so the intervention catches conceptually related prompts, not just exact repeats.
Evaluate adoption patterns across different Gemini surfaces — standalone chat, Workspace, and embedded contexts — to understand where the intervention has the greatest effect.
Takeaways
Interaction decisions have sustainability impact
Individual design choices — where a button lives, what a prompt says — can influence behavior at a scale that translates into measurable energy savings. Sustainability in AI is not just an infrastructure problem; it is a design problem.
Behavior change starts with meeting users where they are
Users restart conversations not out of laziness but because rediscovery is harder than restarting. The right intervention does not ask users to change their habit — it makes the better option feel easier and more obvious than the default.
Efficiency and experience are not in conflict
A search-first layer does not degrade the experience — it improves it by surfacing what users already found useful. Efficiency and helpfulness point in the same direction when the design is grounded in how people actually use the product.
All assets — tell me where each one goes.
Every asset in gemini/ is shown below with its filename. Tell me which figure slot each belongs to (Fig 1.0, Fig 1.1, etc.) and I'll wire them in.
