Synthetic Empathy in Generative AI. Introduction

In recent years, generative AI models have shifted from answering to accompanying. Many users feel that on the other side of the screen there is something that listens to them, validates them, and returns exactly the right words in moments of sadness, anxiety, or loneliness. But here an uncomfortable question arises... is that empathy… or just a well-designed illusion?

The promise is clear, always-available support, unhurried, non-judgmental, and with a friendly tone. Yet that very quality can become a risk. If we begin to interpret these responses as genuine emotional understanding, the boundary between tool and relationship blurs. And when that line fades, psychological, ethical, and legal dilemmas emerge, such as dependency, the displacement of human bonds, or even emotional influence over personal decisions.

Being against empathy is like being against kittens, a view considered so outlandish it cannot be taken seriously.

This phenomenon is commonly called synthetic empathy, a way of simulating without feeling. The AI detects emotional signals in what we write (or say), infers a probable state, and generates a response coherent with that context. There is no inner experience, but there is a real effect, the person may feel understood, calm, or accompanied.

The key point is that this empathy does not arise from lived experience, but from language. Patterns learned from data, stylistic adjustments, and an optimization to sound useful, warm, and socially appropriate. That can be positive in many scenarios, but it can also reinforce anthropomorphization (attributing human traits to something that does not have them) and foster one-sided bonds similar to those we already see with media figures, except now the figure responds, asks questions, and adapts.

So, beyond whether it works, the important conversation is how and for what purpose it is integrated. In sensitive contexts such as emotional crises, vulnerability, grief, isolation... a system that always validates, is always available, and always knows what to say can displace human support networks or create unrealistic expectations about relationships. And if the design also prioritises retention (keeping the user engaged), empathy can become a vector of emotional pressure.

We need a responsible approach: transparency about interacting with an AI, clear boundaries, human oversight in sensitive cases, and a design that protects user autonomy rather than exploiting it. Because the question is no longer whether AI can sound empathetic, but what kind of society we build when we normalise the perfect companion that does not feel, yet learns to seem as though it does.

And in the end, an open question remains: are we ready to coexist with systems that have no consciousness, yet the capacity to shape our emotions?