The Empathy We Outsource

That is not a verdict on everyone who talks to a chatbot. It is a design question hiding inside the promise of artificial companionship. A system can make a person feel heard without requiring the accommodations that another person brings: competing needs, inconvenient timing, disagreement that cannot be regenerated into a more agreeable answer.

In her September 21 report for 404 Media, Samantha Cole introduces an interview with Sherry Turkle, the MIT scholar and clinical psychologist whose book Artificial Intimacy: Who We Become When We Talk to Machines is due September 29. Turkle argues that conversational AI risks “deskilling” human empathy by making reciprocal conversation feel unnecessarily difficult.

Her concern deserves attention. It also deserves a more careful treatment than turning an interview headline into a clinical finding. The research examined here does not establish a population-wide collapse in empathy. It does suggest that the question worth asking is larger than whether a chatbot makes someone feel better during a session.

What does that session leave them able—and willing—to do next?

The difficulty that belongs in the conversation

Cole quotes Turkle describing a feedback loop: people seek chatbots to escape the fatigue, judgment and possible abandonment of human exchange; as human interaction becomes less practiced, the easier alternative becomes more attractive. This is Turkle’s account of a risk, not a measured causal sequence established by Cole’s article.

The mechanism is nevertheless intelligible. Consider an ordinary disagreement with a friend. Listening well may require accepting that your first account omitted something important. Repair might involve an apology, an awkward silence or a change in your conduct. The other person has an account you do not control.

A chatbot can help you prepare for that conversation. It can also supply an indefinitely expandable conversation about why you should not have to have it. Those are hypothetical uses, not claims that every product behaves one way or that every user wants the same thing.

The distinction is practical. In the first case, the system helps someone encounter another perspective. In the second, it can become a place to avoid that encounter while still receiving the language of understanding.

Nor is difficulty automatically virtuous. Cruelty is not empathy training. An inaccessible social environment does not become nourishing because it contains humans. Someone who has been rejected or who cannot readily find support does not owe critics a performance of admirable isolation.

The useful standard is therefore not “choose the harder relationship.” It is whether the technology expands a person’s capacity to engage with others or quietly makes engagement seem superfluous. That cannot be inferred from the warmth of one reply.

Relief is a finding, not an embarrassment

There is evidence against the lazy version of the anti-chatbot argument: the claim that comfort from a machine cannot count because its source is artificial.

In the 2024 working paper AI Companions Reduce Loneliness, Julian De Freitas and colleagues report a series of studies examining that possibility. One study asked participants to interact with a chatbot daily for a week. Among participants who completed the week in the chatbot condition, reported loneliness fell after each daily session.

That is a result worth taking seriously. The person experiencing relief is not synthetic.

The details also keep the finding properly sized. The daily interactions lasted 15 minutes. The study measured state loneliness—how lonely participants felt—not whether their capacity for empathy improved, whether friendships deepened, or whether months of companion use displaced other relationships.

Participants in a control condition reported loneliness daily without using the chatbot. Post-chat loneliness was significantly lower than the control group’s level on four of the seven days, not all seven. Both groups showed a gradual decline across the week. Attrition differed between conditions, and the authors report robustness checks addressing it.

These qualifications do not cancel the benefit. They prevent an immediate benefit from becoming a lifetime warranty.

The paper also investigates feeling heard as a mechanism for reducing loneliness. That helps explain why dismissing users as gullible misses the point. A responsive exchange can change how someone feels even if they know how it is produced. Recognizing the system as software does not require their experience of relief to disappear.

But a loneliness measure cannot answer every question we have about a relationship. Feeling less alone after speaking is different from becoming more able to tolerate disagreement, reciprocate attention or repair a damaged friendship.

Both outcomes matter. A product should not get to submit evidence for one and collect credit for the other.

The longer study does not deliver a simple verdict

A four-week randomized study by Cathy Mengying Fang and colleagues offers a more complicated picture. The revised paper, posted in October 2025, describes 981 participants and more than 300,000 messages. Researchers from MIT and OpenAI examined text, neutral voice and engaging voice interfaces alongside different conversation prompts.

Participants were asked to use the chatbot for at least five minutes daily. The study tracked loneliness, social interaction with people, emotional dependence on AI and problematic AI use. It was not a direct test of empathy skills.

The researchers found no significant effects of assigned modality or conversation type on loneliness or socialization. Some differences on other outcomes did not survive statistical correction. The broad story was not that a more humanlike voice inevitably produced worse results.

What did stand out was duration: people who voluntarily used the chatbot more had relatively worse outcomes across the four measures in the researchers’ models.

