Static Evaluation Fails When Systems Adapt at Runtime

Generative AI is replacing traditional search results, but nobody knows how it changes which sources get traffic or whether it retrieves the same information.

Method: Analyzed 11,500 real-user queries across Google Search, AI Overviews (AIO), and Gemini. AIOs appear for 51.5% of queries, displayed above organic results. Retrieved sources show <0.2 Jaccard similarity between systems. Traditional Google significantly favors institutional sites (.gov, .edu); generative search significantly favors Google-owned content. Websites blocking Google's AI crawler are significantly less likely to appear in AIOs despite content availability.

Caveats: Single snapshot study. Doesn't track how these patterns evolve as Google adjusts AIO algorithms or revenue models.

Reflections: Do websites that allow AI crawling see measurable traffic increases, or does AIO cannibalize clicks regardless? · How do generative engine optimization techniques perform when sources differ so dramatically across systems? · What revenue framework could sustain publishers when generative search surfaces their content without click-through?

GDPR mandates data minimization, but regulators lack scalable methods to determine what data access is actually necessary. Expert audits don't scale and may not reflect user expectations.

Method: Built DePRa, a system that repositions everyday users as privacy evaluators through contextual explanations, category-based selection, intuitive rating interface, and preference-based adjustment. Tested with 200 mobile app users rating sensitive data access. System successfully captured user privacy opinions and enabled comparison with expert assessments, demonstrating feasibility of democratized privacy evaluation.

Caveats: Doesn't validate whether user ratings predict actual privacy harms or behavioral outcomes. Feasibility shown, but effectiveness unproven.

Reflections: Do user privacy ratings correlate with actual harm outcomes (breaches, misuse, behavioral manipulation)? · How stable are user ratings across time and context—do they reflect durable preferences or momentary reactions? · Can democratized ratings influence app developer behavior or regulatory enforcement at scale?

Blind and low-vision users can describe immediate surroundings with assistive tech, but have no way to track how environments change across revisits—creating safety risks and cognitive burden.

Method: StateScribe uses dual-layer memory (episodic scene + object-centric temporal) to track real-world changes. Maintains F1-score of 83.1% across 11 revisits while staying under 1.54s latency and 54MB memory across 110 revisits. Provides live scene descriptions plus temporal change summaries like "The shop on your right has a 'CLOSED' sign; it was open at this time last week." User study with nine BLV participants showed improved change awareness in three real-world locations.

Caveats: Tested in three locations over limited revisits. Long-term memory decay, privacy implications, and multi-user environments unexplored.

Reflections: How should the system prioritize which changes to surface when memory grows large—by recency, safety relevance, or user intent? · What privacy controls do users need when AI systems build persistent spatial memory of their routines? · Can temporal memory extend beyond changes to other capabilities like predictive wayfinding or social context?