AI Empathy Evolution: From Polite Responses to Relationship-Based Intelligence

When the AI Empathy Research Initiative was launched in 2024, the dominant question was still whether artificial intelligence could display empathy at all. Many scholars and executives were understandably skeptical. Could an AI assistant such as ChatGPT or Claude respond empathically in a way that enhanced consumer experience and satisfaction without sounding manipulative, scripted, or inauthentic?

Only two years later, the debate has shifted. The most relevant question is no longer simply whether AI can display empathy. It is how AI agents should behave empathically, in which contexts empathy creates value, and where empathic behavior may become unnecessary, counterproductive, or even harmful.

This shift matters for executives because empathic AI is moving from research laboratories into customer service, healthcare, financial services, education, retail, and companionship. Organizations are beginning to ask not only whether AI can automate tasks, but whether AI can manage emotions, reduce frustration and support better decisions. This requires a more precise understanding of what AI empathy is and what it is not.

AI empathy can be defined as the capacity of AI agents to recognize and adapt to the cognitive needs and emotional states of a human interlocutor. This does not mean that AI “feels” compassion in the human sense. Without subjective experience, AI cannot truly share another person’s emotional state. Affective empathy presumes a capacity for experience-sharing that current AI systems do not possess. But AI systems can increasingly detect emotional cues, interpret conversational context, and generate responses that users perceive as understanding, supportive, and socially appropriate.

In other words, AI empathy is not human empathy transferred into a machine. It is a designed interaction capability. The managerial question is whether this capability can improve human experience, and under what conditions.

Research over the past two years has produced a more nuanced picture. One important finding is what Wenger et al. (2026) describe as a preference–perception gap in AI empathy. People may rate AI-generated empathic responses as highly empathic, yet still prefer to receive empathy from humans. This finding is important because empathy is not only evaluated as message quality. It also signals relationship potential, care, connection, and reciprocity. A beautifully phrased response may make someone feel heard in the moment, but human empathy carries additional social meaning.

This idea is further developed by Perry (2026), who explains why people often discount AI empathy. Genuine human empathy indicates costly commitment and likely future support. By contrast, an AI-crafted message may feel comforting in the moment, but it does not necessarily guarantee future care. As a result, users may down-weight AI empathy as a signal of closeness. Human empathy can function as a signal of future care and commitment, while AI empathy may be perceived as a “cheap signal” of automated care. This directly points to the importance of a relationship-based approach to AI empathy.

A second lesson concerns the risk of sycophancy. Rehani et al. (2026) study whether AI systems such as Claude are genuinely empathic or simply sycophantic. The key insight is that perceived empathy is strongly correlated with sycophancy. Making AI more empathic may also make it more likely to agree with users and avoid pushing back. In other words, what feels like understanding may also reduce the AI’s ability to challenge users.

This is highly relevant for executives. In leadership development, coaching, financial advice, healthcare, or customer service, a useful AI agent should not merely validate the user. It should know when to support, when to clarify, and when to push back.

A third lesson comes from healthcare-related studies. Howcroft et al. (2025) report that thirteen out of fifteen studies found statistically significantly higher empathy ratings for AI chatbots than for human professionals. However, the authors also highlight important limitations. The reviewed studies are based on text-only interactions and often rely on proxy raters rather than real patients experiencing ongoing care. This means that non-verbal cues, tone of voice, and long-term relational context are largely absent.

This is a crucial point. Much of what we call empathy is not only semantic. It is carried through timing, tone, hesitation, memory, and continuity. A text response may look empathic on a screen, but real empathic interaction often depends on voice modality and relationship history.

This is where the next generation of AI empathy research is moving.

At the AI Empathy Lab, our research focuses on how AI agents can behave in ways that users perceive as emotionally appropriate, trustworthy, and supportive.

One stream of our work examines AI empathy in service failure contexts. AI voice assistants are increasingly used in high-stakes domains such as banking, where interaction failures can quickly damage user experience. Our findings show that AI empathy has conditional value: it matters less in smooth interactions, when users mainly want the task completed, but becomes critical when something goes wrong (Efthymiou et al., 2025).

