AI Systems Become Infrastructure Before Users Can Verify Them

MLLMs fail at professional software interfaces. Targets are 3-5x smaller than consumer apps, and 4K displays expose models trained on low-res screenshots.

Method: ScreenSpot-Pro tests GUI agents on Photoshop, Blender, and CAD tools at native resolutions up to 3840x2160. Current best models hit only 25-40% accuracy on small UI elements (buttons under 50px), compared to 70%+ on consumer interfaces. The benchmark isolates three failure modes: resolution degradation during preprocessing, inability to parse dense toolbars, and confusion in multi-window layouts with overlapping coordinate systems.

Caveats: Benchmark focuses on static screenshots, not dynamic interactions like drag-and-drop or modal dialogs that appear mid-workflow.

Reflections: Can fine-tuning on high-res professional UI datasets close the 30-point accuracy gap? · How do coordinate system transformations in multi-monitor setups affect grounding performance? · What's the minimum viable resolution for acceptable accuracy in dense interfaces?

Anger drives disinformation sharing. Users repost false content when emotionally activated, bypassing deliberation entirely.

Method: Two nudges outperformed controls: an emotion-labeling prompt ('You seem angry—take a moment') and a reappraisal nudge ('Consider another perspective'). The emotion-labeling intervention reduced sharing intent by 18% when users reported strong anger (p<0.01). The mechanism is cognitive reappraisal—forcing users to name their emotion creates a 3-5 second pause that disrupts automatic sharing. Critically, the nudge only worked for high-anger content; neutral posts saw no effect.

Caveats: Study used hypothetical sharing scenarios, not real platform behavior. Habituation effects unknown after repeated exposure.

Reflections: Does the nudge lose effectiveness after 5-10 exposures as users learn to dismiss it? · Can real-time sentiment analysis accurately flag anger without misclassifying passionate agreement? · What's the optimal delay duration before habituation sets in?

Dyslexic readers struggle with phonological decoding and working memory load. Existing tools either distort meaning (text-to-speech) or require expensive tutoring.

Method: LARF uses GPT-4 to pre-annotate text with three layers: phonetic breakdowns for complex words, contextual synonyms for ambiguous terms, and sentence-level summaries for long passages. In a 42-participant study, dyslexic readers using LARF improved comprehension scores by 23% and reduced reading time by 19% compared to plain text. The key mechanism is cognitive offloading—annotations reduce working memory demands by externalizing phonological and semantic processing.

Caveats: Tested only on English narrative text. Technical documentation and non-Latin scripts untested. Annotation density may overwhelm some users.

Reflections: What's the optimal annotation density before cognitive load increases instead of decreases? · Can users customize which annotation types to display based on individual dyslexia profiles? · Does long-term use improve unaided reading ability or create dependency?