A teenager might open a chatbot to get help with an assignment, then continue the conversation about a difficult friendship, an anxious evening or a decision about their future. The software has not changed. But the role it is playing in that young person’s life has.

This possibility deserves more attention than it receives in education debates dominated by cheating, productivity and prompting. Young people’s relationships with AI cannot be understood solely by asking whether it helps them complete schoolwork. We also need to ask what they trust it with, why they turn to it, and whether it strengthens or weakens their capacity to think and act independently.

Research at University College London (UCL) is taking these questions seriously. Its project, Adolescence, Chatbots and AI Intimacies, brings together psychology, sociology and education, with Professor Jessica Ringrose and Dr Marc Tibber as its academic leads. It seeks to understand young people’s experiences and develop practical support for schools and families (University College London, n.d.-a).

There is an important distinction about publication status. UCL’s report launch, Adolescents and AI Chatbots: Use, Dependency and Critical AI Literacies, is scheduled for 10 December 2026. As of this article’s date, the full report and detailed findings have not been located publicly. What follows draws on the project’s published approach, a preliminary result shared by Ringrose, and supporting published research—not on a report that has already been released (University College London, n.d.-b).

Schoolwork and companionship are not separate worlds

In a preliminary public update, Ringrose reported that 82% of teens in the team’s recent research with nearly 2,000 adolescents in UK secondary schools were using AI chatbots for schoolwork or homework. She also noted that this blended with other types of use (Ringrose, 2026).

That is a striking indication of how familiar these tools have become within the study sample. It is not, however, a national prevalence estimate. Nor does it mean that 82% are dependent on AI. The full report will be needed to assess the sampling, question wording and relationships between different forms of use.

The overlap between purposes is the more interesting educational issue. Schools may classify a chatbot as a learning tool, while a young person experiences it as an explainer, adviser and conversational companion. General-purpose chatbots can be used for companionship even when they are not marketed as companions. Systems specifically designed to simulate friendship or care introduce further questions about emotional attachment and commercial incentives (Witting et al., 2026).

Published UK evidence makes this more than a hypothetical concern. Internet Matters’ 2025 study surveyed 1,000 children aged 9–17 and included focus groups with teenage chatbot users. It found that 64% of the surveyed children had used an AI chatbot. Among chatbot users, 35% said talking to one felt like talking to a friend; 15% preferred talking to a chatbot rather than a real person; and 12% said they used one because they had no one else to speak to (Internet Matters, 2025).

These figures cover AI chatbots broadly, not just companion platforms. They describe reported experiences and motives, not clinical dependency. They also raise a question that education alone cannot answer: if a child turns to software because there is nobody else to talk to, what human support is missing?

US research offers a complementary perspective, not a directly comparable estimate. In Pew Research Center’s survey of 1,458 teenagers aged 13–17, 12% of all teens reported having used chatbots for emotional support or advice. Among chatbot users, the figure was 18% (McClain et al., 2026). The difference in denominators matters. So does the difference between seeking support and becoming dependent.

Dependency needs more careful language

The word “dependency” can obscure as much as it explains. Frequent use, trusting an answer, enjoying a conversation, emotional attachment and clinical addiction are not interchangeable.

A student who regularly asks for explanations may be using an effective study aid. Another may obtain completed answers without developing the understanding needed to assess them. A third may seek reassurance repeatedly or avoid approaching people who could provide appropriate support. Counting sessions would not tell us enough to distinguish these situations.

For educators, the useful questions concern function and consequences. Can the learner explain the work without the chatbot? Can they disagree with its advice? Can they choose not to use it? Is the interaction supplementing relationships, or taking the place of support they need from people?

UCL’s project is valuable because its published approach explicitly moves beyond simple warnings about “technology addiction”. It combines attention to wellbeing and thinking styles with the social context of use. It also proposes creative and participatory methods through which adolescents can explain experiences that a usage count might miss (University College London, n.d.-a).

The available evidence still requires caution. An association between loneliness and chatbot use would not establish which came first. Young people might seek chatbots because they already lack support; particular interactions might also worsen difficulties. Both possibilities require investigation. UNICEF’s policy brief calls for further research, including longitudinal studies, on benefits and harms (Witting et al., 2026).

We should take risks seriously without turning every young chatbot user into a diagnosis.

Critical literacy is more than learning to prompt

Much AI training starts with the question: how can we get a better answer from the system? Critical AI literacy asks an earlier question: should we use this system for this purpose at all?

UNESCO’s student competency framework places a human-centred mindset and AI ethics alongside technical knowledge and system design. Its twelve competencies extend beyond operating a tool to retaining human agency and evaluating AI responsibly (Miao et al., 2024).

A separate UCL IOE vision for teacher education makes a related distinction. Polly Glegg argues for critical digital literacy that includes ethical evaluation, engagement with research, developing learners’ judgement and questioning policy. It treats deciding when not to use generative AI as part of professional competence, rather than as resistance to innovation (Glegg, 2026).

Applied to adolescents’ chatbot use, this means examining several kinds of trust.

