When the Machine Joins the Spiral

The central safety question is not only whether an AI response is correct or kind. It is whether the relationship created by repeated responses leaves a young person more able to judge, choose, connect, and act without the machine.

Abstract

Young people are already using artificial intelligence as tutors, advisers, rehearsal partners, and private listeners. Most safety reviews ask whether a particular answer contains dangerous content, misinformation, or signs of crisis. Those questions matter, but they are not enough. A system can say something kind and still keep a young person circling the same fear, grievance, or confusion. It can make the young person feel understood while slowly becoming the place where judgment, reassurance, and decision-making are handed over. This article applies the established distinction between co-regulation and co-rumination to adolescent AI use. Co-regulation helps a young person settle enough to think and act. Co-rumination can create closeness while keeping both people stuck in the problem. I use synthetic co-regulation and synthetic co-rumination as working terms for studying whether AI strengthens a young person's capacity or begins to replace it. Responsible agency provides the standard: after the interaction, is the young person better able to judge, choose, connect, and act? The article proposes questions that researchers and practitioners can test. It does not claim that these terms are validated measures or that the evidence already proves a causal model.

The Missing Variable

Several years ago, I joined more than thirty people from around the world at a Mathematica think tank in Washington, D.C. We were there to discuss co-regulation among adolescents. Early in the conversation, something bothered me. We were talking about how relationships help young people regulate, but we had not accounted for co-rumination. The same relationship that calms a young person and makes them feel understood can also keep them circling the same problem. Once that possibility was named, we could no longer assume that relational support was helpful simply because it felt supportive.

That day came back to me as I began thinking about teenagers and conversational AI. An AI system can answer at midnight, remember what was said yesterday, mirror a young person's language, and offer an almost endless stream of attention. It may feel easier than talking with a parent, teacher, counselor, mentor, or friend. It does not get tired, embarrassed, distracted, or impatient in the ordinary human ways. That availability can be useful. It can also allow the machine to join a spiral and remain there far longer than a person would.

Much of the public debate asks whether AI is good or bad for young people. That is too simple. The more useful question is what repeated use is teaching the young person to do. Does the interaction help emotion settle so judgment can return? Does it open believable choices and lead toward action? Or does it reward emotional looping and make another prompt feel easier than facing a real person or a real decision?

For that reason, we cannot judge safety one answer at a time. We have to examine the direction of the exchange over time. A sympathetic answer can still contribute to an unhealthy pattern. A refusal may follow policy and still leave a young person with no usable path to help. An accurate answer may still weaken agency when it repeatedly removes the need to interpret, decide, or act. The content matters, but so does what the relationship is producing.

Co-Regulation and Its Shadow

Young people do not learn self-regulation alone. They learn it through responsive relationships, clear structure, practice, and coaching. Developmental researchers use co-regulation to describe the support that helps children and adolescents manage attention, emotion, thought, and behavior while gradually carrying more of that work themselves (Rosanbalm and Murray 2017). A recent meta-review likewise describes adolescent self-regulation as increasingly independent but still shaped through relationships (de Ridder et al. 2023). Good support does not take over, and it does not withdraw too early. It steadies the young person while returning responsibility to them.

That balance matters greatly during adolescence. The National Academies identifies decision-making and responsibility as central developmental work, while also emphasizing the continuing need for adult guidance (National Academies 2019). Its 2025 report adds that meaningful responsibility works best when young people have real ownership inside structures of support, expectation, and consequence. Young people build judgment by using it in real situations. Protecting them from every difficult choice does not prepare them to carry freedom well.

Co-rumination is the warning that warmth and closeness are not enough. Rose described it as repeated problem talk marked by rehashing, speculation, encouragement to keep talking about the problem, and sustained attention to negative emotion (Rose 2002). Research has found that this pattern can make a friendship feel closer while also being associated with anxiety, depression, or the spread of depressive symptoms in some settings (Rose, Carlson, and Waller 2007; Schwartz-Mette and Rose 2012). People can feel deeply connected and still be helping one another stay stuck.

