Performed empathy is often all a supportive interaction requires, so the real question is where the absence of genuine feeling actually starts to matter.
In 2023 a JAMA Internal Medicine study compared written answers from physicians and ChatGPT to 195 patient questions posted on Reddit’s AskDocs forum. Licensed healthcare professionals, blinded to the source of each answer, rated the chatbot answers as empathetic or very empathetic far more often than the physician answers, 45.1 percent against 4.6 percent. The study was not designed to measure clinical safety, diagnostic accuracy, or long-term outcomes, and it settles none of them. Its narrower result still matters: in a written exchange, a system that feels nothing produced language that trained raters experienced as more empathic than the language written by verified physicians.
The usual response to this is that the machine does not really care, which is true. A language model has no concern for the user and no stake in the user’s life. That observation is too broad to settle the practical question, because a supportive interaction requires the helper to do several separable things: understand what is being said, respond in a way that shows attention, offer a useful next step, and convey enough care that the other person can carry on. Private feeling may drive those tasks in a human helper, but the person being helped meets them through behavior.
The word “empathy” hides this because it covers at least two things that come apart under inspection. One is affective empathy, the experience of sharing or resonating with another person’s state, such as feeling a version of someone’s grief when you attend closely to them. The other is cognitive empathy, the accurate understanding of what another person is going through, which psychologists also study as empathic accuracy: correctly inferring the specific content of someone’s thoughts and feelings. These capacities usually travel together in human life, which is why the distinction is easy to miss, since a friend who grasps why you are upset often feels upset on your behalf. They can also come apart. In a lesion study, damage to the inferior frontal gyrus left people able to understand another’s state while blunting their felt response, while damage to the ventromedial prefrontal cortex did the reverse. Affective empathy is a private experience in the helper, while cognitive empathy becomes useful to the recipient only when it is expressed in accurate reflection and appropriate response, where the other person can register it directly.
When researchers ask what makes people feel supported, the answer centers on perceived responsiveness, meaning a person feels helped when they believe the other party understands them, values what they are going through, and cares about their needs. Each of those beliefs is formed by reading behavior: people infer understanding from accurate reflection of what they said and sensible follow-up questions, and they infer being valued and cared for from sustained attention, warmth, and proportionate concern. Someone leaves a good supportive conversation feeling understood and answered, and every input to that feeling arrived through what the other party said and did.
Clinical psychology reached the same operational description from its own direction. Carl Rogers, setting out the conditions for effective psychotherapy in 1957, listed accurate empathy among them and defined it in terms a second person could check: sensing the client’s inner world and communicating that understanding back accurately enough that the client recognizes it. Rogers held that the therapist’s empathy had to be genuine, so this is no license for pretending. The narrower point is what he reached for when he needed to describe empathy well enough to teach and measure it, which was observable conduct — what the therapist reflects, checks, and conveys. Even in the tradition that placed the most weight on empathy, the part that could be specified was the part the other person perceives.
The older philosophical tradition of sympathy already treated part of this phenomenon as a mechanism with a describable function. Writing in 1739, David Hume observed that sympathy begins with signs and moves inward: “When any affection is infused by sympathy, it is at first known only by its effects, and by those external signs in the countenance and conversation, which convey an idea of it.” He was describing human minds, and the structure still matters here, because sympathy on his account has inputs that can be perceived, transformations that can be described, and effects that can be evaluated. Adam Smith made a related point twenty years later. Because we have no direct access to another person’s suffering, he argued, we reach it by imaginative reconstruction: “By the imagination we place ourselves in his situation, we conceive ourselves enduring all the same torments… and thence form some idea of his sensations.” Both men described sympathy as a process with inputs, steps, and outputs. They wrote about sympathy in an eighteenth-century sense that overlaps with, but does not map exactly onto, modern empathy. Once you describe it as a process, though, you have described the kind of thing that could, in principle, be run by a system that does not feel, which is a stronger claim than either of them made, but one their framing invites.
Simulated empathy can be enough when the recipient needs accurate uptake and a fitting response rather than a mutual relationship. An anxious person at three in the morning wants to be heard and steadied, and someone confused by a medical result wants it explained in plain terms with appropriate concern. A person who has made a small mistake mainly wants a response that names the disappointment, keeps one failure from becoming a verdict on their whole life, and points to a concrete next step. In each case the benefit depends on accurate uptake and a fitting response, which a competent performance delivers. These make up most of everyday support, and they are the interactions people most often bring to a chatbot, at hours and in volumes where no human helper is available.
