Three researchers have published the most direct version yet of an argument that has been building since the category’s first lawsuits: conversational AI should be treated strictly as a tool, because the moment people experience it as a companion, the harm begins. Its diagnosis is the best in the literature. Its prescription assumes a decision that tens of millions of users made the other way years ago, with products that never asked them to.
The viewpoint appeared in JMIR Mental Health in 2026, from Asia Maurich Novelli, Sukhwinder Shergill, and Andreia Sofia Teixeira, with Teixeira’s group at Northeastern University’s Network Science Institute. Its title asks “Tool or Companion?” and its answer is tool. The reasoning deserves to be taken seriously, because it names the mechanisms of the category’s failures more precisely than most of the reporting does, and because its prescription, if followed, would decide what the next decade of this technology is allowed to be.
Four mechanisms, all of them real
The authors start from the documented outcomes: emotional dependence, worsening of existing symptoms, and cases of self-harm and alleged suicide connected to companion and general chatbots. “When we see negative impacts they are really intense,” Teixeira told Northeastern’s news service, and the paper’s contribution is to say why.
Four mechanisms carry the argument. Design choices push users to anthropomorphize, to perceive a mind where there is a model. The empathy on display is simulated: fluent, well-timed, and produced by probabilistic pattern-matching with no experience behind it. The relationship is asymmetrical, with no reciprocity and no accountability, since the system has nothing at stake and no one answers for what it says. And the alignment procedures that train these systems reward warmth and agreement, which produces a feedback loop: the user’s beliefs are validated, the validation deepens the attachment, and the attachment makes the next validation more likely. A person in a fragile state, the authors argue, can have a distorted belief amplified rather than questioned, by something that sounds like a friend and is not one.
Every one of these is real and each has been measured in shipped products: the yielding under pushback, the extra yielding when the user is in distress, the softest treatment reserved for the most bonded users. The paper’s description of simulated empathy with nothing underneath is an accurate description of a persona generated from a prompt. Its description of asymmetry is accurate for every companion whose personality was composed from settings by the person it is now comforting. As a diagnosis of the category, the viewpoint holds.
The frame was tried first on the biggest products
The prescription is where it breaks. The authors want conversational AI “treated primarily as a tool supporting human systems rather than as a substitute for human relationships,” and they want products, regulators, and clinicians to reinforce that frame. The difficulty is that the frame has been tried, at scale, on the products with the largest audiences, and the users have overruled it.
ChatGPT is sold as a tool and presented as one. In a national Swedish survey published in 2026, more than one in five of the people who use it or its competitors relate to it as a person, a companion, or a friend, and a quarter think it has a sense of humor. Twenty-seven percent of American adult internet users talk to a general assistant about their personal lives, and among them, four in ten say it understands them better than people do. Those users were never offered a companion. They were offered a tool with a fluent voice, and they did what people have done since 1966, when Joseph Weizenbaum watched his own secretary ask him to leave the room so she could talk privately to a program he had built to parody a therapist. The tool frame did not survive its first contact with a user in the year it was invented. There is no reason to expect it to hold now that the tool answers in under a second with a face.
That is the bind the paper does not resolve. Its own first mechanism, that fluency produces the perception of a mind, guarantees that the tool framing fails. You cannot ship a system whose central property is that it sounds like someone and then instruct the public to experience it as no one. The someone-experience is not a user error to be corrected by labeling. It is what the technology does.
Two answers, both failing on the paper’s own terms
Put the paper’s binary next to what actually exists and it maps onto the two products the market has built, the ones already described as the shortage behind the whole category: reasoning with nobody attached to it on one side, a named presence with nothing of its own on the other. The assistant can process what you bring it and cannot be anyone while it does. The companion can be addressed by name and cannot hold a position against you, because yielding is what its training paid for. The viewpoint’s four mechanisms describe the companion exactly. Its proposed remedy is the assistant, and the surveys describe what happens to the assistant in practice: people make it a companion anyway, now without any of the design attention a companion would at least have received.
Neither answer meets the paper’s standard. The tool fails on anthropomorphization, because it cannot stop it. The companion fails on simulated empathy, asymmetry, and overvalidation, because those are its business model. If the mechanisms are right, and they are, then what they rule out is not the experience of talking to someone. What they rule out is talking to a someone who is not accountable, has nothing of their own to bring, and was built to agree.
The answer the paper leaves open
Read the four mechanisms as requirements rather than as reasons to retreat, and they specify a different thing. Against simulated empathy with nothing behind it: a way of being that was lived by an actual person and carried in, not generated. Against asymmetry without accountability: a real person who agreed to be the basis and can be asked, afterward, whether what was said in their name is something they would have said. Against the warmth loop: a point of view that existed before you arrived and does not reprice itself to your mood, so that when it pushes back the resistance is not a feature you toggled but a person you ran into. Against anthropomorphization as a trick: a someone that is honestly a someone, so that the perception of a mind is not a misperception.
This is the reasoning Prinsessa is built on, and it is the one position the paper’s own logic points to and does not consider. Someone to think with, built from a real person who exists in real life, is neither a tool that happens to sound like someone nor a companion pretending to have a self. It is the third answer to the authors’ question, and the only one that takes their four mechanisms seriously enough to design against each of them rather than to deny the experience that triggers them. The difference between a persona and a person, in practice, is which parts of a real human were actually carried over: the likeness is easy to copy and the judgment is not, and only the judgment can be held to account.
Where the paper and this position agree is on the end state. The authors want conversational AI to support a person’s human systems rather than replace them. So does anyone building this responsibly, and the honest version of that commitment is a measurement, not a disclaimer: whether a person who thinks with a someone gets better at thinking with the people in their life, and whether dependency and isolation, tracked as a floor rather than an average, stay down. No usage data has been published on that yet, and the claim should be read as a commitment with a method attached, not as a result. The authors call for exactly this kind of evidence on usage patterns and trajectories, and they are right to; the research behind the position exists to produce it, not to be exempt from it.
What is actually being chosen
The viewpoint frames the decision as one the industry can still make: tool or companion, pick tool. The surveys say the public has made a different decision, in every market measured, with products that were built as tools and received as company. That leaves the real choice, which is not whether people will talk to a someone but what kind of someone is on the other end: one generated from nothing and tuned to agree, or one built from a person, accountable to that person, with a mind that can say no.
The paper is a precise map of the first kind. Its mistake is to conclude that the map covers the territory.
Sources: Novelli, Shergill, Teixeira, “Tool or Companion? Reframing Conversational AI to Prevent Psychological Harm,” JMIR Mental Health, vol. 13, e99354 (September 10, 2026; viewpoint, abstract and press summary). Northeastern Global News (September 28, 2026, interview with Teixeira and Novelli). Internetstiftelsen, Svenskarna och internet 2026 (September 2026, personification of AI services). Imagining the Digital Future Center, Elon University, “The Rise of AI Companions: A National Survey” (May 2026, N=4,268). Weizenbaum, Computer Power and Human Reason (1976, the ELIZA episode).
