The Verification Bottleneck Arrives

GenAI tools promise to boost creative output, but scattered evidence makes it unclear whether they actually improve idea quality or just increase volume.

Method: Meta-analysis of empirical studies reveals GenAI increases creative output quantity but shows mixed effects on quality. The key mechanism: GenAI excels at divergent thinking tasks (generating many alternatives) but struggles with convergent tasks requiring evaluation and refinement. Human-AI collaboration outperforms either alone when humans retain editorial control—the AI generates raw material, humans curate and synthesize.

Caveats: Effect sizes vary wildly across domains and task types. What works for marketing copy may fail for product design.

Reflections: How does prolonged GenAI use affect baseline human creativity over time? · What collaboration patterns maximize quality while maintaining efficiency? · Can we predict which creative tasks benefit from AI augmentation versus pure human work?

Companies deploy algorithmic hiring tools with gender-balanced shortlists, yet final hires remain homogeneous. The shortlist constraint doesn't translate to diverse outcomes.

Method: The culprit is human-algorithm correlation: when hiring managers' preferences align with the algorithm's screening criteria, they systematically select the same candidate types from balanced shortlists, nullifying diversity gains. The solution inverts conventional wisdom—deliberately reduce correlation by using algorithms that screen on orthogonal criteria to human judgment. Example: if humans favor pedigree, train algorithms on work samples instead.

Caveats: Requires organizational willingness to hire candidates who don't match traditional manager preferences—a cultural, not just technical, shift.

Reflections: What's the optimal correlation coefficient between algorithm and human criteria? · How do you maintain quality standards while deliberately diverging from manager preferences? · Can this framework extend beyond gender to other diversity dimensions?

Mobile LLM agents automate phone tasks by interpreting screens and executing actions, but their broad permissions create attack surfaces that traditional apps don't have.

Method: Security analysis across system-level agents (YOYO), third-party agents (AutoDroid), and web agents reveals three exploit vectors: prompt injection via malicious UI elements, privilege escalation through chained actions, and data exfiltration via screenshot analysis. Concrete example: attackers embed invisible text in webpages that hijacks agent instructions, causing it to transfer funds or leak credentials—agents can't distinguish legitimate UI from adversarial content.

Caveats: Mitigation strategies significantly degrade agent autonomy and user experience, creating a fundamental security-convenience tradeoff.

Reflections: Can we develop UI authentication methods that agents can verify but attackers can't spoof? · What's the minimum viable permission set for useful mobile agents? · How do users mentally model agent trustworthiness versus traditional app permissions?