The academic case for MindLens Lab.

MindLens Lab is the follow-up to a literature review published as “Artificial Intelligence in Emotional Intelligence Training for Autism” (Kim, 2025, Curieux Academic Journal; also on the UNESCO Learning Planet Institute Youth Fellow platform). The review evaluated four AI tools designed to help autistic users build emotional intelligence. Across very different designs — wearable, classroom platform, recognition engine, social robot — the same four limitations kept appearing: an annotation bottleneck, weak transfer to real-world settings, risk of overdependence on AI feedback, and unresolved ethical concerns around emotion data.

The natural next step felt like building a fifth tool. Several were sketched — a wearable, a classroom platform, a conversational AI tutor. Each sketch inherited the same data substrate the review had spent thirty pages critiquing. So Phase 1 of MindLens Lab shifted: instead of another tool on top of single-label-per-moment training data, build the substrate from the ground up.

Researchers and practitioners broadly accept that emotion reading is layered and contextual; people misread each other, and many viewers see more than one thing at once. But the data emotion AI trains on is still labelled one-emotion-per-moment, and the benchmarks grade against single answers. The concept and the substrate don't match — and the tools built on that substrate inherit the mismatch, no matter how sophisticated they become.

MindLens Lab Phase 1 builds the missing substrate: many readers, the same short social moments, structured carefully enough that later phases — validated study materials and adaptive tools for the people the project was designed to serve — can rest on it.

Five formal documents anchor the project. The originating paper named what the field was missing. The Phase 1 pre-registered plan committed the analytic frame in advance. Two Phase 1 preprints report what the resulting dataset shows — one on the human side (accepted at Curieux), one on the AI side. And a Phase 2 pre-registration commits, before analysis, to how four vision-language models will be compared against the plural human distributions that Phase 1 established.

A note on terminology: a preprint is a manuscript posted publicly (with a citable DOI) before peer review; a paper in the strict sense has been accepted through a journal's peer review. The P1 human-side preprint is now both — accepted at Curieux, still linkable as a Zenodo preprint. The other three are preprints / pre-registrations.