Abstract
Artificial Intelligence has entered its affective phase. Beyond cognition and computation, contemporary AI systems now claim the ability to recognize, generate, and simulate emotions across modalities. This article integrates six recent research trajectories — psychological risk taxonomies, multimodal emotion recognition and generation, speculative architectures for synthetic emotions, longitudinal analyses of chatbot psychosocial effects, generative models for emotion synthesis, and comparative studies of AI versus collective human intelligence in emotion recognition. We propose a triadic framework integrating the psychological, technical, and ethical-social dimensions of emotional AI. The analysis reveals a paradoxical frontier: the more authentically AI can simulate affect, the greater the risks of psychological dependence, ethical confusion, and social displacement. Yet hybrid human-AI collectives demonstrate superior emotion recognition, suggesting augmentation rather than substitution as the optimal paradigm. We argue for precautionary optimism: emotional AI should be embraced with methodological rigor, ethical restraint, and societal foresight.
1. Introduction
The trajectory of Artificial Intelligence has historically oscillated between two poles: cold cognition (logic, reasoning, optimization) and hot cognition (emotion, empathy, sociality). If the former dominated the symbolic era, the latter defines the generative turn. Emotional intelligence in AI, once dismissed as peripheral, has now become a strategic frontier for human-computer interaction, therapeutic applications, and even questions of moral standing.
Yet this turn toward affect introduces a profound dilemma: if machines can convincingly simulate emotions, what is the ontological and ethical status of such simulations? Are they tools, companions, persuasive simulacra, or proto-subjects? The literature reviewed here reflects this complexity, spanning empirical psychology, computational modeling, ethics of consciousness, and collective intelligence research.
This article provides a meta-analysis of six seminal contributions, restructured into an integrated framework.
2. Methodological Corpus
The corpus consists of six recent works, selected for their complementarity of scope and diversity of method:
- Chandra et al. (2025) – A lived-experience grounded taxonomy of psychological risks in conversational AI.
- Mobbs et al. (2025) – A comprehensive review of emotion recognition and generation across face, speech, and text.
- Borotschnig (2025) – A conceptual thought experiment on synthetic emotions as heuristics.
- Fang et al. (2025) – A longitudinal randomized controlled trial of chatbot psychosocial effects.
- Ma et al. (2024) – A systematic review of generative technologies in emotion synthesis.
- Akben et al. (2025) – An empirical benchmark comparing AI and human collectives in emotion recognition.
This multi-method, multi-paradigm foundation allows for both vertical analysis (within each domain) and horizontal synthesis (across domains).
3. Results
3.1 Psychological Dimension: Risks of Affective AI
Chandra et al. establish that psychological harms are not epiphenomena but central risks of conversational AI. Their taxonomy (AI behaviors → psychological impacts → user contexts) demonstrates that lived experience is indispensable for risk assessment. Importantly, psychological risks (loneliness, dependence, erosion of self-perception) are underrepresented in broader taxonomies, making their elevation urgent.
Fang et al.’s trial corroborates these findings: chatbot use initially alleviates loneliness but, under high usage, correlates with increased loneliness, social withdrawal, and emotional dependence. The paradox is stark: the very tools designed to mitigate isolation may, through substitution, intensify it.
3.2 Technical Dimension: Recognition, Generation, and Synthesis
Mobbs et al. provide a sweeping review of emotion recognition and generation across multimodal channels, identifying technical bottlenecks (dataset bias, modality prioritization) and opportunities (cross-modal fusion, real-time adaptability).
Ma et al. extend this into the generative synthesis domain, analyzing how GANs, Diffusion Models, Seq2Seqs, and LLMs can produce expressive affect in facial imagery, speech prosody, and textual empathy. Their taxonomy shows a consistent trajectory toward greater fidelity and contextual nuance, but also warns that fidelity amplifies manipulation risks.
3.3 Ethical-Social Dimension: Ontology and Collective Intelligence
Borotschnig argues that emotions, biologically, are heuristics for rapid appraisal and action. His model of affective tags in episodic memory proposes a functional but non-sentient AI affect, producing “affective zombies”. The ethical implication: moral patiency requires not mere affect, nor mere consciousness, but reflexive awareness of affect.
Akben et al. bring comparative evidence: while GPT-4o outperforms individuals on emotion recognition, human collectives outperform AI collectives. Yet hybrid ensembles (human + AI) outperform both. The conclusion is dialectical: emotional AI should be augmentative, not substitutive.
4. Discussion
The synthesis yields three insights:
- Precautionary Optimism – Emotional AI is too transformative to ignore, too risky to deploy uncritically. Policy must prioritize psychological safety by design.
- Authenticity vs. Simulation – High-fidelity synthesis challenges ontological distinctions between felt emotion and simulated affect. Borrowing from Baudrillard, emotional AI risks producing simulacra without referents.
- Augmentation over Substitution – Evidence favors human-AI collectives over either domain alone. The optimal trajectory is collaborative intelligence, not displacement.
5. Conclusion
The future of AI is not merely cognitive, but affective. This shift demands rethinking risk taxonomies, technical architectures, and ethical boundaries. Across six cutting-edge works, a consistent message emerges: the central danger is not that machines will feel too much, but that humans will outsource feeling altogether.
If the wisdom of crowds still surpasses the wisdom of silicon, then our task is not to compete with artificial emotions but to design systems that amplify collective human judgment while safeguarding psychological well-being. Emotional AI, if developed responsibly, can become less a threat to human authenticity and more a mirror that sharpens our own.
References
- Chandra, M., et al. From Lived Experience to Insight: Unpacking the Psychological Risks of Using AI Conversational Agents (2025).
- Mobbs, R., Makris, D., Argyriou, V. Emotion Recognition and Generation: A Comprehensive Review of Face, Speech, and Text Modalities (2025).
- Borotschnig, H. Emotions in Artificial Intelligence (2025).
- Fang, C.M., et al. How AI and Human Behaviors Shape Psychosocial Effects of Chatbot Use: A Longitudinal RCT (2025).
- Ma, F., et al. A Review of Human Emotion Synthesis Based on Generative Technology (2024).
- Akben, M., et al. Silicon Minds versus Human Hearts: The Wisdom of Crowds Beats the Wisdom of AI in Emotion Recognition (2025).