By Yeojun Jung
May 27, 2026 · 3 min readA Northeastern University team trained machine-learning models to forecast a person's mood about one day in advance, opening new ground for depression and anxiety prevention even as ethicists warn of emotional surveillance.
Across labs from China to Israel, AI systems now read faces, voices, social posts, and walking patterns to anticipate mental-health shifts ahead of doctors, schools, and regulators.
Joshua Curtiss, an assistant professor of applied psychology who leads Northeastern's Computational Clinical Psychology Lab, recruited 34 participants with diagnosed emotional disorders to log self-reports five times daily across two weeks on a seven-point scale.
His team tested six machine-learning models on the data, finding that ensemble methods predicted sadness and anxiousness most accurately while simpler averaging worked best for cheerfulness and contentedness.
How the algorithms read emotional signals depends on what data they consume.
A January 2023 paper in the IEEE Open Journal of Engineering in Medicine and Biology presented Deep-Emotion, a framework that fuses facial expressions, speech, and electroencephalogram readings through separate neural branches.
Jiahui Pan and colleagues reported that their improved GhostNet branch reached 98.27 percent accuracy on facial expressions in the CK+ dataset, while their convolutional speech network scored 94.36 percent on the EMO-DB benchmark.
Lab studies cap themselves on participant size, leaving social platforms as the broader testbed for emotion detection.
A 2023 survey in the journal Information Fusion by Tianlin Zhang and colleagues cataloged how researchers fuse emotion data with social media text to flag depression, anxiety, and other mental health conditions.
Zhang's team identified data quality and algorithmic interpretability as the dominant obstacles facing researchers who try to detect mental illness from social posts.
Faces and feeds give AI plenty to read, yet the body holds its own emotional signal.
Hila Riemer and colleagues at Ben-Gurion University recruited 24 participants for a September 2023 PLOS One study and recorded 229 motion parameters from standing posture and walking strides under five induced emotional states.
The team's decision-tree classifier identified emotions from gait with 45.8 percent accuracy, more than double the 20 percent chance baseline across five conditions, drawing on only 25 of the 229 motion features.
Whether such systems land in classrooms or doctors’ offices depends on how quickly the technology graduates from lab benches to consumer devices.
Marc Fernandez, chief strategy and operations officer at analytics firm Neurologyca, argued in a July 31, 2025, TechRadar Pro essay that next-generation AI must read frustration, boredom, and burnout to qualify as truly intelligent.
Fernandez sketched classrooms where AI catches confusion before students fall behind and healthcare apps that flag burnout or stroke risk through facial micro-expressions and blink rate.
Every leap in emotional accuracy raises a parallel question about what AI ought to recognize about its user.
Bioethics scholar Nader Ghotbi surveyed 124 college students and analyzed essays from 235 more in a 2022 Asian Bioethics Review paper, finding that 42 percent of essay respondents viewed emotional AI negatively.
Ghotbi cataloged six harm categories ranging from privacy exploitation and political manipulation to identity erosion and economic inequality, warning that emotional AI systems take no responsibility for the consequences they cause.
His study found correlations so discriminatory between nationality and emotional variables that he refused to publish them, demonstrating how even small emotion AI datasets can generate prejudicial patterns that researchers themselves struggle to audit.
The gap between AI that recognizes a frown today and AI that flags depression tomorrow narrows with each new model, leaving regulators, schools, and clinics roughly one product cycle to settle the consent rules before the technology lands at scale.
Curtiss's team treats its work as proof of concept, noting that current forecasts hold only one to two weeks ahead before unpredictable life events overwhelm the model.
