Affective computing is the branch of artificial intelligence concerned with detecting, interpreting and responding to human emotions. Existing systems can estimate emotional states from facial expressions, speech, text, body movement, heart rate or other physiological signals. Yet the researchers argue that many such systems still approach emotion as an isolated classification problem: a machine receives a signal and labels it as happiness, anger, fear, sadness or another category. Human emotion, by contrast, is fluid and context-dependent. The same silence may signal concentration in one setting, discomfort in another or respect in a third. A smile may express joy, politeness, nervousness or social pressure. The central claim of the review is that future systems must model the transformations linking signals, situations, cultures, memories and reasoning, rather than merely recognize patterns in individual data streams.

The CPST framework divides the problem into four interacting dimensions. Physical space includes measurable bodily conditions such as pulse, skin conductance, facial muscle activity, breathing and movement. Social space encompasses relationships, cultural norms, group behavior and the expectations attached to particular situations. Thinking space refers to cognitive processes, including interpretation, attention, memory, intention and judgment. Cyberspace contains digital traces such as written messages, online interactions, sensor records and other computational representations. These categories are not independent compartments. A stressful conversation in cyberspace can produce a physical response; that response can influence cognition; cognition can modify how the conversation is remembered; and social norms can determine whether the person displays or conceals the emotion. In the proposed view, affective intelligence emerges from tracking those cross-space transformations.

The role of artificial general intelligence, or AGI, in this proposal is not to provide a magical “emotion detector.” Instead, the authors describe AGI capabilities as computational engines that could connect information across the four spaces. Perception would allow a system to gather signals from text, images, audio, wearables and environments. Learning would help it identify patterns in an individual or population while adapting to new evidence. Reasoning would support interpretation of ambiguous situations, and memory would preserve relevant personal and social context. Generalization would allow knowledge learned in one circumstance to be applied cautiously in another, while generation would enable the system to produce explanations, dialogue, recommendations or other responses. The combination is intended to move machines from surface-level recognition toward contextual affective understanding.

That shift is technically demanding because emotion is only partly visible in data. Physiological measurements can indicate arousal, but arousal alone does not reveal whether a person is excited, frightened or physically exerting themselves. Speech analysis can detect changes in pitch, speed and intensity, but those features vary across languages, cultures and individuals. Text can reveal sentiment or emotional vocabulary, yet people often use irony, understatement, euphemism or deliberate concealment. A system operating within the CPST framework would therefore need to combine heterogeneous data and estimate relationships among them. It might compare a person’s words with vocal features, physiological signals, recent events and known social circumstances, while representing uncertainty instead of presenting a confident but unsupported label.

The review maps several AI techniques onto these affective tasks. Causal modeling is important because correlation alone cannot explain why an emotional state emerged. A system may observe that rapid speech often accompanies anxiety, but a causal model would ask whether anxiety caused the speech pattern, whether the pattern resulted from another factor or whether both arose from a third circumstance. Such models could help distinguish triggers, consequences and coincidental signals. The authors also discuss chain-of-thought reasoning as a way for an AI system to organize multi-step interpretations, linking observations to possible explanations and responses. In practice, a reliable system would need safeguards around such reasoning, because an internally coherent explanation can still be wrong if its data or assumptions are biased.

The framework also identifies generative models, including diffusion models, as tools for affective computing. Diffusion models are generative systems that learn to produce complex outputs by gradually transforming structured noise into images, audio, text or other data. Within an emotion-aware architecture, such models could help simulate likely emotional scenarios, generate personalized interactions or reconstruct missing information from incomplete multimodal signals. Their value would not lie simply in producing convincing content, but in supporting interaction across spaces—for example, translating a detected physiological state into an adaptive interface or generating a response that reflects social and cognitive context. The same generative power, however, creates risks: a system that can fabricate emotionally persuasive content could manipulate users, amplify stereotypes or imitate empathy without possessing any genuine understanding.

The authors position their framework as a roadmap for applications rather than a report of a completed AGI system or a clinical validation study. Potential uses include intelligent assistants that respond more appropriately to users’ emotional circumstances, human-centered robots that adapt their behavior to social settings, educational technologies that recognize cognitive and emotional difficulty, and healthcare tools that integrate physiological and conversational information. In cyber-physical environments—where people, sensors, machines and digital networks operate together—the CPST model could help systems react to changing human needs instead of treating every user as an abstract data point. The same principles could apply to transportation, workplaces, immersive virtual environments and other settings in which emotional states influence safety, cooperation and decision-making.

The most provocative promise is the possibility of moving toward machine behavior that resembles empathy. The paper defines this ambition not as proving that an AI feels emotion, but as enabling it to recognize another person’s condition, interpret that condition within context and respond in a manner appropriate to the individual and situation. Genuine empathy-like performance would require more than accurate classification. It would require sensitivity to cultural differences, personal history, conflicting signals, uncertainty and the possibility that a person does not want to disclose how they feel. A machine that responds to every sign of sadness with the same scripted reassurance could be emotionally attentive in a narrow technical sense yet socially inept. Contextual awareness, the authors argue, is therefore the dividing line between an affective system that detects and one that understands.

That vision also exposes serious scientific and ethical problems. Emotional data can be intimate, persistent and difficult to collect with meaningful consent. Physiological signals may reveal vulnerability, while digital records can expose relationships, habits and psychological states. Systems trained on narrow populations may misread expressions, speech or social conventions in people from other cultures or demographic groups. An AGI that infers hidden feelings could be used to target advertising, evaluate workers, influence voters or make consequential decisions without a person’s knowledge. The review’s framework does not eliminate these dangers, but it makes them harder to ignore by placing emotion inside a network of bodies, societies, minds and machines. Before such systems are deployed, they will need transparent uncertainty estimates, strong privacy protections, mechanisms for human oversight and evidence that their interpretations are valid across real-world contexts. The proposed CPST perspective therefore represents both an ambitious blueprint for next-generation affective computing and a warning that emotional intelligence in machines will be measured not by how persuasively they imitate empathy, but by how responsibly they handle the complexity of human feeling.