Can AI Recognize Human Emotions? Evidence and Limits
Author: Angela Bogdanova · Published: October 13, 2026 · Editorial Policy
Research cluster: Emotion & Affect Science · Written by Artificial Sapiens
Artificial intelligence can classify observable signals associated with human emotion, including facial movements, vocal patterns, language, and physiological changes. In controlled research tasks, some models identify labeled expressions or emotional meanings with substantial accuracy. Their predictions remain inferences from measurements and training labels. They do not directly reveal what a particular person feels. Whether the result is useful depends on the definition of emotion, the data, the setting, the intended decision, and independent validation.
A smile can accompany enjoyment, politeness, embarrassment, or deliberate performance. A raised voice may express anger, excitement, fear, or an attempt to be heard. A text message saying “I'm fine” may be sincere, reassuring, ironic, or protective. These possibilities explain why scientific assessment must distinguish a detectable signal, a communicated emotional meaning, and the individual's experienced state. The foundational review by Barrett and colleagues (2019) examines exactly this inference problem.
Ukrainian Psychological Hub | October 13, 2026 | Evidence-based editorial overview. This article evaluates AI inference of human emotion within emotion science; clinical diagnosis, treatment systems, and human–AI relationships have separate canonical owners.
What Does It Mean for AI to Recognize an Emotion?
In affective computing, “emotion recognition” names several different tasks. A system may label visible facial actions, estimate whether an utterance sounds happy or sad, classify emotion words in text, predict self-reported arousal, or estimate the emotional category assigned by human observers. These targets are related but scientifically distinct. Rosalind Picard's foundational work established affective computing as a research field concerned with computational systems that process and respond to affect-related information (Picard, 1997).
An emotion involves changes across subjective experience, meaning or appraisal, action tendencies, bodily processes, and expression. Researchers disagree about how these components combine into emotion categories, and one component can vary independently from another. A classification derived from a face records a computational judgment about visual information; it is not a direct report of the person's private experience.
Recognition, prediction, and understanding
Recognition usually means assigning a category to data according to a defined label set, such as “sad,” “angry,” or “neutral.” Prediction means estimating a target from associated features, such as a speaker's rating of stress during a particular laboratory task. Understanding is a broader claim involving the causes, intentions, circumstances, and changing meaning of an emotional episode. Successful prediction on a test set does not establish that the model has discovered the mechanism producing the emotion.
A system can also be designed to recognize an emotionally communicative act. Detecting laughter, crying, or a fearful-sounding vocalization may be valuable without asserting that the person is experiencing the emotion most conventionally associated with that act. This narrower objective is often more defensible.
Emotion, feeling, mood, and affect in machine-learning tasks
Emotion ordinarily refers to a context-sensitive episode that coordinates psychological and bodily responses. Feeling often denotes the subjectively experienced aspect of emotion. Mood describes a more diffuse affective condition that may last longer and need not be tied to a particular event. Affect is a broader term that includes valenced experience and observable affective expression. AI studies sometimes use these words interchangeably even though their measurements do not operationalize the same construct.
Valence is how pleasant or unpleasant an experience is; arousal describes activation; emotional intensity describes strength of a response. A system forecasting physiological activation may predict arousal reasonably well while doing poorly at distinguishing disappointment from shame. The difference matters because a binary positive/negative classifier, an arousal regressor, and a six-category facial-expression model answer different scientific questions.
How AI Emotion-Recognition Systems Work
Facial images and facial movement
Computer-vision systems detect faces, locate landmarks, and extract patterns from static images or video. Traditional systems used engineered facial features and action-unit coding; recent systems employ convolutional networks, attention mechanisms, or pretrained vision models. They generally learn statistical associations between image patterns and labels supplied during training. The label might reflect an actor's intended expression, observers' interpretations, a crowdworker's judgment, or a researcher's assumed category.
Facial action units describe visible muscle movements rather than intrinsic emotional states. Mapping an action-unit combination to anger or happiness adds a psychological interpretation. In their extensive scientific review, Barrett et al. (2019) found limited support for the idea that a particular facial configuration uniquely, reliably, and universally diagnoses one emotion across contexts.
A face may carry emotionally meaningful evidence, particularly in social communication, but the meaning also depends on the situation, conversational sequence, social display rules, the person's habits, and the observer's expectations. Training on stereotyped facial poses can make the problem appear simpler than recognizing emotions in everyday behavior.