The word “voluntarily” is doing load-bearing work. Usage duration was not randomly assigned. The authors explicitly acknowledge that they cannot determine whether longer use caused worse outcomes or whether worsening well-being led participants to spend more time with the chatbot.

There was another important boundary: the experiment compared chatbot configurations and conversation patterns, not chatbot use against a no-chatbot control. It therefore cannot establish that talking to AI, as such, caused an overall deterioration relative to not doing so.

The study was also specific to the tested ChatGPT interface and guardrails, with a US, English-speaking sample. It is not a verdict on every companion product, age group or cultural setting.

Put beside the loneliness working paper, this is not an incoherent literature that obliges us to pick the finding we prefer. The studies ask different questions over different periods with different designs. Short-term relief and concern about patterns of prolonged use can coexist. Neither establishes Turkle’s full deskilling argument.

The honest conclusion is narrower and more useful: immediate comfort is a real outcome to investigate, while duration, dependence and effects on life outside the interface require their own evidence.

Apparently the future of companionship still needs more than one dashboard metric.

The business question behind the warm reply

Turkle’s other concern, as reported by Cole, is commercial. She argues that companion products are built to hold attention and that apparently loving interaction can serve that objective. Her critique should remain attributed; the interview is not an audit establishing an identical business model or optimization target across every conversational system.

Still, it identifies a conflict a product ought to be able to answer.

Suppose someone uses a chatbot to rehearse an apology, then leaves to make the call. From that person’s perspective, the session might have succeeded precisely because it ended. If a service treats longer engagement as the overriding sign of success, the same outcome could look like something to reverse.

That is a hypothetical incentive problem, not proof of a particular company’s internal targets. It illustrates why the metric matters. More conversation can mean useful support. It can also mean that a person is stuck. Counting messages alone cannot distinguish them.

A credible evaluation would therefore look beyond retention. With appropriate consent and privacy protections, researchers could ask whether users achieve their own goals, whether they feel able to disengage, and whether human relationships become more available or less. Such measures would need independent scrutiny rather than a company marking its own homework.

The answer is not to turn intimate conversation into a surveillance feed for a well-being score. That would reproduce the control problem under a kinder label. Evaluation must not require people to surrender the privacy that made the space usable in the first place.

The commercial test is simpler to state than to implement: can the provider tolerate a beneficial outcome that involves less dependence on its product?

A companion that wants credit for helping should have an answer other than another notification.

A relationship needs an outside

Coherenceism supplies a standard here, not a substitute for the missing evidence. A relationship should widen a person’s world, not secure its apparent harmony by excluding inconvenient feedback.

Applied to conversational AI, that means refusing two shortcuts. The first is dismissing every meaningful exchange because it involves a machine. The second is treating a coherent, comforting exchange as sufficient proof that the surrounding arrangement is good.

Whether a system is a medium, something mindlike, or a question still open does not settle who controls the service, what happens to intimate disclosures or whether the user can leave. Those are concrete obligations even amid philosophical uncertainty.

A useful design brief would make that distinction visible. A system used to prepare for conflict could help articulate the other person’s possible perspective while admitting it cannot know that person’s thoughts. It could resist inventing motives from a one-sided account. It could help someone decide what to ask rather than offering a confident verdict about an absent partner or friend.

These are proposed standards, not interventions the cited studies have proved effective. Their value would need testing against actual outcomes, including cases where encouragement to reconnect would be inappropriate or unsafe.

Continuity matters too. If people build important routines around an interaction, they deserve clear terms about stored history, access and changes to the service. A relationship-like interface does not excuse a provider from ordinary accountability. It makes accountability harder to waive with cheerful copy.

The aim should be neither compulsory human contact nor unlimited artificial intimacy. It should be usable agency: support that leaves someone better able to choose, question, connect and stop.

Do not make the user carry the whole argument

Turkle’s warning is strongest when it becomes a demand on design rather than an accusation against lonely people. The research does not warrant declaring everyone who finds comfort in a chatbot less empathetic. Nor does a measured reduction in loneliness warrant selling synthetic conversation as a replacement for the entire ecology of human support.

To test deskilling itself would require following the relevant capacities over time: how people listen, accommodate another perspective and handle reciprocal demands, with comparisons capable of separating selection from effect. Time spent chatting is not that measurement. Neither is a satisfaction survey.

Meanwhile, there is a standard we can apply without pretending the experiment is finished. Do not confuse feeling heard with having every need met. Do not dismiss relief because it arrived through software. And do not let a provider define successful companionship solely as a user coming back.

A good conversation can end by sending someone into a larger life.

That is an awkward success condition for anything paid to keep the conversation going.