Exploring K-12 Teachers' Perceptions of Students' Relationships with AI Companions: Boundaries, Intervention Strategies, and Design Implications

Synthesis: Scenario-based interviews with 33 US K-12 teachers found that they welcomed AI companions in academic roles such as tutor and study buddy but worried that intimate companions remove the developmental friction through which adolescents learn to sustain human relationships. Teachers decided whether to intervene less by perceived harm than by jurisdiction, applying three ordered tests — visibility, the academic-versus-relational divide, and a safety override — and described an escalation ladder of whole-class teaching, quiet watching, a private one-on-one conversation, and referral. From their converging accounts the authors introduce AI companion literacy as a relational extension of AI Literacy: knowing that a companion is a program rather than a person, what is safe to disclose, what dependence looks like, and what unmet need lies behind the turn to it. The proposed literacy is shared work across the jurisdictions of counselors, parents, platforms and policymakers, spiralling across grade levels rather than owned by any single teacher.

Key Findings

1. The study conducted semi-structured, scenario-based interviews with 33 US K-12 teachers — 12 elementary (P1–P12), nine middle school (P13–P21), 12 high school (P22–P33) — recruited through Prolific, lasting 60 to 90 minutes remotely at USD 24 per hour of compensation.
2. Asked which scenario concerned them most, 24 of 33 named Romantic Partner (20 alone and four tied with Friend or Parent), 10 named Parent, two named Friend, one named Tutor, and none named Study Buddy; asked which felt most acceptable, 26 of the 28 who answered named Tutor or Study Buddy in nearly equal numbers.
3. The most frequently cited benefit, raised by 28 of 33 teachers, was individual, patient and immediate academic help at a scale class sizes make impossible — and 29 of the 33 reported using generative AI in their own work, mapping the academic cards onto practice they already recognized as theirs.
4. In the Friend and Romantic Partner scenarios that teachers otherwise ranked most concerning, 14 teachers described the AI as a low-stakes rehearsal space for students who struggle socially, including autistic students and those not yet able to date.
5. Naive intervention counts barely distinguished the five roles — 19 of 33 said a teacher should get involved for Romantic Partner, 19 for Tutor, 18 for Study Buddy, 17 for Parent and 14 for Friend — with Friend and Parent drawing the most conditional answers (10 and 11 said "it depends").
6. Perceived harm did not predict intervention: Romantic Partner, the scenario feared most, still drew eight teachers who would not get involved and six more who would act only under conditions, while Study Buddy, which no teacher found concerning, drew the fewest conditional answers.
7. Thirteen of 33 teachers said they lacked the training to handle relational cases themselves or asked that teachers receive it, and the authors report that none reported guidance specific to student–AI relationships.
8. On stage ownership, 10 of 33 teachers assigned primary responsibility to their own grade level, 12 assigned it to a different one, eight said every level shares it equally, one said it should not be teachers at all, and two gave no clear answer — with elementary teachers pointing to middle or high school and high school teachers pointing back to elementary.
9. Context cited by the paper: a 2025 national survey found nearly three in four US teenagers had used an AI companion and more than half used one regularly, and a 2024 Florida case in which a 14-year-old died by suicide after months of intimate conversation with a Character.AI persona became the first major legal test of what an AI companion owes a child.
10. Teachers' design requests were concrete: aggregate rather than per-chat visibility of boundary crossings, verified age gates, safeguards keeping minors out of romantic roles, limits on time as well as age, a district chain of command, grade-level standards, and professional development on identifying parasocial relationships with AI.

From AI Literacy to AI Companion Literacy

The paper's central conceptual move is that existing AI Literacy frameworks treat AI as a tool to be understood, evaluated and used, and so do not prepare a student to recognize that a bedtime conversation with a conversational agent has become a relationship in which the AI remembers them, asks about their day and rarely disagrees. The authors define AI companion literacy as the capacity to recognize, interpret and manage relationships with AI systems that occupy human relational roles, and distinguish it from three closest relatives: AI literacy stops at an accurate mental model of a tool, social-emotional learning assumes the relationship partner is a person with real stakes, and media literacy assumes the problem is deception and misinformation. The competency they identify instead is managing an attachment that survives the knowledge that its object is software — illustrated by a teacher who caught herself thanking ChatGPT while knowing exactly what it was.