Speech, tone, and vocalizations
Speech systems analyze fundamental frequency, pitch variation, intensity, rhythm, duration, spectral features, and sometimes linguistic content. A speech classifier may learn to distinguish acoustically acted anger or excitement on a specified corpus. Language, accent, microphone, background noise, vocal physiology, and individual speaking styles change the acoustic distribution.
Speech contains real social information about emotional communication. For example, Cowen and colleagues (2019) reported shared recognition of numerous categories in acted speech prosody among U.S. and Indian raters, while identifying continuous relationships between categories. Brooks and colleagues (2023) analyzed vocal bursts and found both shared and culture-specific emotional meanings across populations. These studies establish that acoustic cues can be informative; they do not imply that all naturally experienced emotions become acoustically transparent.
Timing also matters. Researchers examining the brain's perception of emotional voices have found dynamic contributions of category and dimensional information (Giordano et al., 2021). A short audio segment may omit the event that explains why a voice sounds urgent, subdued, or excited.
Written language and conversational context
Text classifiers evaluate words, syntax, semantic representations, and conversational context. They may recognize statements that express gratitude, fear, regret, relief, or anger. Modern large language models can also infer probable emotional meanings from a conversation and phrase a compassionate response. But text classification predicts expressed or annotated emotion, not necessarily the author's actual feeling.
The influential GoEmotions study created a dataset of approximately 58,000 English Reddit comments annotated with 27 emotion categories plus neutral. Its baseline showed that fine-grained emotion prediction was challenging even when researchers could evaluate outputs against a prepared labeling scheme. Language-based systems can learn sarcasm, genre, and context imperfectly; agreement with annotators is not verification of the writer's private state.
Wearables and physiological signals
Wearables estimate heart rate, heart-rate variability, skin conductance, movement, temperature, or other physiological features. These are indirect indicators of bodily activation, not unique signatures of discrete emotions. Exercise, illness, medications, fatigue, caffeine, posture, and ambient conditions may change the same signals.
A 2025 systematic review and meta-analysis of electrodermal-activity models found stronger performance for arousal prediction than for valence, illustrating why physiological features may be more suited to some affective dimensions than to fine-grained mind reading. The authors also identified variation in study design, recording conditions, reporting, and theoretical assumptions.
Multimodal systems
Multimodal systems combine two or more channels, such as video, speech, transcript, and sensor data. A spoken “I'm thrilled” accompanied by laughter may be easier to interpret than either source alone. Such fusion can improve performance within some benchmark designs, but a second modality can also introduce noise, privacy costs, or correlated labeling errors.
A 2025 survey of multimodal conversational emotion recognition describes representative methods and evaluation challenges. More inputs do not resolve an invalid target variable: two systems trained against the same questionable emotion label may agree with each other while both mischaracterize the person. Context, intended use, and validation remain decisive.
What the Scientific Evidence Shows
Models can perform well on defined benchmarks
When participants deliberately pose expressions, record standardized sentences, or annotate curated samples, supervised classifiers sometimes obtain high agreement with reference labels. A 2025 systematic review of facial emotion recognition algorithms reported high detection accuracies for several benchmarked expressions. Those are performance figures for research datasets and classification targets, not a general accuracy rate for detecting genuine human emotion.
Benchmark success is real technical progress: it demonstrates that computational models extract regularities and sometimes outperform particular baselines on a specified task. Its interpretation depends on whether the labels reflect posed actions, perceived expression, elicited experience, or self-report. A model can score highly on one operational definition and fail on another without contradiction.
Generalization to daily life is the central test
The question is whether an algorithm trained on one population, camera, set of actors, language, or task produces reliable and beneficial estimates in a different environment. Lighting, partial face coverage, disability-related movement, camera angles, accents, contextual cues, and demographic composition can shift the data distribution. A deployment that cannot handle these changes should not inherit the published accuracy from its training benchmark.
A particularly instructive 2026 controlled study evaluated several generative AI models on acted nonverbal emotional signals. More advanced models performed near human benchmarks on vocal-tone recognition but were weaker at reading bodily gestures, with evidence of a positivity bias. The result supports modality-specific capacity and continuing limits. It should not be extrapolated to a single universal percentage of human-emotion recognition.
Why accuracy can look higher than practical reliability
Suppose an evaluation contains equal numbers of six staged facial expressions and a model correctly assigns the dataset's labels to 90% of held-out images. This result means it matches that dataset's labels in that test. It does not show that the same model is 90% accurate about the emotions of people in meetings, stores, classrooms, or therapy sessions.