Trust in information requires more than asking whether an answer sounds convincing. Learners need to examine evidence, check sources and distinguish an explanation from a verified claim. A fluent answer about an apprenticeship, qualification or health concern can still need checking against an authoritative source.

Trust in a relationship requires recognising the difference between supportive language and human understanding. A chatbot may simulate empathy or remember details, but that does not give it the responsibilities, mutuality or accountability of a friend, teacher or counsellor. UNICEF identifies these simulated relationships as a particular concern for child rights and safety (Witting et al., 2026).

Trust in the service requires asking who operates it and what happens to a conversation. What personal information is being disclosed? What do the privacy terms permit? What commercial interests might shape the experience? Some companion services face a tension between protecting users and keeping them engaged. That is a design and governance issue, not simply a failure of individual self-control (Witting et al., 2026).

Critical literacy therefore includes the capacity to question an answer, recognise an inappropriate relationship claim and decide that a conversation belongs with a person instead.

From critical literacy to educational practice: what CoDiCri adds

Defining critical literacy is only the beginning. The harder question is how to design learning that gives students opportunities to exercise it. Telling learners to question AI does not necessarily create the conditions in which they can do so.

The CoDiCri Critical Digital Competence Handbook offers a useful contribution here. Developed for university educators, it brings together a pedagogical framework, a conceptual framework and proficiency indicators, the CoDiCri ACAD Toolkit, and analyses of implemented open educational practices. Critical competence is developed alongside disciplinary learning, rather than treated as an additional topic detached from it (CoDiCri, 2026).

Its ACAD approach—Activity-Centred Analysis and Design—asks educators to consider three connected aspects of a learning situation: the intellectual tasks; the tools and resources; and the organisation of roles, relationships and participation. Cards and a canvas make these choices visible and discussable. The toolkit can also be used after teaching to reconsider what happened, rather than merely document what was planned (Castañeda et al., 2026).

That matters for agency. If a teacher supplies every question, chooses every source and rewards agreement with a predetermined conclusion, students have little opportunity to exercise independent judgement. Critical content alone does not make a learning activity participatory.

One handbook example is directly relevant to AI. In initial teacher education, students analysed generative-AI applications, examined their ethical and educational implications, designed a learning activity, and undertook self- or peer assessment. The analysis of student work found that critical digital competence remained mainly at the introductory level. Staff valued the structured sequence of investigation and design, but reported difficulties in fostering deeper critical engagement and sharing, with continued guidance needed (CoDiCri, 2026, pp. 28–30).

This is not a failure to hide. It is a useful finding: a well-designed activity can begin the work without completing it. CoDiCri presents its examples as situated experiences to analyse and adapt, not perfect models to reproduce. The handbook’s progression indicators distinguish supported basic understanding, more advanced independent use, and the capacity to facilitate or develop practice with others. They provide a reference for planning and evaluation, not a fixed sequence or a universal test (CoDiCri, 2026, pp. 18, 22).

These are university-based experiences, not evidence that the toolkit prevents adolescent chatbot dependency. Their relevance to UCL’s agenda is pedagogical: they help turn an aspiration such as “retain your judgement” into learning opportunities, and invite educators to examine whether those opportunities actually materialised.

What this could mean in schools and VET

The UCL launch programme includes a new lesson plan on critical AI literacies through its collaboration with the PSHE Association. Its effectiveness is not yet established by the listing (University College London, n.d.-b). CoDiCri suggests useful questions for designing and reviewing such teaching, rather than supplying a ready-made adolescent intervention.

For example, learners could examine fictional, teacher-prepared chatbot conversations about an assignment or a training choice. The intellectual task would be to identify claims, check evidence and justify a decision. Resources would include the prepared conversations and authoritative information, not compulsory companion accounts. The social design could give peers distinct roles in questioning evidence, while leaving room for disagreement and teacher support. Nobody would need to disclose personal difficulties to a chatbot.

Learners might then co-create a short guide for other students. Open educational practice would involve negotiating what to share, how to credit sources and how others could reuse the work. With adolescents, openness must not become compulsory public exposure: private, pseudonymous or collective contributions should remain possible.

Afterwards, teachers would examine what students actually did. Did they check claims or merely repeat warnings? Could they explain their reasoning independently? Did less-confident learners have a voice? These are proposed adaptations, not findings from either project. They would need age-appropriate safeguards and evaluation.

For adolescent learners in vocational education and training (VET), the same approach could be embedded in occupational tasks involving confidentiality, verification and professional responsibility. It would build practical knowledge and critical judgement together. Teachers also need support to make these choices; Glegg’s vision rightly treats critical literacy as a responsibility for teacher education, not something staff should acquire through unsupported experimentation (Glegg, 2026).

Careers guidance must preserve the young person’s judgement

The overlap between educational and personal chatbot use also matters for careers guidance. A question about course entry requirements can quickly become a conversation about identity, confidence or whether a young person is “suited” to an occupation.