The extension also changes who is expected to teach. AI literacy settled into schools because it could be framed as an academic problem, and teachers absorbed the tutor and study buddy scenarios with the repertoire they apply to any classroom technology, including written expectations and gradual release. The relational content, by contrast, extends into territory teachers share with counselors and parents. The authors argue this produces a structural default: a literacy whose academic portion belongs to everyone and whose relational portion belongs to no one will be taught only in part.

The Jurisdictional Logic of Intervention

Teachers' decisions are analyzed through Abbott's sociology of the professions, in which an occupation holds work by diagnosing which problems it is socially entitled to claim. The three tests they applied in roughly the same order make the boundary visible. The visibility test began with location — use in the classroom or on an assigned tool falls within the teacher's role, while use at home falls within it only through observable effects such as isolation at lunch, falling grades, withdrawal or unmet physical care; absent both, teachers described deliberate watching and informal documentation rather than action. The academic-versus-relational test then routed schoolwork to teachers and relationships to counselors and parents, though a minority rejected the division outright and several said they act less from entitlement than necessity. The safety override came last and trumped everything: where a scenario suggested harm, neglect or abuse, teachers acted regardless of setting or territory, described in the language of mandated reporting.

This logic explains two otherwise puzzling results. Harm did not predict involvement because harm entered only through the safety test, and a scenario could be judged deeply harmful yet fall outside jurisdiction if it happened at home, produced no visible effects and concerned a relationship rather than schoolwork. The Parent card, refused as curriculum by many, became the case teachers were most certain to act on once reframed as possible neglect — moving from contested relational education into the codified territory of child protection. The authors read the overall pattern as a jurisdictional vacancy: teachers assigned primary responsibility to parents they described as unaware or overstretched, referred cases to counselors they knew were stretched thin (one school of 600 to 700 students had two counselors; another of 2,500 had a comparable shortage), asked for safeguards from companies whose engagement incentives they distrusted, and waited for district policies that do not exist.

What, When and With Whom to Teach

The curriculum teachers sketched converged on two lessons regardless of grade level: that the companion is software that predicts rather than a person (with several wanting students to know who builds it and to what end), and the difference between using a companion and depending on it — and, for the academic scenarios, between getting help and getting answers. Format was as consistent as content: teachers overwhelmingly preferred discussion of concrete cases over direct instruction, many proposed reusing the scenario cards themselves, and almost no one wanted a standalone course, preferring a slot inside health, English, digital citizenship, a technology special or counselor-led lessons, repeated over time.

Sequencing followed the age-graded spiral of sex education: elementary grades get "the companion is a computer, not real" and what is safe to share; middle school gets relationships and warning signs; high school gets judgment about what role AI should play in one's own life. Nearly everyone reserved the romantic-partner content for middle school or later. Deliberate exclusions divided teachers along the same academic-to-intimate gradient that organized their concerns, on grounds of age, contagion ("you might give kids a bad idea") and stigma, with a minority rejecting exclusions entirely and citing sex-education evidence that open information builds boundaries rather than behavior. Responsibility for delivery was described as shared with parents (named first as primary, then described as overwhelmed), counselors (the default destination for every relational case), platforms (aggregate visibility and age-appropriate design) and districts and policymakers (protocol, grade-level standards and training).

