Capability Isn't the Bottleneck Anymore

Video See-Through headsets cause cybersickness because raw camera feeds distort scale perception and exaggerate motion parallax. Users get nauseous from the mismatch between what they see and what their vestibular system expects.

Method: Geometry-aware passthrough renders the real world using depth maps and 3D reconstruction instead of raw video feeds. The system corrects for the physical offset between cameras and eyes (interpupillary distance mismatch) and eliminates the scale distortion that makes objects appear closer or farther than they are. In a 36-participant study, this reduced cybersickness scores by 31% compared to standard passthrough while maintaining the same visual quality.

Caveats: Requires accurate depth sensing hardware. Performance degrades in environments with reflective surfaces or poor lighting where depth reconstruction fails.

Reflections: Does the cybersickness reduction persist across multi-hour sessions, or does adaptation occur? · Can simplified depth correction (e.g., planar approximation) achieve similar results with lower computational cost? · How does this interact with other cybersickness mitigation techniques like dynamic FOV restriction?

In bandwidth-constrained regions, web search fails because results are formatted for desktop browsers and require loading heavy web pages. Teachers in Sierra Leone couldn't get practical lesson plans from Google.

Method: A WhatsApp-based AI chatbot delivered text-only responses instead of web links. Over 17 months tracking 40,350 queries from 529 teachers, the AI responses scored 2.3x higher on relevance and 4.1x higher on actionability compared to google.com.sl results. The key mechanism: bypassing HTML rendering entirely and delivering markdown-formatted text that loads instantly on 2G connections.

Caveats: Study focused on teachers with specific pedagogical queries. Generalizability to other domains (health, commerce) and query types (transactional, navigational) remains untested.

Reflections: Does the relevance advantage hold for non-educational queries where web search excels (e.g., local business information)? · What happens when AI responses require visual content (diagrams, maps) that can't be text-only? · Can hybrid approaches (AI triage + selective web links) optimize for both bandwidth and comprehensiveness?

Diffusion models generate images that look photorealistic at first glance, but contain subtle artifacts that reveal their synthetic origin. The problem: we don't know which artifacts humans actually notice.

Method: Analyzed 749,828 human judgments on 450 AI-generated and 149 real images from 50,444 participants. Detection accuracy was only 59.7%—barely above chance. The most detectable artifacts were anatomical implausibilities (malformed hands, wrong number of fingers) and physically impossible lighting. Photorealistic textures and correct perspective fooled participants even when other elements were wrong. Crucially, participants who left comments explaining their reasoning had 12% higher accuracy than those who didn't, suggesting detection requires deliberate scrutiny.

Caveats: Study used static images. Detection rates may differ for video where temporal inconsistencies become visible, or in professional contexts where stakes are higher.

Reflections: Do detection rates improve with training, or is the 60% ceiling fundamental to human perception? · Which artifacts are most detectable in video versus static images? · Can automated tools replicate the 12% accuracy boost from deliberate scrutiny?