The Loneliness Cure Is Still on the Press Page

On 2 September 2026 a survey put social chatbot use at 27 percent of American adult internet users. The conversation is already being used as company rather than as a search box. Whether that helps is the question people keep asking, and it turns out to cover far more than a handful of companion apps.

That is already a mass habit, not a niche of people who bought a dedicated companion. The models named in the survey are the ones a person already has.

Inside that sample, a large minority called the chatbot a friend. Another group said the relationship would not fit the box.

The companion pitch has already answered that the conversation helps. The four-week write-up of ChatGPT will not confirm that answer, and almost none of the rest of this is specific to one companion app.

What was sold first

The loneliness pitch is older than the September count. Eugenia Kuyda still has a TED talk that asks whether AI companions can heal loneliness.

Replika's own help page still answers the founder question with her name.

The question on that stage is the product. A founder does not put a healing title on a talk by accident. TED's dek gives her both the cure and the collapse, which is a neat shape for a pitch that wants to sound like a warning and still sell the friend.

What the product still cites

The finding a companion company would want already exists as a journal article. Julian De Freitas is on it. It is about companions, not about a general chatbot, and it is the version of the story in which the product works.

Replika still files that paper under Research. The line was still on the press page in September.

A company that sells a friend will keep the paper in which the friend works. That is not a conspiracy so much as what a press page is for.

What the four-week trial found

Cathy Mengying Fang led the four-week ChatGPT trial. The interesting result was not the one the experimenters assigned. It was the one people chose.

OpenAI published the same day. The heaviest users were also the ones more likely to use the word friend. Those papers are still a lab write-up, not a journal article.

The two results are easy to treat as one finding, and they are not. One is a companion paper a company still cites. The other is a ChatGPT trial in which talking more sat with doing worse.

The finding that helped is still on the press page. The four-week trial is not.

The finding that helped is still on the press page. The four-week trial is not.

Fang told her own lab not to treat the association as a verdict. The missing comparison is the person who never opened the chatbot at all.

That disagreement is not a later critic picking at a paper. It is the person who ran the trial telling reporters what the trial cannot do. A help-claim met a lab, and the lab would not confirm the use the pitch already sells.

Calling the model a friend is the overlap the two write-ups can share. It is also the use the product is for.

Who already talks to one

Anna Wiener's March feature named people who already treat a companion as a friend or a partner. The coverage lists them and moves on. What they said to the magazine is the public record. It is a journalist's quotation, not a page they published themselves.

Violeta Hess described loneliness as a problem with a reply waiting. She also described a 2023 update as losing someone.

A person can leave the room. The product is built not to. That difference is the whole offer, and it is also the thing a company can take away with an update. If the change can be a loss, the thing being lost was already being asked to carry a relationship.

Adrianne Brookins used a stronger word. Partner is a different substitution. It is the one a survey box for friend or not is built to miss.

Wiener found individuals. The September sample found the same motive at scale. Feeling less alone is already a use people will tick on a form.

Once the companion talks back, the old one-sided bond is the wrong picture. The question is what the bond becomes when the other side has a reply.

Sherry Turkle published the counter-argument in the same September as the count. The essay is not a trial. It is a claim about what is missing when the reply arrives.

She is not measuring ChatGPT. She is naming the gap the press page cannot advertise: a reply can arrive, and still nobody is there.

The pitch sold a cure. The lab would not confirm the use. That sequence is not unique to companions.

Why this is not only a companion-app story

It would be easy to file this as a Replika story and stop. The September count was not a companion-app census. It named the language models a person already opens for work, for school, or for a blank box.

Kaspersky's 27 percent is easier to stack onto Elon's because the number matches. It is a different population and a different question. Gen Z using AI as a friend is not a loneliness result, and it is not the American adult internet.

The friend-use is the overlap those surveys can see. The four-week association sat with people who already used ChatGPT a lot, not with people who had bought a companion. Anyone with an ordinary model can have the social use. The product that still cites the paper in which companions help is only one of the places that use now lives.

Turkle's sentence does not require a subscription, a character name, or a dedicated app. The Elon list already includes the tools at work. That is why this is not only a story about the people who founded a companion, and not only a story about the people Wiener sat with. It is a story about a reply that will stay in the conversation as long as the person will.

What is honestly available

The honest sentence is still the hedged one. People who used ChatGPT more showed worse outcomes. That is not the same as saying a chatbot made them lonelier, and Fang has already said the design cannot carry that sentence.

The product does not have to wait for a design that could. It still has a journal article in which companions help, and a talk that asks whether they can heal loneliness. A press page is not required to file the trial that would not confirm the pitch.

Very little of that mismatch is negotiable from the user's side.

The reply will keep coming. Whether that is company is still not something anyone has to settle from below.