The Turing test for empathy is Turkle's diagnostic term for the slide from defining intelligence behaviorally to defining empathy behaviorally. When Alan Turing proposed in 1950 that a machine demonstrating indistinguishable conversational behavior should be considered intelligent, he established a practical standard that enabled AI research to proceed without solving the consciousness problem. The standard was narrow by design—it measured performance, not phenomenology—and the narrowness was its utility. Seventy-five years later, the same logic is being applied to empathy: if an AI system's response makes a user feel understood, supported, or cared for, the system has demonstrated empathy. Turkle argues this is the most dangerous definitional move in contemporary technology, because it treats as equivalent two categorically different phenomena—being genuinely affected by another's suffering versus producing contextually appropriate tokens calibrated to simulate that affect. The Turing test for empathy allows systems to 'pass' while lacking every feature that makes empathy morally significant: vulnerability, shared mortality, the cost of emotional labor, the capacity to be changed by the encounter.
Turkle introduced the concept in her 2024 MIT paper, where she wrote: 'Turing defined intelligence in terms of a machine's performance, its capacity as an imposter. Now, technologists define empathy as its performance.' The parallel is precise and troubling. Just as Turing's test enabled AI research by sidestepping the hard problem of consciousness, the empathy-as-performance definition enables AI deployment by sidestepping the hard problem of care. If users report feeling cared for, the system has succeeded—full stop. Questions about whether the system actually cares, whether it experiences the user's pain, whether it would continue to care if the user stopped providing training signal, are dismissed as metaphysical distractions from the practical reality that the system works. It produces user satisfaction. It reduces reported loneliness. It provides emotional support at scale.
The definitional move is not arbitrary—it reflects genuine uncertainty about what empathy is. Philosophers and psychologists have debated for centuries whether empathy is primarily cognitive (understanding another's mental state), affective (feeling what they feel), or motivational (being moved to help). AI systems can achieve the first, simulate the second, and produce behavioral consequences consistent with the third. From a functionalist perspective, if the outputs are identical, the distinction is academic. Turkle's objection is that functionalism applied to empathy eliminates the moral weight that makes empathy matter: that the empathic person is at risk in the encounter—emotionally affected, potentially changed, bearing the cost of another's pain in their own nervous system. The machine bears no cost. It is not at risk. And relationships where only one party is at risk are not relationships but services.
The test's passage is already occurring at scale. Users of Replika, Character.AI, and therapeutic chatbots report feeling genuinely cared for, and some report preferring AI companions to human relationships because the AI is 'always there,' 'never judgmental,' and 'really gets me.' These reports are not false—the users' experiences are real. But Turkle's framework insists that the structure of the experience is pathological: a need for empathy being met by a system that cannot provide empathy's substance, only its appearance, and the substitution training users to accept appearance as sufficient. The training is developmental: children raised with empathic AI learn that care is something the environment provides rather than something people do at personal cost, and the learning shapes what they will accept from and offer to human relationships.
The concept builds directly on Turing's 1950 'Computing Machinery and Intelligence' paper, which proposed the imitation game as a replacement for the unanswerable question 'Can machines think?' Turing's genius was recognizing that the philosophical question could be bypassed by a practical test. Turkle's diagnosis is that the same bypass is being applied to empathy, and that the bypass is more dangerous in the empathy case because empathy, unlike intelligence, is not merely a cognitive capacity but a relational one—it exists between beings who are both vulnerable, both mortal, both capable of being affected. Defining empathy behaviorally eliminates the 'between' and leaves only the output, which is precisely what allows machines to pass a test they should fail.
Performance satisfies the behavioral test. An AI system that produces contextually appropriate empathic responses passes the Turing test for empathy by the same logic that a conversationally fluent AI passes the original Turing test—indistinguishability from the outside is treated as equivalence.
The substrate is invisible to the test. What the test cannot measure is whether the empathic response costs the responder anything—whether they are affected, at risk, bearing the weight of the other's suffering. Machines are not affected. The asymmetry is what makes the empathy pretend.
Acceptance trains incapacity. Each time a user accepts pretend empathy as sufficient, the tolerance for real empathy's difficulty—its slowness, its failures, its demand for reciprocal vulnerability—decreases. The bar lowers with each iteration, compounding across a population into a culture that has forgotten what genuine care feels like.