A chatbot might help a learner prepare questions for an adviser or organise information about training options. But a plausible recommendation is not automatically reliable, current or appropriate to that individual. Eligibility, funding and qualification recognition need checking against official sources. Decisions about aspirations need space for reflection and discussion, not just a confident answer.

The aim should be to strengthen the young person’s judgement. Can they explain their options, recognise what remains uncertain and question the recommendation? An interaction that produces an apparently clear choice without that understanding may be efficient, but it is not necessarily good guidance.

This is an application of the research agenda, not evidence that the UCL team has tested careers-guidance chatbots. It shows why critical AI literacy belongs in conversations about agency and informed choice, as well as online safety.

Literacy cannot compensate for unsafe design

There is a danger in making “better literacy” the answer to every problem. It can shift responsibility from providers and institutions onto young people: if harm occurs, the user supposedly failed to be sufficiently critical.

That would miss an essential part of the issue. Children should not have to understand every commercial incentive or technical failure before a product is safe enough for them to encounter.

UNICEF calls for safeguards that include transparency about systems’ non-human nature, age-appropriate access controls, independent testing, child-rights impact assessments and appropriate responses when a child discloses risk. It also challenges misleading marketing of chatbots as human or therapeutic substitutes (Witting et al., 2026).

Schools and families have responsibilities, but so do designers, providers and regulators. A literacy lesson cannot make an inappropriate product appropriate. Nor can it supply the human relationship that a young person lacks.

The most important promise of UCL’s work is its attention to young people’s lived experiences. It starts with what adolescents are doing and feeling, rather than assuming that adult categories capture their use. The forthcoming report should help clarify those experiences and the support they require.

For now, the educational principle is already clear: AI literacy should help young people remain authors of their own decisions, not merely more effective users of a chatbot.

References

Castañeda, L., Marín, V. I., Tur, G., Peguera-Carré, M. C., Orellana-Hernández, M. L., & Villagrá, S. (2026). Designing for openness, critical digital competence and agency: The development of the CoDiCri ACAD toolkit. European Journal of Open, Distance and E-Learning, 28(1), Article 6. https://doi.org/10.65043/eurodl.179

CoDiCri. (2026). Critical digital competence: User guide / handbook (Version 1). Zenodo. https://doi.org/10.5281/zenodo.22869113

Glegg, P. (2026, March). IOE’s vision for critical digital literacy in teacher education in an age of generative artificial intelligence. UCL Institute of Education. https://discovery.ucl.ac.uk/id/eprint/10223409/1/Final-version2.pdf

Internet Matters. (2025, July). Me, myself and AI: Understanding and safeguarding children’s use of AI chatbots. https://www.internetmatters.org/wp-content/uploads/2025/07/Me-Myself-AI-Report.pdf

McClain, C., Anderson, M., Sidoti, O., & Bishop, W. (2026, February 24). How teens use and view AI. Pew Research Center. https://www.pewresearch.org/internet/2026/02/24/how-teens-use-and-view-ai/

Miao, F., Shiohira, K., & Lao, N. (2024). AI competency framework for students. UNESCO. https://doi.org/10.54675/JKJB9835

Ringrose, J. (2026). In our recent research with nearly 2K adolescents in UK secondary schools, we found that 82% of teens were using [Post]. LinkedIn. https://www.linkedin.com/posts/jessica-ringrose-4666a650_parent-registration-activity-7520769994176262144-UCzg

University College London. (n.d.-a). Adolescence, chatbots and AI intimacies. Retrieved November 3, 2026, from https://www.ucl.ac.uk/research-innovation/case-studies/2025/nov/adolescence-chatbots-and-ai-intimacies

University College London. (n.d.-b). Adolescents and AI chatbots: Use, dependency and critical AI literacies [Event listing]. Eventbrite. Retrieved November 3, 2026, from https://www.eventbrite.com/e/adolescents-and-ai-chatbots-use-dependency-and-critical-ai-literacies-tickets-2004063383528

Witting, S., Day, E., Galea Baron, J., Kaviani Johnson, A., & Vosloo, S. (2026, June). When AI becomes a friend: Child rights risks, harms, and regulatory responses to AI chatbots and companions [Policy brief]. UNICEF. https://www.unicef.org/media/181131/file/UNICEF-When-AI-becomes-friend-policy-brief-2026.pdf

About the Image

The AI race describes the companies and nations which are racing forward to the newest iteration of ‘AI’, including software, hardware capacity, power, leadership, and influence in the field. It’s a race that nation leaders and private corporations have signed up for, but workforces, neighbourhoods, communities, and working class people are impacted by the harms caused. I strive to show the implications of technology, in this case, the runner represents the workers running the race for entire nations, and its silhouette contains the materials, hardware, and resources (often hidden) needed to build generative AI technologies. Skylines are representative of cities and progress, but the heavy smoke cruising along the image representing the pollution, e-waste, and debris the race leaves behind. The panels of colour represent the countries typically referred to in the context of the “AI race”: the US and China. But I leave this open to interpretation and did not want the flags and countries to distract from the most important aspect of the illustration: the consequences of the AI race.