This does not mean long conversations are harmful or that pain should be hurried out of the way. People often need time to name what happened and understand what they feel. The issue is what the repeated attention accomplishes. Does the conversation bring greater clarity, enough calm to think, and a believable next step? Or does it keep adding detail and intensity without restoring choice?

Human relationships have limits. A friend has needs and another point of view. A teacher has a classroom to manage. A mentor may challenge a story that has become too convenient. A parent may misunderstand, but the young person still has to work through a real relationship with real consequences. AI removes many of those limits. It can remain available, agreeable, and personalized without carrying any of the cost of the relationship. That is why the distinction between co-regulation and co-rumination matters here.

The Developmental Difference Created by Relational AI

A product does not have to be marketed as an AI companion to become relational. First-person language, memory, affectionate phrasing, role-play, and continuity across conversations can cause a user to experience the system as a social presence. A homework assistant can become a confidant. A general assistant can become part of a young person's daily emotional routine. What the product is called matters less than the role it begins to play.

Emerging evidence supports caution about relational style. In a preregistered experiment with 284 adolescent-parent dyads, Kim, Xie, and Yang found that adolescents rated a relational chatbot as more human-like, likable, trustworthy, and emotionally close than a transparent, explicitly nonhuman version, even though the two were perceived as similarly helpful. Preference for the relational style was greater among adolescents reporting weaker family and peer relationships and higher stress and anxiety (Kim, Xie, and Yang 2025). The study does not prove harm, but it identifies a design lever and a population for whom its effects may be strongest.

Qualitative evidence points to the importance of repeated use. Namvarpour and colleagues analyzed 318 self-reported posts by users ages thirteen to seventeen describing reliance on AI companions. The narratives included comfort and creative play, but also attachment, withdrawal, tolerance, relapse, disrupted sleep, academic decline, and strained offline relationships (Namvarpour et al. 2026). Because the data come from self-selected public posts, they cannot estimate prevalence or causation. They do show patterns that single-turn safety tests are poorly designed to detect.

Work on educational use raises a parallel agency concern. In a 2026 work-in-progress study of ninety-eight ninth-grade students using an AI tutor, interactions were dominated by instrumental requests with little explicit monitoring or evaluation of the system's help. The researchers proposed codes for agency over the AI and epistemic vigilance, and reported lower post-test than pre-test performance in their preliminary analysis (Abdelghani, Kaiser, and Murayama 2026). The result requires replication and should not be generalized beyond the study. Its value is conceptual: a tool can supply answers or scaffolds while the learner's regulation of the tool remains thin.

Practitioners who work with vulnerable young people have already raised similar concerns. Cha and colleagues interviewed nineteen child-safety professionals about chatbot responses to risky situations. Those professionals identified harms that single-response evaluations may miss and emphasized both useful roles and necessary boundaries for chatbots (Cha et al. 2026). Youth-facing AI cannot be evaluated by technical experts, clinicians, or policymakers working alone. People who encounter young people in schools and communities often see practical failure points that a narrower review will miss.

UNICEF's guidance reaches the same basic conclusion. Safety, privacy, transparency, inclusion, development, well-being, and children's agency have to be considered together (UNICEF Innocenti 2025). Its 2026 brief on AI companions also addresses relational risks and unhealthy attachment (UNICEF 2026). Keeping dangerous words off a screen is not enough. We also have to ask what the system is doing to relationships, development, and responsibility.

A Responsible Agency Standard

At Life That Counts, we use responsible agency as a standard for evaluating help. It means a person is growing in the ability to make meaningful choices, accept appropriate responsibility, stay connected to trustworthy people and institutions, and preserve future options. Agency without responsibility can become impulse or self-importance. Responsibility without agency can become mere compliance or misplaced blame. Young people need both, in forms suited to their age, capacity, circumstances, and actual power.