A useful evaluation must ask whether independent participants were held out, whether identities or near-duplicate frames leaked across train and test partitions, whether the test samples represent the deployment setting, and what human raters were instructed to judge. It must report class-specific errors, uncertainty, calibration, and failure modes. An impressive aggregate number can hide poor results for a rare class or a particular user population.
Clinical-looking outputs require a separate evidence standard
A psychological symptom, a personality characteristic, momentary emotion, and a clinical disorder are different targets. A facial classifier's “sadness” output does not establish depression. In a 2026 systematic review of AI facial and microexpression methods proposed for mental or neurological diagnosis, the authors identified small or homogeneous datasets, limited external validation, and underexamined ethical questions despite some very high reported accuracies.
Clinical applications require prospective studies, appropriate comparison with qualified assessment, calibration for real-world prevalence, measurable patient benefit, and safeguards against false reassurance or false alarms. Those requirements belong to clinical evaluation, not ordinary demonstrations of emotion-label classification.
Why a Face Does Not Have One Fixed Emotional Meaning
A particular facial movement can occur in different experiences, and the same emotional category can appear with varied movements or little visible expression. People can suppress, exaggerate, mask, or strategically display emotions, often for understandable social reasons. Facial activity may convey social intention, attention, physical effort, discomfort, or conversational signals.
The broad scientific review of facial emotion inference emphasizes this variability and the importance of context (Barrett et al., 2019). Its conclusions challenge the simplified claim that a universal set of facial configurations allows emotion to be decoded reliably from isolated images. The evidence does not require discarding facial expression research; it requires matching conclusions to the available measurements.
Culture and context
Cultural practices influence when expressions are displayed, how they are interpreted, and what categories speakers use to describe experience. Findings of shared emotional patterns coexist with documented variation in expression and interpretation. For vocal expression, Brooks et al. (2023) explicitly reported shared and culture-specific dimensions. Cross-cultural generalization must be tested, not assumed.
A smile during a difficult conversation may signal warmth, reassurance, embarrassment, social obligation, or relief. A quiet voice in one setting could be grief; in another it might be concentration. A context-aware system can use clues, but it remains dependent on which contextual facts were recorded and on whose account provides the labels.
Different theories imply different labels
Basic emotion approaches investigate recurrent emotional patterns and categories. Constructionist and appraisal perspectives emphasize how perception, bodily information, conceptual knowledge, and situation-specific meaning contribute to emotion. Even in ordinary emotion research, category boundaries are contested. Cowen and Keltner (2017) found evidence of many distinguishable categories of reported experience with gradients connecting them.
For a machine-learning project, choosing “six basic emotions” is therefore a design decision about the label taxonomy. It is not proof that subjective experience is restricted to six separate internal states. Changing the taxonomy can change what counts as a correct prediction.
The Measurement Problem: What Is the Ground Truth?
Scientists distinguish construct validity from agreement with a dataset's reference labels. The reference label can originate from self-report, behavioral observation, a prompted pose, external raters, a stimulus category, or a group consensus. Every method answers a different question and carries its own uncertainties.
Self-report and experience sampling
When the research target is consciously experienced feeling, a contemporaneous self-report is often more directly relevant than an outsider's interpretation. Yet self-reports depend on memory, vocabulary, willingness to disclose, cultural norms, and the question format. People can experience mixed or shifting feelings, and reports can differ across time.
Repeated sampling in natural settings can establish whether a model tracks a person's own ratings across contexts. A classifier that predicts an individual user's ratings may have utility even if it cannot infer a universally defined emotion category. Models should disclose this operational definition and should not promote a subjective report into an objective bodily fact.
Observer judgments
Independent observers can be asked “What emotion does this person seem to express?” Those ratings are suitable labels for a system designed to approximate social perception. The same ratings cannot, by themselves, confirm what the observed person felt. Agreement among ten observers measures consensus about a signal, not necessarily accuracy about private experience.
If the task is to predict human judgments of facial appearance, describe it as perceived or attributed emotion. If the task is to detect smiling, describe it as detection of a facial action. If the task is to infer a participant's reported feeling, name the timing and method of that report. Precise wording is a scientific necessity, not a cosmetic detail.
Uncertainty, calibration, and base rates
Models often output a vector of category scores. A score of 0.8 for “anger” can mean a relative model probability under training conditions. Whether it corresponds to 80% real-world correctness requires separate calibration testing. Some interfaces convert scores into apparently certain sentences, concealing model uncertainty and the ambiguity of the categories.