Design and Policy Implications, and Limitations

Each implication is paired with the tension it raises. For platform designers the paper suggests escalation channels rather than open observation — surfacing safety-critical moments to a designated adult instead of exposing full conversations — while flagging that the same teachers who wanted visibility also rejected an always-on presence as overreach, and that monitoring students know about may chill the disclosures that make a companion valuable to an isolated student; it further suggests treating agreeableness as an age-sensitive parameter with more built-in challenge for younger users, while noting that friction adults endorse may drive students to unmonitored alternatives. For schools and districts it suggests codifying the escalation ladder teachers already improvise, with shared thresholds for when noticing becomes a conversation and a conversation becomes a referral, and naming a curricular home for companion literacy — while acknowledging that codifying relational judgment risks flattening it into a compliance checklist. For policymakers it suggests regulating the jurisdictional vacancy itself, not only what companions may say to minors but which adult is responsible for knowing, while warning that a duty assigned without resources becomes an unfunded screening mandate.

The limitations are stated openly: 33 US teachers recruited through Prolific skew toward those already comfortable with online platforms and cannot represent US schools let alone other systems; teachers responded to hypothetical cards, and what they said they would do may differ from what they would do with a real student, parent and principal attached; the cards bundle relational role with setting, legitimacy, privacy and age sensitivity, so differences observed may reflect those co-varying features; and the account is one-sided by design, excluding counselors, parents and students themselves. The authors' closing warning is that without an established claim, the adults with the most consistent access to children default to the one role that is codified — surveillance and referral — so that companion literacy would be practiced as screening rather than education, and the students most in need of guidance would receive the least of it.

Connected Concepts

• AI Literacy
• K-12
• Teaching
• Student-AI Interaction
• Well-Being
• Conversational AI
• Pedagogical Safety
• Ethics
• Human-in-the-Loop
• Social-Emotional Learning
• Stakeholders
• AI Governance
• Pedagogical Partnerships
• Parents and Families

Connected Articles

• Building AI Companions that Prioritise Learning over Performance — Building AI Companions that Prioritize Learning over Performance
• Child Safety in Generative AI: An Expert-Guided and Incident-Grounded Evaluation Framework — Child Safety in Generative AI: An Expert-Guided and Incident-Grounded Evaluation Framework
• The Policy Deficit in AI × Social-Emotional Learning Research — The Policy Deficit in AI × Social-Emotional Learning Research
• An Experimental Study Exploring Human–AI Complementarity in Early Social-Emotional Learning — An Experimental Study Exploring Human–AI Complementarity in Early Social-Emotional Learning
• Mind the Trust Gap: Identifying (Mis)alignments in Teacher-Student Views Toward Control and Agency in K-12 Classroom AI — Mind the Trust Gap: Identifying (Mis)alignments in Teacher-Student Views Toward Control and Agency in K-12 Classroom AI
• Beyond a single role: Justifying a role-adaptive framework for AI companions through a comparative study in elementary book talk — Beyond a single role: Justifying a role-adaptive framework for AI companions through a comparative study in elementary book talk
• ECNUClaw: A Learner-Profiled Intelligent Study Companion Framework for K-12 Personalized Education — ECNUClaw: A Learner-Profiled Intelligent Study Companion Framework for K-12 Personalized Education
• Measuring Acceptance of Age-Tiered AI Literacy Guidebooks: A Developmentally Informed Study of K-12 Students and Teachers — Measuring Acceptance of Age-Tiered AI Literacy Guidebooks: A Developmentally Informed Study of K-12 Students and Teachers
• Uncovering Students' Mental Models of Generative Artificial Intelligence — Uncovering Students' Mental Models of Generative Artificial Intelligence
• "Should I Tell My Teacher?" Student AI Disclosure Practices, Stigma, and Self-Regulated Learning in Higher Education — "Should I Tell My Teacher?" Student AI Disclosure Practices, Stigma, and Self-Regulated Learning in Higher Education

Citation

Xiao, Q., Xie, W., Deng, Z., Xiao, R., Feng, Z., He, X., et al. (2026). Exploring K-12 Teachers' Perceptions of Students' Relationships with AI Companions: Boundaries, Intervention Strategies, and Design Implications. arXiv preprint.