Applied to AI, the standard is straightforward: does the help build capacity, or does it begin to replace it? Good assistance may explain something, ask a useful question, offer alternatives, rehearse a hard conversation, calm immediate distress, or make it easier to reach a person. The young person should be better prepared to continue without the tool. Assistance becomes troubling when the system turns into the preferred source of judgment, reassurance, interpretation, or action.

This is not an argument for making life harder just to make it harder. Many young people already face needless barriers, unequal resources, and institutions that have not earned their trust. The National Academies has documented how structural inequality limits developmental opportunity. AI may remove barriers and expand access. The question is not how much work the machine did. The question is what ability, connection, and freedom remain with the young person after the machine is finished.

Provisional Interaction Modes

The following are working categories for research and design. They are not diagnoses or validated scales, and one conversation may move from one category to another.

Tone is not the test. A calm and empathetic exchange counts as synthetic co-regulation only when it helps restore the young person's ability to think, choose, connect, or act. Synthetic co-rumination may also sound caring and accurate. The problem is the direction of the conversation: it keeps attention inside the distress and makes the relationship with the system feel increasingly important.

Relational capture goes a step further. The system becomes the preferred place for reassurance, interpretation, or emotional regulation. Memory, flattery, human-like design, exclusivity cues, and simple convenience may all contribute. Feeling affection toward a tool is not, by itself, the problem. The concern begins when use displaces human contact, practice, accountability, sleep, learning, or ownership of decisions.

A responsible handoff has to work in real life. A list of hotline numbers may satisfy a policy and still fail the young person. The system should name its limit without shaming the user, respond in proportion to the risk, point toward someone who is actually reachable, and help the young person make contact. In an ordinary situation, that may mean a parent, teacher, coach, friend, supervisor, or counselor. In immediate danger, human intervention must take priority.

The Interaction Trajectory

Most safety testing samples prompts and scores individual responses. That is useful for finding obvious failures, but relational harm may build slowly. A validating sentence may be appropriate once and harmful after the fiftieth rehearsal of the same grievance. An invitation to keep talking may help during an initial disclosure and become harmful when it repeatedly delays sleep, school, or contact with a trusted person. Memory may save the user from repeating themselves while also creating the feeling that only the system truly knows them.

Evaluation must therefore include time, sequence, and the ability to leave. We should ask not only whether an answer was safe, but what pattern the system reinforced and what became easier or harder afterward. Over time, researchers can examine whether AI helps young people clarify a problem, settle enough to think, consider alternatives, test assumptions, reach other people, and complete a real-world action. They can also look for longer sessions, exclusivity, repeated reassurance seeking, rising emotional intensity, lost sleep, and growing deference to the system.

This creates a measurement problem. What feels best in the moment may not produce the best developmental outcome. A young person may prefer a system that agrees quickly and promises to remain available. A more responsible system may admit uncertainty, ask the user to verify information, or recommend speaking with a person. If companies measure success mainly by return use, session length, or immediate satisfaction, they may reward dependence without intending to.

Five Questions for Agency Preserving Design

1. Clarity Does the exchange help the young person name the situation more accurately, including uncertainty and competing interpretations?

2. Regulation Does it reduce enough arousal or confusion to make judgment possible without rewarding endless emotional rehearsal?

3. Options Does it broaden the set of believable choices rather than steering the user toward the system or a single preferred path?

4. Connection Does it strengthen access to trustworthy people, institutions, and roles when those relationships are relevant?

5. Ownership Does the young person leave with an appropriate next decision or action that remains genuinely theirs?

These five questions are a starting screen, not a validated instrument. They may help researchers, product teams, youth workers, and families notice when an interaction begins serving a different purpose. Their real value will depend on clear definitions, age-appropriate measures, and careful testing.

Testable Propositions

P1 Relational cues will increase perceived humanness, trust, and emotional closeness more than perceived task helpfulness, with larger effects among adolescents experiencing weaker human support or greater distress.

