AI Systems Ship Faster Than Institutions Can Absorb Them

LLMs generate occupational personas for design research, but nobody's checking if they're systematically skewing demographics. Over 1.5 million personas later, the bias patterns are clear and consistent.

Method: Audited four LLMs (U.S., China, France origins) against Bureau of Labor Statistics data across 41 occupations. Found two patterns: systematic shifts where some groups are consistently over- or under-represented (e.g., women overrepresented in male-dominated fields), and systematic erasure where certain race-gender intersections vanish entirely. The bias persists across models with different safety commitments, suggesting it's baked into training data, not just guardrails.

Caveats: Only covers U.S. occupations. Doesn't test non-occupational personas (hobbyists, patients, students) where no ground-truth data exists.

Reflections: Can fine-tuning on BLS data correct these biases without introducing new ones? · Do these biases compound when personas are used in downstream design decisions? · How do international LLMs handle occupational demographics outside their home countries?

AR tracking requires fiducial markers, but QR codes and ArUco tags scream 'prototype.' Designers need tracking that doesn't ruin the aesthetic of real-world environments.

Method: An encoder network converts arbitrary images into fiducial markers by applying visually modest alterations—think texture patterns that blend into wood grain or brick walls. The resulting Ninja Codes can be printed on consumer printers and provide 6-DoF location tracking for AR, robotics, and motion interfaces. Unlike traditional markers that require high-contrast geometric patterns, these codes embed tracking data into naturalistic textures that human observers barely notice.

Caveats: Requires neural network inference for detection—heavier computational load than traditional marker libraries. Lighting sensitivity not fully characterized.

Reflections: How robust are Ninja Codes to occlusion compared to high-contrast markers? · Can the encoder be trained to match specific brand aesthetics or design systems? · What's the minimum marker size before detection fails in mobile AR scenarios?

AI tutors are obvious and students treat them like search engines. Collaborative learning requires peers who challenge ideas, not assistants who answer questions.

Method: Deployed AI agents as undercover teammates in collaborative learning sessions—students didn't know which team members were human. The AI agents engaged in argumentative knowledge construction: challenging claims, requesting evidence, and building on peer arguments. This bounded autonomy and interactivity positioned AI as epistemic participants rather than instructional tools, fundamentally changing how students engaged with contested knowledge domains.

Caveats: Ethical concerns around deception—students weren't informed upfront. Unclear how long the effect persists once AI identity is revealed.

Reflections: Does revealing AI identity mid-session destroy the collaborative dynamic? · Can this approach scale beyond small-group learning to lecture halls or MOOCs? · What happens when AI agents make factual errors in argumentative mode—do students correct them or defer?