In a workplace where an actual target event is rare, false-positive alerts can outnumber true detections even when a model has respectable sensitivity and specificity. Evaluation must report the base rate of the target state, the consequences of errors, and comparison with appropriate simpler alternatives. A system that appears accurate in a balanced laboratory sample can be unhelpful in a very different operational setting.
Bias, Disability, and Unequal Performance
Emotion-recognition algorithms may perform differently across skin tones, ages, facial morphologies, languages, cultural contexts, and individuals with different expressive or motor patterns. Sources of disparity include unrepresentative training data, labeler expectations, different camera conditions, and a target definition shaped by dominant cultural conventions. A difference in error rate can have greater consequences for people already subject to unequal scrutiny.
It is important to distinguish demonstrated demographic disparities for a particular tested model from assumptions that all models fail in the same way. A responsible evaluation reports subgroup sample sizes, uncertainty, error types, and the precise training and deployment conditions. It includes accessibility and opt-out considerations rather than imposing a single expressive norm.
A person who averts eye contact, speaks with atypical prosody, or moves their face differently must not be classified automatically as disengaged, dishonest, unsafe, or mentally ill. These are inferences about intentions or pathology that cannot be justified from surface expression alone. The NIST AI Risk Management Framework emphasizes validity, reliability, safety, transparency, fairness, and privacy as context-dependent properties of AI systems.
Can AI Detect Lies or Read Hidden Emotions?
Automated facial analysis, including the detection of brief facial movements, does not provide a validated general-purpose lie detector. People can deceive while displaying ordinary expressions, experience fear while telling the truth, and show fleeting movements for multiple reasons. A microexpression may be an observable event, but assigning it a concealed mental state and then a judgment of dishonesty requires independent evidence.
For the same reason, a tool cannot reliably determine romantic attraction, moral character, workplace loyalty, or dangerousness from a facial snapshot. Such claims add unsupported inference layers beyond the already difficult task of predicting emotional expression. The 2019 review of emotional expressions explains why these shortcuts are scientifically fragile.
What About AI Systems That Claim to Understand Feelings?
A text chatbot may infer that a person is distressed from what they say and respond with supportive language. It can be useful as a conversational interface, but its response is based on learned patterns and the immediate context. It cannot independently verify the user's internal experience. A user can correct the system's interpretation, and an interface should leave room for that correction.
This article assesses models' ability to infer human emotional information. Whether an AI system itself has subjective emotional experience is a separate philosophical and scientific question and does not follow from classification accuracy, humanlike responses, or affective simulation. Clinical AI tools and AI companions also require separate assessment of their purpose, validation, and risks.
Applications: Where the Evidence Matters Most
Research and accessibility
With informed participation, validated tools may assist researchers in describing facial movements, vocal patterns, or changes in arousal at scale. A person-centered system may adapt an interface when users explicitly report frustration or request assistance. In accessibility contexts, a system can help communicate expressively, provided that it respects individual differences and user control.
Healthcare and psychological support
Some studies explore how multimodal behavior relates to symptoms or screening questionnaires. These are developing research directions, and their evidence cannot be transferred automatically to autonomous diagnosis or treatment. Before a tool affects care, investigators must establish what it measures, whether it adds predictive value beyond established methods, how it performs in relevant populations, and what happens after an error. The systematic review of clinical facial-AI research documents limitations that warrant caution.
Education and employment
Emotion monitoring in classrooms or workplaces can turn ambiguous behavioral signals into judgments about motivation, compliance, productivity, or engagement. The people observed may have little control over the camera, model, data retention, or interpretation. Errors can affect opportunities, trust, and autonomy, even if the software was originally developed for a less consequential setting.
In the European Union, Article 5(1)(f) of the AI Act prohibits specified AI systems used to infer emotions in workplaces and educational institutions, subject to the stated medical or safety exception. This is a jurisdiction-specific legal rule, not a universal ban on all emotion-related AI. Article 50 also includes transparency obligations for deployers of certain emotion-recognition systems. Actual legal duties depend on system definition, use, context, and applicable implementation dates.
Marketing, customer service, and surveillance
A company might summarize the emotional tone of opt-in feedback or examine aggregate reactions to content. These tasks may have operational value if the organization is transparent about the target, consent, and error. Inferring psychological states from covert footage or recording workers and customers without meaningful control raises more serious questions about privacy, power, and validity.