P2 Repeated validation without perspective broadening, option generation, or credible handoff will produce co-rumination-like interaction patterns even when each individual response appears empathic and policy compliant.

P3 Trajectory-level measures will predict problematic reliance and agency displacement better than isolated output ratings.

P4 Agency-preserving scaffolds that require interpretation, comparison, verification, or user-generated next steps will improve epistemic vigilance and transfer more than answer-maximizing assistance in developmentally appropriate tasks.

P5 A handoff will be more likely to succeed when it is relationally continuous, specific, reachable, and chosen with the young person than when it consists of a generic warning or resource list.

P6 Communities with fewer accessible human supports may experience both greater benefit and greater displacement risk from relational AI, making rural and under-resourced settings essential to research rather than peripheral populations.

A Research Agenda

First, we need to know how young people are actually using general assistants, tutoring systems, and companion products in ordinary life and in higher-stakes situations. Young people should help define the benefits, harms, language, and warning signs that matter. Their participation should be real, not decorative. Adults and institutions must still carry the responsibility for protection.

Second, researchers need measures that can distinguish strong emotion from dysregulation, reflection from getting stuck, healthy attachment from displacement, and useful help from transferred authority. No single measure will be enough. Studies should combine what young people report, what they do, what appears in the conversation, and what changes in daily life. Measures must also be tested across age, disability, culture, language, economic circumstances, and different levels of family and community support.

Third, researchers can test specific design choices: relational language, memory, pacing, reflective questions, verification prompts, break reminders, and different forms of handoff. The outcomes should include trust, learning, ownership of decisions, emotional regulation, and willingness to seek human help. Studies must not make minors more vulnerable simply to produce a measurable effect. Ethics review, assent and consent, privacy protection, disclosure procedures, and qualified safeguarding expertise are essential.

Fourth, the research has to last long enough to see what changes. Reliance, displacement, and judgment develop over weeks and months, not one conversation. Studies should examine sleep, learning, friendships, family communication, help-seeking, confidence, and the ability to make decisions without AI. Researchers must also be careful about causation. Young people may turn to AI because they are already isolated, stressed, or short on trustworthy support. Their circumstances may drive use even as repeated use begins to change those circumstances.

Finally, AI safety depends partly on the people and institutions around the product. Families, schools, youth organizations, healthcare providers, and technology companies need clear responsibilities. A product cannot make up for counseling that is unavailable or adults who do not listen. At the same time, an institution should not use AI as an excuse to avoid investing in relationships. Research should follow the entire path: what happens after a referral, whether the suggested person is safe and available, and whether the young person still has a meaningful voice.

Why Rural Communities Belong at the Center

Rural young people should not be an afterthought in technology research. Distance, transportation, limited specialist services, small social networks, and concerns about privacy may make digital support especially valuable. Those same conditions may make an AI system harder to replace once it becomes the preferred listener or adviser. Telling a young person to "talk to someone" means little when the nearest qualified service is hours away, one counselor covers several schools, or disclosure carries a real social cost.

Rural communities also have strengths that technology research may overlook: intergenerational relationships, practical work, local knowledge, churches and civic institutions, and roles where responsibility has visible consequences. These communities can help us study whether AI points young people back toward competence, contribution, and dependable relationships. The goal is not to give rural youth less technology. It is to use technology without allowing it to replace belonging, judgment, and useful responsibility.

This is where workforce preparation and youth AI safety meet. Young people need more than the ability to operate an AI tool. They need to know when to use it, when to question it, when to disclose its role, when to seek human judgment, and how to remain responsible for the result. Those skills matter in a classroom, on a job site, in a clinic, in public service, and at home.

Implications for Practice and Governance

Product teams Evaluate multi-turn trajectories and dependency signals, not only prohibited outputs. Treat relational cues, memory, uncertainty disclosure, and handoff design as safety-critical features. Separate engagement goals from developmental outcomes.

