Research & Experiments
AI UX Experiments
I run structured experiments on how people interact with AI-powered interfaces — with a particular focus on trust, disclosure, and how users recover when the AI is wrong.
AI transparency is a UX problem
Whether to disclose AI involvement, when to show confidence levels, and how to handle contradictory outputs are design decisions, not just engineering ones.
Failure modes are the real design surface
Most AI UX work focuses on the happy path. I focus on the 20% of interactions where the model is wrong, uncertain, or misses intent — that's where trust is built or broken.
Test with real uncertainty
Usability testing AI features requires deliberately inducing failure states and watching how people respond. Staged tests where the AI always succeeds tell you nothing.
Experiments
Patterns for AI Disclosure in Consumer Interfaces
When should a UI tell a user that what they're seeing is AI-generated? I tested 12 disclosure patterns across product recommendation, search summaries, and content suggestion surfaces. Spoiler: timing matters more than wording.
When AI Gets It Wrong: Designing Graceful Fallbacks
LLMs fail interestingly — they hallucinate, misread intent, and hedge when they shouldn't. This experiment maps the failure modes I encountered in 6 months of shipping AI features and what UI patterns best recover user trust.
30 Days of AI-Assisted Wireframing
I used Claude, Midjourney, and a prompt-to-Figma pipeline to assist wireframing for a real client project. A daily log of what worked, what didn't, and where human judgment was irreplaceable.
Trust Signals in AI Chatbots
Analysing the micro-copy, interaction timing, and visual language that correlates with users rating a chatbot as 'helpful and reliable' vs 'untrustworthy.' Based on a 200-participant remote study.
Methodology note
These experiments are not academic research — they're practitioner-led, design-focused investigations. Sample sizes are typically 20–50 participants per study, recruited from user panels or network screeners. I publish methods alongside findings so they can be critiqued and replicated.