Artificial Emotion: A Survey of Theories and Debates on Realising Emotion in Artificial Intelligence PubID: pubid: 0000–0000/00$00.00 © xxxx IEEE
Yupei Li, Qiyang Sun, Michelle Schlicher, Yee Wen Lim, Björn W. Schuller · 2025
Read Artificial Emotion: A Survey of Theories and Debates on Realising Emotion in Artificial Intelligence PubID: pubid: 0000–0000/00$00.00 © xxxx IEEE by Yupei Li, Qiyang Sun, Michelle Schlicher, Yee Wen Lim, Björn W. Schuller (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-10286.txt.
Introduction
Whether AI should hold inner emotion-like states is surveyed by Yupei Li and co-authors in a 2025 arXiv paper. Machines already read and fake human feelings well. The team asks if giving them felt states of their own would help, as feelings help people. They call that step Artificial Emotion and sort the claims around it.
Early signs of such states in learning systems get a pass. So do builds that let feeling steer the works, and ways to wire feelings in. Gains could span smarts, drive, and social fit. Risks and ethics get a full look before the team's own take.
Catalog record
| CREATOR | Yupei Li, Qiyang Sun, Michelle Schlicher, Yee Wen Lim, Björn W. Schuller |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 39 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-10286
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Artificial Emotion: A Survey of Theories and Debates on Realising Emotion in Artificial Intelligence PubID: pubid: 0000–0000/00$00.00 © xxxx IEEE Yupei Li, Qiyang Sun, Michelle Schlicher, Yee Wen Lim, Björn W. Schuller Abstract Affective Computing (AC) has enabled Artificial Intelligence (AI) systems to recognise, interpret, and respond to human emotions—a capability also known as Artificial Emotional Intelligence (AEI). It is increasingly seen as an important component of Artificial General Intelligence (AGI). We discuss whether in order to peruse this goal, AI benefits from moving beyond emotion recognition and synthesis to develop internal emotion-like states, which we term as Artificial Emotion (AE). This shift potentially allows AI to benefit from the paradigm of ‘inner emotions’ in ways we – as humans – do. Although recent research shows early signs that AI systems may exhibit AE-like behaviours, a clear framework for how emotions can be realised in AI remains underexplored. In this paper, we discuss potential advantages of AE in AI, review current manifestations of AE in machine learning systems, examine emotion-modulated architectures, and summarise mechanisms for modelling and integrating AE into future AI. We also explore the ethical implications and safety risks associated with ‘emotional’ AGI, while concluding with our opinion on how AE could be beneficial in the future. I Introduction There is a growing scholarly interest in Artificial General Intelligence (AGI), particularly as Large Models (LMs) have demonstrated remarkable performance across a broad spectrum of general tasks mumuni2025large . Notably, the capacity of AGI systems to recognise and understand emotions, often referred to as Artificial Emotional Intelligence (AEI), has been extensively investigated within the field of Affective Computing (AC) schuller2021review ; schullerspeech . However, remaining on this perspective is insufficient for advancing AGI beyond its present functional capabilities ng2020strong . A deeper and more intrinsic understanding of emotion in AI in terms of Artificial Emotion (AE) is required to support the development of AGI and next generation oncoming AI systems that realise human-like affective processes. Despite extensive discourse on the concept of AE in the literature, which is sometimes referred to as synthetic emotion murray1993toward ; schroder2001emotional ; gunes2011emotion , a clear and unified definition has yet to emerge. For instance, dubois2010natural examines the neurobiological mechanisms underlying human emotional processing to derive insights that could inform the design of emotion-capable AI, whereas zhou2021emotional introduces the notion of emotional thinking patterns, defined as the ability to integrate emotional responses into information processing and decision-making, which is an essential component of human-like cognition. Building on these perspectives, we define AI emotion, or AE, as a technical integration of emotion states into the internal representational and decision-making processes of AI. Note that this does not directly imply that the AI ‘feels’ the emotions in a human-like manner, but rather alludes to the exploitation of the concept of emotions as biological insipiration similar to many other concepts in AI such as artificial neural networks take inspiration from the mammalian brain. This leads to the question why emotion could be important for AI and how we could realise emotion within AI. commonsense2023role This perspective suggests that emotions are