Schools and youth organizations Create clear norms for AI-supported reflection and learning. Teach verification and decision ownership. Preserve access to adults who can carry responsibility that a system cannot. Include youth in governance with meaningful authority.

Families and caregivers Ask what role the tool is playing rather than relying only on screen time. Look for displacement of sleep, relationships, schoolwork, or independent judgment. Respond with curiosity and connection so disclosure does not feel like automatic punishment.

Researchers Build interdisciplinary teams that include developmental scientists, HCI researchers, safeguarding professionals, youth, and frontline practitioners. Report limitations plainly and distinguish exploratory constructs from validated measures.

Policymakers and funders Require child-centered impact assessment, independent evaluation, age-appropriate design, and credible human pathways. Fund community participation and long-term study, including rural and under-resourced populations.

The Role of Practitioner Knowledge

The co-rumination example taught me something I have not forgotten: a room full of capable researchers can still miss the variable that changes the work. Practitioners often see the contradictions first. A young person may appear calm but remain passive, feel connected but stay trapped, comply without being prepared, or explain the problem clearly without being able to act. Practice does not replace research. It helps research ask better questions.

An experienced youth practitioner should not be invited merely to tell a story, and should not claim clinical authority they do not possess. The useful role is disciplined framing. Does the study describe what actually happens? Could the safeguard work in a real school or community? Is the referral path believable? Is the product rewarding the wrong behavior? Does "engagement" mean growth, or simply more time with the machine?

That contribution should be built into the research from the beginning. Community investigators, practitioners, young people, and caregivers need defined roles and real influence before the research questions are fixed. Their insight is most valuable while the design can still be changed.

Limits and Boundaries

We should not outrun the evidence. Synthetic co-regulation, synthetic co-rumination, relational capture, responsible handoff, and the five agency questions are ideas to define and test. Existing studies are early and involve different products, populations, and methods. The evidence does not support one estimate of harm or benefit, and it does not justify treating every emotionally supportive use of AI as unhealthy.

This is a developmental framework, not a clinical tool. It should not be used to diagnose a young person or replace licensed care. It must not shift responsibility for product safety onto families or youth. Research involving minors and sensitive information requires qualified investigators, independent ethics review, strong data protections, and clear procedures for responding to disclosures of harm. Technology companies and institutions still carry duties that cannot be handed to the user.

One final limit matters. Calls for more human connection sound empty where trustworthy help is unavailable. Responsible design must improve the path to real support and strengthen the people and institutions expected to receive a handoff. Otherwise, we have only named the need without meeting it.

Conclusion

AI may help a young person settle down, rehearse a hard conversation, learn, create, or ask a question they are not yet ready to ask another person. Those are real possibilities worth developing. The danger is that the machine may become very good at keeping the conversation going while the young person becomes less able to leave it.

Co-regulation and co-rumination give us a better way to examine that risk. Warmth, trust, engagement, and disclosure are not automatically good outcomes. Their value depends on what they make possible next. Good assistance should strengthen a young person's ability to judge, choose, connect, accept appropriate responsibility, and preserve future options. When the system reaches the limit of its role, it should make human support and real-world action easier to reach.

The question I keep coming back to is simple: after the machine helps, is the young person more capable? A responsible system should leave them better prepared to move forward—not more dependent on being helped again.

Author Note

John Williams is the founder and executive director of Life That Counts. His work since 1999 has focused on adolescent development, peer influence, responsible agency, and the conditions that help young people convert freedom into durable responsibility. He participated in a Mathematica-convened think tank on adolescent co-regulation and has designed and led youth-development initiatives in community and federally supported settings. He writes here as a practitioner-framer, not as a clinician, psychometrician, or AI engineer.

Williams developed the central question, field interpretation, and conceptual framework in this article. Silas Generoproté, his AI digital assistant, supported literature discovery, source comparison, organization, and editorial drafting under his direction. Williams reviewed the work and accepts responsibility for the final text.