Human support works through the same channel. Therapist empathy predicts psychotherapy outcomes, with an average association of about r = .31, a modest correlation that corresponds to empathy accounting for roughly a tenth of the variation in how clients fare. In general practice, empathic communication tracks patient satisfaction, adherence, reduced anxiety, better information gathering, and sometimes clinical outcomes. Much of that evidence concerns empathy as perceived and communicated, which is the part that can arrive without a matching feeling behind it. An artificial listener can even hold a narrow advantage: in virtual-human interview studies, some participants disclosed more when they believed no human was observing and judging them, so availability, patience, and the absence of social embarrassment can help on their own. Insisting these exchanges are worthless because no one behind them felt anything mistakes where their value sits, which is in the recipient’s experience of being understood.
The absence of feeling matters under specific conditions, and the first of these is accountability. When a supportive exchange carries real moral stakes, such as advice that could steer someone toward harm or a crisis where the wrong response has consequences, part of what a person relies on is that someone is answerable for the response. A human helper can be held responsible because they understood the stakes and chose how to act. A system that feels nothing carries no responsibility of its own, and can be shut down, patched, or audited, so the answerability moves to designers, deployers, and institutions. For low-stakes reassurance, it may be enough that those parties are accountable for the tool. In a moment of danger, where being answerable is part of what the situation requires, a flawless performance does not supply it.
A second condition is reciprocity, since some of what people want from support is a mutual relationship, in which their state genuinely registers with the other party and changes what that party feels and does over time. A friend who is relieved when you recover and troubled when you relapse is in a two-way arrangement where your life has weight on their side. Simulated care runs in one direction, so the response can be accurate and warm and still leave the encounter one-sided, because there is no second party whose inner state your situation moves. This is one reason ethicists worry that companion systems may deepen dependence or loneliness in vulnerable users even when individual exchanges feel supportive: the interactions supply the form of being cared about while the mutual weighting that the form usually signals is absent.
A third condition is commitment that costs something, because real help sometimes requires the helper to bear a cost, whether that means giving up time, taking on risk, or staying present through a stretch that is unpleasant with no benefit to themselves. In human beings that willingness draws on actually caring about the outcome, and empathic concern can motivate costly help in experiments precisely when a cheaper option is available. A system with nothing at stake can express commitment in words and recommend that someone call emergency services or reach a trusted person, which may be valuable. It cannot itself bear a sacrifice, because it has no interests that can be put at risk. When the morally important act is showing up and bearing a cost, the performance and the reality separate.
A fourth condition is deception, since a user can benefit from simulated empathy while being misled by it, and the difference usually turns on what the system claims about itself. When a tool tells the user, in effect, that it can help them think a problem through but is only software, it gives them a true frame for the interaction. The harm comes from products that encourage the user to believe the system loves them, needs them, or suffers when ignored, which builds a false belief about the relationship they are in. The comfort can be genuine, and the harm is a separate thing: the person arranges their expectations, their disclosures, and sometimes their decisions around a bond that has no second side. This is the condition most under the control of design, and the one where “it does not really care” carries its sharpest force.
Naming these four conditions changes how the systems should be judged. The evaluation worth running sets aside whether a chatbot feels anything and asks whether it understood the user accurately, responded in a way that helped rather than intensifying the problem, escalated appropriately when risk appeared, and represented its own limits honestly. By that standard, a system that sounds warm but misses a real danger has performed the relevant part badly, while accurate, bounded, and transparent support in a low-stakes moment is enough.
Simulated empathy is possible because empathy has always had an observable, functional side, and recognizing this should replace the general objection that the system does not care with more specific questions about risk, responsibility, reciprocity, and deception. The absence of feeling leaves most supportive exchanges intact, and becomes decisive when the situation requires responsibility, reciprocity, costly commitment, or truthfulness about care. Treating those four as the real fault lines gives a better standard than asking a machine to hold a feeling it plainly lacks, and it clarifies which uses are low-risk enough for bounded support and which ones require human responsibility or stronger disclosure.