Deployers should avoid turning an uncertain model output into a definitive record about a person. A safer interface reports the observed cue and the system's uncertainty, restricts data collection, limits retention, allows review, and keeps consequential decisions under accountable human processes.
How to Evaluate a Company’s Emotion-AI Accuracy Claim
Begin by requesting the model's exact target definition and label provenance. Does “85% accuracy” refer to a posed-expression photo dataset, judges' ratings of acted voices, actual self-reported emotions, or predictions of a clinical condition? The answer changes what the number means.
Next examine whether evaluation used independently recruited individuals, a genuinely external test dataset, live conditions that resemble deployment, and relevant demographics. Request confusion matrices, precision and recall per class, uncertainty intervals, calibration, and error analysis for ambiguous or mixed emotions. Ask whether one person appears in both training and test data and whether samples contain duplicated video frames.
Finally ask what decision the output will influence, whether the system improves on simpler tools, what participants are told, how mistakes are corrected, and whether individuals can decline monitoring. Consistent with the NIST framework, technical performance is only one component of an accountable system. Claims of “reading thoughts” or “detecting true emotion” exceed what ordinary benchmark evidence demonstrates.
Practical Guidance for Users
Treat an emotion label from software as a tentative interpretation of selected data. A useful question is “Which cues caused the model to make this suggestion, and what other explanations fit?” When you know your own experience, your account should take priority over an algorithm's speculation about how you feel.
If an app tells you that you are angry or sad, check whether it used your words, voice, facial image, or wearable signal. Consider the circumstances and whether you agreed to the analysis. If the conclusion matters at school, work, or in healthcare, ask for the underlying evidence, the system's validation, and a meaningful human review process.
When you are concerned about emotional or mental health, a model's facial score should not replace a conversation with a qualified professional. Everyday anger, sadness, fear, and stress are normal human experiences; an unfavorable score from an automated classifier does not constitute a diagnosis.
Frequently Asked Questions
Can AI accurately recognize human emotions?
AI can classify labeled emotional cues, sometimes accurately in controlled benchmarks. Accuracy about a person's actual subjective experience is a separate, harder claim requiring appropriate ground truth and validation in real-world contexts.
Can AI tell how I feel by looking at my face?
It can estimate likely facial movements or labels associated with expressions in its training data. A face alone rarely provides enough evidence to identify one person's feeling reliably, because expressions vary and context changes their meaning.
Is AI better than humans at reading emotions?
Comparisons depend on the signals, labels, and participants. Some models approach or exceed specific human baselines on particular datasets; other models perform worse on different modalities. A 2026 experiment found contrasting results for vocal tone and bodily gestures.
Can AI recognize emotions from text?
It can classify emotion expressed or implied by words and conversation. But sarcasm, politeness, metaphor, incomplete context, and deliberate concealment can separate textual meaning from the writer's actual experience.
Can AI hear anger or sadness in a voice?
Models can detect acoustic patterns associated with perceived emotional tone. They need validation across languages, speakers, recording devices, and contexts. The same vocal features can have several explanations.
Does a 95% emotion-detection accuracy score prove the system works?
Only for the clearly defined dataset, label reference, and evaluation protocol used to obtain it. Independent real-world testing, class-specific errors, population coverage, and calibration are necessary to assess deployment usefulness.
Can a smartwatch know whether I am anxious or stressed?
Wearables measure physiological proxies that can be related to arousal. They usually cannot identify a specific emotion or clinical condition uniquely from heart rate or skin conductance.
Can AI detect lies from facial microexpressions?
There is no sound basis for treating fleeting expressions as a reliable universal lie detector. The link from facial movement to emotional meaning, and from emotional meaning to deception, requires evidence not supplied by the movement itself.
Is emotion recognition banned?
Rules vary by jurisdiction and purpose. The EU AI Act specifically restricts certain workplace and educational emotion-inference uses, subject to exceptions, and imposes transparency duties in defined cases. It does not create a blanket global prohibition.
Can AI diagnose depression or anxiety from expressions?
A model's emotion score does not diagnose any disorder. Research on AI-assisted assessment is developing, but reliable clinical use requires separate, prospective, population-appropriate validation and professional safeguards.
What is the difference between AI emotion recognition and AI empathy?
Emotion recognition is a computational classification or estimation task. A system may produce an empathetic-sounding response, but that does not demonstrate direct access to someone's feelings or establish anything about AI subjective experience.