In these failure contexts, AI empathy functions as an emotional regulator. It helps reduce frustration and anger, stabilizes the interaction in real time, increases user satisfaction, and lowers verbal aggression. These effects are strongest during failures, suggesting that empathy primarily acts as a buffer against negative reactions rather than as a universal enhancer of all AI interactions.

A second stream of our research studies empathic AI in sustainable consumer decision-making. We explored how empathic and persuasive AI voice assistants can steer users toward more sustainable choices. In this research, participants ordered food through a custom voice assistant powered by a LLMs, choosing between a plant-based and a meat-based burger and between bicycle and car delivery.

The findings suggest that empathic AI assistants can encourage sustainable consumer choices. However, trust in AI competence plays a complex role (Mari et al., 2025). Trust can mediate behavior change, but the research also identifies a paradox: without explicit moral direction or sustainability-framed persuasion, greater trust in AI may cement habitual choices instead of changing them. In other words, if a trusted AI assistant presents an option without a clear sustainability frame, users may infer that their usual choice is acceptable. Trust alone does not necessarily produce better decisions. It must be combined with appropriate persuasion.

A third stream examines AI empathy in voice-based shopping. Across experiments, we show that AI empathy in voice assistants increases consumer decision satisfaction (Mari et al., 2026). Empathy works directly by improving the emotional experience of the interaction, and indirectly by enhancing perceived transparency and reducing perceived manipulative intent.

Importantly, these effects depend on product type. For hedonic products, such as scented candles, empathy has a stronger direct emotional impact. For utilitarian products, such as batteries, empathy works mainly through cognitive evaluations such as transparency and trust. This suggests that the future of empathic AI will not be one-size-fits-all. Effective agents will calibrate their emotional behavior depending on the task, product category and user emotional state.

Most current AI empathy is momentary. The user says something, the AI detects an emotional cue, and the system responds with a supportive phrase. This can be useful, but it is also limited. Users may think: “The AI is saying this because it is programmed to sound empathic.” This creates skepticism, manipulation inferences, and sometimes backlash against “fake empathy.”

Memory may change this dynamic. Empathy without memory can feel like polite performance. Memory without empathy can feel like cold personalization. But empathy combined with memory may become relationship-based intelligence.

Consider the difference between an AI saying, “I’m sorry you are having a difficult morning,” and an AI companion saying, “You mentioned last time that mornings are often difficult for you. How are you feeling today?” The second response is not only emotionally appropriate in the moment. It shows recognition across time.

This is why our current research studies how an empathic AI companion’s use of memory affects consumer experience. The central idea is that empathy becomes more credible when it is situated in a shared interaction history. Memory can legitimize empathic responses because it gives the AI a reason to respond in a particular way. The response no longer appears generic. It appears grounded.

This also creates new ethical and managerial questions. How much should an AI remember? Which memories are useful, and which are intrusive? When does continuity build trust, and when does it create discomfort? How should organizations explain memory use transparently? These questions are especially important in applications involving older adults, healthcare and financial advice. 

AI empathy has moved from a philosophical question to a managerial one. The question is no longer whether machines can truly feel. The question is whether organizations can design AI agents that respond to human emotion in ways that are useful, credible, ethical, and contextually appropriate. That is the work ahead.

For executives, the practical implication is clear: empathic AI should not be understood as a “nice tone of voice” layer added at the end of system design. It is a strategic design capability that requires decisions about context, memory and boundaries. As the technological possibilities of AI become almost limitless, the decisive capability shifts back to managers: domain expertise. Leaders need to understand the logic of the customer, patient, employee, or learner experience well enough to decide where empathy creates value, where it may be unnecessary, and where it may even be harmful.

During our executive education modules, we therefore work on emotion-based AI interventions. Participants design and test AI interactions that adapt to a user’s emotional state. The objective is to understand how context-adaptive empathy can influence user emotions, behavior, or decisions. This type of exercise helps leaders move beyond abstract debates and confront the real design trade-offs: when should the AI comfort, when should it challenge, when should it explain, and when should it remain neutral? In this sense, the future of empathic.

AI will not be shaped by technology alone, but by managers who know how to design meaningful experiences around it.

Text: Alex Mari