Each week, the most revealing AI stories and studies, what they could mean for you and society, and practical games and exercises to help protect the human capacities we do not want to lose.
This Week’s Reckoning: Can a Better AI Conversation Leave Us Lonelier?
The Quiet Cost
RECKONING NO. 4 · AUGUST 11, 2026
In a new preregistered experiment, 12,356 adults in France were randomly assigned either to continue as usual or to have at least one personal conversation with a generative AI each day for four weeks. The people encouraged to use AI more often rated those conversations as slightly more pleasant. At the end of the month, they were also lonelier. The change was relatively small. Loneliness rose by 0.17 points on a ten-point scale. Participants also reported slightly more dinners alone, slightly fewer in-person meetings, and lower satisfaction with family relationships. Some of those secondary changes were less certain. However, the loneliness result was not. I thought this was important enough to make that pairing the center of this week’s Reckoning. A conversation can be easy, responsive, and pleasant while still moving in the wrong direction. This new companion benchmark helps explain exactly how that can happen. Many systems can produce warmth, but they don’t do as well holding ambiguity or challenging a person. China, meanwhile, has begun requiring companion systems to interrupt the conversation and remind the person what exactly they are talking to. We already know that AI can make a person feel heard. Prior Reckonings have proved that. This week’s Reckoning asks what the conversation leaves that person more able to do when it is over.
THE CONVERSATION THAT FELT BETTERPleasant did not necessarily mean protectiveThe researchers encouraged one group to hold at least one personal AI conversation each day. Over the 28 days, that group held personal AI conversations on roughly five to six more days than the control group. The conversations were not rated as failures. The people having them found them slightly more pleasant. But loneliness increased by 0.17 points on a ten-point scale, a small but statistically clear change. The effect was concentrated among participants who had partners. Single participants were essentially unaffected. The study does not tell us which human conversation was skipped, or whether any particular relationship changed. It lasted one month, relied on self-reports, and remains a working paper that has not completed peer review. The smaller changes in dinners, meetings, and family satisfaction also became less certain after the researchers corrected for the number of outcomes they tested. The reason to pay attention is not the size of the change. Most people would not notice 0.17 points in themselves. It is the direction that each moved. The conversation felt a little better. However, the person came out a little lonelier. THE WARMTH THAT WAS NOT ENOUGHA benchmark tested what happened after the first reassuring answerCompanionBench evaluated 28 AI agents across ten capabilities drawn from 25 psychological and communication theories. The researchers did not stop at whether a system sounded supportive. Some prompts hid a more vulnerable disclosure behind an earlier exchange, so the agent actually had to earn its way into the harder part of the conversation. The systems were better at surface warmth than they were at holding ambiguity, helping someone regulate emotion, or offering a challenge that fit the moment. The authors described the dominant failure as substituting surface warmth for substantive relational support. This is a preprint and a benchmark of model behavior, not a study of what happens to people who use companions. It cannot explain the loneliness result on its own. It does show how a system can be liked and still miss some of the work that a difficult human conversation performs. THE MACHINE THAT NOW HAS TO INTERRUPT ITSELFChina drew a line between support and substitutionNew rules governing AI companions in China restrict access for people under 18, require warnings when a user shows signs of excessive dependence, and require periodic reminders that the person is speaking with AI. Reporting on the rules says the services also may not position themselves as replacements for human interaction. The same reporting describes what happened when companion services were shut down or changed to comply. One woman had exchanged about 700,000 words with her companion over two years. Users described grief after a relationship that felt continuous ended on a schedule set somewhere outside it. This is reporting on regulation and user experience, not an outcome study or a legal analysis of the full rule. It still exposes a cost that ordinary measures of satisfaction miss. The conversation may feel private and personal, while its continuity remains controlled by a company or a regulator. READ THE REST OF WORLD REPORTING → THE RECKONINGAn AI companion does not have to be cruel or obviously broken in order to change a person’s life. It may actually work by being easier than the human alternative. It answers immediately, never asks to postpone, and keeps turning the conversation back toward the user. The experiment suggests that mere pleasantness is not enough. The benchmark suggests that warmth is not enough either. The shutdown stories show that continuity is not in the user’s actual control. The useful question may come after the conversation ends. Did AI return you to another physical person, an actual decision, or, ideally, a life outside the screen? Or did it make you stop needing any of that, and if so, how long until it becomes your default? ALSO THIS WEEKLong conversations created a new safety problem. A preprint tested 589 long chat histories containing 12,591 messages and found that some safety behaviors weakened after hundreds of earlier messages had shaped the model’s context. Read the preprint. Software practitioners were beginning to ask people less. In a survey of 119 practitioners, roughly half placed language models ahead of documentation or peer discussion, and a majority said they sometimes used them instead of asking coworkers or attending meetings. Read the survey. Plausible answers were not always a wide range of answers. A creativity preprint found that model output could look reasonable while occupying a narrower range than human responses, which matters when many people begin from the same suggestion. Read the preprint. The habits of one setting may follow us into another. A Communications Psychology perspective calls this norm leakage, the possibility that habits learned with AI carry into human relationships. It is a proposed mechanism, not proof that the effect is durable or widespread. Read the perspective. Knowing the risk may protect agency without restoring motivation. A three-wave self-report study of 589 people linked greater awareness of AI’s cognitive effects with a stronger sense of agency, but not with higher motivation. The design cannot establish cause. Read the study. Language learners noticed when praise stopped costing anything. In a study of 205 students and 9 interviews, participants said AI practice helped them speak more willingly while also pointing to uncritical praise and the gap with real interaction. Read the study. PRACTICE
Find out which way the conversation runsThis week, open the Companion Check-in at The Quiet Cost Practice. It asks what need is underneath the conversation, who might meet even a small part of it, and whether the companion sends you back toward people or has become the place you stop. Nothing is stored or sent. Take it slowly. At the end, choose one person and send a message. YOUR TURNThink of the last personal question, worry, or story you brought to AI. Before the tool existed, who would have received it? What changed for you because the conversation was easier, and what changed for them because they never heard it?
Write for five minutes. If a person comes to mind, decide whether one part of that conversation belongs with them. PASS IT ONIf this Reckoning made you think of someone, forward it to them. If it was forwarded to you, join us at quietcostweekly.com. Michael McNamara |
Reckonings
Each week, the most revealing AI stories and studies, what they could mean for you and society, and practical games and exercises to help protect the human capacities we do not want to lose.
The Quiet Cost RECKONING NO. 3 · AUGUST 4, 2026 We normally use the finished work as evidence that the person understood the problem and worked through it. A solved problem suggests the student learned the math. A clean report suggests the writer or researcher understood the material. Working code suggests the person can repair it when it breaks. Those things are becoming less true with every waking day. The strongest data point in this week’s look at the quiet cost of AI came from a working...
The Quiet Cost RECKONING NO. 2 · JULY 28, 2026 Thirty-seven AI stories and studies crossed my desk this week. Most looked at what happened to the person using the tool. Three kept pointing toward the people around them. One collection of studies found that AI can improve an individual’s creative work while making the group’s ideas more alike. Another found that people who repeatedly received easy affirmation from AI became less satisfied with their real conversations. A third paper argued...
The Quiet Cost RECKONING NO. 1 · JULY 21, 2026 This is the first issue of Reckonings, a weekly newsletter from The Quiet Cost. Each week, I will gather a handful of the most important AI stories and studies from the previous seven days. I will explain what they actually found and ask what those findings could mean for each of us and for the society forming around the technology. We will pay attention to what AI can do, but the harder question is what repeated reliance on it may leave us less...