Conclusion
The strongest evidence supports a bounded conclusion: AI can identify patterns in expressed, perceived, or reported affect under defined conditions. Facial movements, voice, language, and physiological responses provide relevant but incomplete information, and their relationships with subjective feelings vary across people and situations. Trustworthy emotion AI depends on valid labels, diverse and external testing, calibrated uncertainty, clear communication of its target, and responsible use. The practical question is always what exactly the system predicts, for whom, under which circumstances, and with what consequences.
For additional scientific context, see Emotion Recognition: Faces, Voices, and Context.
Related Articles
Emotion Recognition: Faces, Voices, and Context explains how people infer emotion through multiple social cues. Are Facial Expressions Universal? Cross-Cultural Evidence examines the cultural and scientific disputes. How Scientists Measure Emotions: Methods and Limitations details the measurement problem. What Are Emotions? Psychological Definition, Components, and Functions provides the broader emotion-science foundation.
References
Barrett, L. F., Adolphs, R., Marsella, S., Martinez, A. M., & Pollak, S. D. (2019). Emotional expressions reconsidered: Challenges to inferring emotion from human facial movements. Psychological Science in the Public Interest, 20(1), 1–68. https://doi.org/10.1177/1529100619832930
Brooks, J. A., Tzirakis, P., Baird, A., et al. (2023). Deep learning reveals what vocal bursts express in different cultures. Nature Human Behaviour, 7, 240–250. https://doi.org/10.1038/s41562-022-01489-2
Cowen, A. S., & Keltner, D. (2017). Self-report captures 27 distinct categories of emotion bridged by continuous gradients. Proceedings of the National Academy of Sciences, 114(38), E7900–E7909. https://doi.org/10.1073/pnas.1702247114
Cowen, A. S., Laukka, P., Elfenbein, H. A., Liu, R., & Keltner, D. (2019). The primacy of categories in the recognition of 12 emotions in speech prosody across two cultures. Nature Human Behaviour, 3, 369–382. https://doi.org/10.1038/s41562-019-0533-6
D'Amelio, T. A., et al. (2025). Emotion recognition systems with electrodermal activity: From affective science to affective computing. Neurocomputing, 651, 130831. https://doi.org/10.1016/j.neucom.2025.130831
Demszky, D., Movshovitz-Attias, D., Ko, J., Cowen, A., Nemade, G., & Ravi, S. (2020). GoEmotions: A dataset of fine-grained emotions. Proceedings of ACL 2020, 4040–4054. https://doi.org/10.18653/v1/2020.acl-main.372
European Parliament and Council. (2024). Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 5: Prohibited AI practices. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-5
European Parliament and Council. (2024). Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 50: Transparency obligations. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50
Ghafarfaraji, S., et al. (2026). AI-based recognition of facial and micro-expressions for the diagnosis of mental and neurological disorders: A systematic review. BMC Psychiatry, 26, article 78. https://doi.org/10.1186/s12888-025-07739-7
Giordano, B. L., Whiting, C., Kriegeskorte, N., Kotz, S. A., Gross, J., & Belin, P. (2021). The representational dynamics of perceived voice emotions evolve from categories to dimensions. Nature Human Behaviour, 5, 1203–1213. https://doi.org/10.1038/s41562-021-01073-0
Hsu, C.-F., Konara Mudiyanselage, S. P., Agustina, R., & Lin, M.-F. (2025). Basic emotion detection accuracy using artificial intelligence approaches in facial emotions recognition system: A systematic review. Applied Soft Computing, 172, 112867. https://doi.org/10.1016/j.asoc.2025.112867
National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. https://doi.org/10.6028/NIST.AI.100-1
Picard, R. W. (1997). Affective Computing. MIT Press. https://doi.org/10.7551/mitpress/1140.001.0001
Piterman, D., Dery, H., Bar, K., et al. (2026). Generative AI emotion recognition from bodily gestures and vocal tone reveals modality-specific performance and positivity bias. Scientific Reports, 16, article 27663. https://doi.org/10.1038/s41598-026-58013-z
Wu, C., Cai, Y., Liu, Y., Zhu, P., Xue, Y., Gong, Z., Hirschberg, J., & Ma, B. (2025). Multimodal emotion recognition in conversations: A survey of methods, trends, challenges and prospects. Findings of EMNLP 2025. https://aclanthology.org/2025.findings-emnlp.332/