far more complex than simple, momentary feelings; rather, they are deeply intertwined with the richness of human experience. Consequently, to genuinely comprehend and replicate human emotional life, an artificial intelligence system must be exposed to human emotional contexts and narratives. Such exposure is essential if AI is to engage in authentically meaningful interactions with humans. Similarly, hildt2019artificial explores the social implications of AEI, emphasising the importance of equipping AI with emotional capacities in order to fulfill social roles. They note that current developments in AE primarily focus on forming unidirectional emotional bonds, projecting lifelike qualities, and attributing human-like characteristics to machines. However, the other directional emotional bonds should be AI having emotions to effectively communicate with humans. AE has established both conceptual and technological foundations across diverse fields within learning systems. Emotion can be modelled either as a conditional factor for prediction aslam2022channels or as a primary target for inference saxena2020emotion . Furthermore, architectures such as Titans behrouz2024titans , which memorise learning materials during the testing phase, provide further evidence that machines can retain information and emulate aspects of human affective processing. Such capabilities have been manifested in numerous existing techniques, reflecting the growing convergence between memory-augmented learning and AE. Fig. 1: The structure of this review: Initially, we examine an exploration of the technical aspects, focusing on emotions in deep learning and architectures influenced by emotion. Following that, we review the social discourse surrounding this topic. However, there is considerable debate surrounding the concept of AE for instance whether AI could genuinely have the capacity to understand, process, and even experience human emotions sun2024towards , with both strong supporting and opposing viewpoints. On the one hand, proponents argue that granting AI having emotional capacities is a necessary step toward achieving AGI with greater capabilities gros2011emotionalcontrolconditio . On the other hand, critics express concern that emotionally capable AGI may pose existential risks to humanity, especially if such systems evolve rapidly alongside emerging forms of machine consciousness katirai2024ethical ; stark2021ethics ; lagrandeur2015emotion . This ongoing debate highlights the need for a thorough examination, which we aim to provide in this work. In conclusion, we recognise that the development of AE is ongoing and continues to provoke ethical debate. We complement the existing work from scheutz201412 , who previously explored the emergence of emotional capabilities in AI, and argues that artificial emotions could enhance an agent’s intelligence by improving its understanding of humans, as we review how to technically realise emotion in AI. We begin with an examination of learning approaches to modelling AE in Section II , then turn to focusing on emotion-embodied architectures in Section III . Finally, in Section IV , we review key ethical considerations, and offer our perspective on the necessity of incorporating emotional capabilities into AI systems. The topology of this review is shown in Figure 1 . II Emotion in Learning Approaches Modern learning systems approach AE through two primary paradigms: as conditioning variables that shape behaviour and as goals that guide learning. This section explores how each approach captures distinct aspects of emotion while also confronting fundamental limitations in approximating genuine affect. We begin by examining Reinforcement Learning (RL), a key domain where both paradigms are prominently applied. II-A Emotion as Reinforcement Signal RL offers one of the clearest pathways for artificial systems to develop emotion-like behaviours through direct interaction with consequences where AI systems should ‘feel’ emotions, rather than through superficial mimicry where AI system could just understand which emotion to mimic. The critical distinction lies in how reward supervisions are integrated into the system’s learning processes, shaping behaviour in a manner analogous to emotional conditioning moerland2018emotion . Pure RL usually uses emotion from environmental stakes. When agents learn through direct environmental feedback, emotion-like behaviours emerge organically. DeepSeek-R1-Zero guo2025deepseek discovered novel reasoning strategies through massive-scale RL without supervised warm-up, exhibiting “eureka moments” that gives surprising ideas when reasoning which its developers had not anticipated. This unexpected delight evokes the thought that, during reasoning, it may appear to feel a sense of “satisfaction” or “excitement” about its own line of thought, as if it possessed certain emotions and inspiration. Embodied