Creative AI Works Better When Humans Lead the Conversation

LLM interfaces for creative work default to model-led modes where the AI generates content proactively. Unclear whether human-led modes—where the AI asks questions or offers suggestions instead—produce better creative outcomes.

Method: Randomized controlled experiment with 486 participants compared three collaboration modes: human-led with AI suggestions, human-led with AI questions, and model-led with AI generating content. Measured creative output quality, user satisfaction, and perceived collaboration effectiveness across all three conditions.

Caveats: Single-session creative writing task. Long-term effects on creative skill development or professional workflows unknown.

Reflections: Do human-led modes remain superior across different creative domains (visual design, music composition, code architecture)? · How does the quality gap between human-led and model-led modes change as users gain experience with AI tools? · Can hybrid modes that switch between human-led and model-led based on task phase outperform either pure approach?

Generic LLM assistants ignore individual differences in expertise, thinking style, and psychological attributes. Personalization based on psychometric profiles might unlock better human-AI synergy in creative collaboration.

Method: Built an LLM assistant informed by users' psychometric profiles (demographics, psychological attributes, divergent thinking scores, domain expertise). Tested against baseline LLM in multi-turn creative collaboration tasks. Measured collaboration quality, creative output, and user satisfaction across personalized versus generic conditions.

Caveats: Psychometric profiling adds setup friction. Unclear whether benefits justify the overhead for casual or one-off creative tasks.

Reflections: Which psychometric dimensions contribute most to personalization gains—can we reduce the profiling burden? · Do personalization benefits compound over repeated sessions as the model learns user preferences? · How do users react to AI that explicitly references their psychological profile in its suggestions?

GDPR requires cookie banners to present accept and reject options equally. Many banners use visual manipulation to draw attention to the accept button, but manual auditing doesn't scale.

Method: Applied salient object detection algorithms to 2,579 websites' cookie banners. Measured visual attention bias by comparing saliency scores of accept versus reject buttons. Classified banners as manipulative when accept buttons received significantly higher saliency scores than reject buttons.

Caveats: Saliency detection approximates human attention but doesn't capture reading patterns or cultural differences in visual processing.

Reflections: Do users actually click accept buttons more often on high-saliency banners, or do they ignore visual manipulation? · Can designers use saliency equalization tools to self-audit banners before deployment? · How do saliency-based manipulation tactics interact with banner fatigue and habituation over time?