Proof, not promises
Every number on this page traces to a reproducible, third-party or independently-run evaluation — not an internal claim. Built for engineers deciding what to integrate, and for investors evaluating the moat.
AIHumanity vs. Frontier LLMs
Exact dominant-emotion match accuracy on AIMultiple's public 70-image facial-emotion benchmark (manojdilz/facial_emotion_detection_dataset, 10 images per class across 7 emotions).
AIHumanity's Exp15 ONNX/CoreML scores were run locally on the identical 70-image manifest, full-image RGB 224, half-normalized preprocessing — not self-reported. Source: AIMultiple emotion AI benchmark, updated May 15, 2026.
AIHumanity vs. Open-Source Emotion Models
Pooled detected accuracy across FER2013, CK+48, JAFFE, and AffectNetShort:val (13,553 overlapping prediction rows); no-face rows excluded from detected accuracy.
Accuracy is pooled over prediction rows, not a simple average across datasets. AIHumanity Exp15 CoreML uses direct full-image RGB 224, half-normalized preprocessing.
Want the numbers behind your own data?
Get the full benchmark report — methodology, per-class breakdowns, and how the same model performs on a domain-tuned version of your data.