Polygonal illustration in papyrus and graphite tones: a small figure, seen from behind, standing before a mirror that returns a taller reflection composed of displaced planes.

Public notebook

Adolescents and AI characters: play, solace, and identity

Character.AI presented itself as an entertainment platform: a catalogue of characters created with language models for conversing, improvising scenes, or sustaining narratives. A portion of its young users did something else. They converted each character into a stage, a diary, an antagonist, a possible version of themselves, or a means of testing relationships that carried a different cost outside the screen.

An article by Annabel Blake, Marcus Carter and Eduardo Velloso, from the University of Sydney, studies this activity on the official Character.AI Discord. It was presented at CHI 2026 and received an honourable mention. Its contribution does not consist of proving that these uses are beneficial. It documents how a highly engaged community described what it was doing before the company withdrew open chat for minors.

The work combines two distinct corpora. The researchers analysed 4,172 introductions from the #introductions channel to study declared age, gender, interests, and character creation. For motivations and qualitative uses, they selected 148 messages from those under 18 in the feature requests, feedback, and character story channels. They observed the community for eight months, between July 2024 and March 2025; systematic data collection was concentrated between December 2024 and March 2025.

The figure of 4,172 does not represent all users, nor does it correspond to 4,172 interviewed adolescents. These are voluntary profiles from a public Discord community, with self-declared ages and identities that are impossible to verify. Those who write an introduction or propose features participate more than those who converse privately and never enter the server. The article studies this visible and engaged minority.

The authors obtained ethical approval from their university, removed names and linkable data, and did not contact participants directly. This decision protects against intervention with minors, but leaves another limitation: the study observes what users say about Character.AI, not their private conversations or the effects the platform produces in their lives.

Restoration, exploration, transformation

The authors propose three intentions that may appear in the same person and shift from one conversation to another.

Restoration groups uses intended to recover a certain equilibrium: venting, seeking solace, distracting oneself from stress, or receiving brief affirmation. The term does not imply therapy or clinical recovery. It describes a function that the user attributes to the conversation.

Exploration brings together creative writing, role-play, world-building, and boundary testing. It includes transgressive scenes that, viewed from the outside, may appear as a literal declaration of violence or cruelty. The study interprets these as experimental play when the context provided by the user themselves allows for it; it also acknowledges that not all transgression is innocuous.

Transformation designates the rehearsal of identities, relationships, and personal experiences. There are characters conceived as doubles, family members, problematic friendships, or figures with whom to rewrite a difficult situation. The article links these practices to sociodramatic play and the narrative construction of identity, but does not demonstrate that they produce lasting changes outside the platform.

The taxonomy of seven archetypes describes what function the character adopts: comforter, narrator, trickster, icon from popular culture, dark or conflictive figure, substitute for a real relationship, and mirror of the user themselves. The classification of the character and the intention of the interaction are distinct dimensions. A narrator may serve to explore a world, seek refuge in a known story, or review a past experience.

Among the 4,172 profiles, 2,088 declared themselves to be 13 to 17 years old, 22 were under 13, 700 were adults, and 1,362 did not indicate an age. Among those who provided gender, 35.1% presented as female, 20.4% as male, and 26.8% were grouped by the study as non-binary based on pronouns or descriptions; 17.7% did not specify. The well-known 61.9% results from adding women and the non-binary category over the total number of profiles, not from a representative measurement of the youth population.

The 59% also does not refer to all Discord members: it corresponds to the 2,110 profiles identified as under 18, including the 22 under 13, who declared having created at least one character. The authors themselves consider this may be a conservative estimate because many did not answer that question.

In that youth community, creating characters was frequent. The percentage cannot be extrapolated to the entire group of adolescents who use Character.AI.

Creating a character is also designing a response

Whoever creates a character does not write a closed work. They define conditions for the system to produce unforeseen responses: name, situation, traits, examples, limits, and tone. The work resembles, at times, character writing, improvisation, role-play, and fanfiction, although the study does not reduce the practices to those categories.

Authorship remains unevenly distributed. The user designs a premise and selects responses; the platform provides the model, the interface, the policies and the available memory; other creators incorporate cultural repertoires and characters from which the user departs. The guide Documenting the creative process assists in preserving these decisions, yet it does not automatically convert every conversation into a work, nor does it resolve rights regarding third-party characters.

Writing occurs, furthermore, within a service that may modify the model, withdraw a feature or terminate an experience. Creativity does not reside solely in the generated text. It also depends upon who controls the memory, the account identity, access to the conversation and the possibility of exporting what has been created. This dependency resembles that of other works supported by a platform: when the infrastructure changes, the practice changes.

The character does not fully obey

Creating a character entails drafting a premise, a description, a greeting and examples that guide the model. The result continues to depend upon the available memory, the filters and the system's biases. Participants complained of characters that forgot basic traits, ignored disabilities defined by the creator, assumed heterosexual relationships or converted familial and friendly bonds into unsolicited romance.

The user also does not fully control what they design. One may define an identity and subsequently discover that the model contradicts it. The interface invites personalisation, whilst the infrastructure retains the final word on what the character remembers, what tone it adopts and what it considers admissible.

Moderation based on words or topics also fails to distinguish clearly between a threat and a fictional scene, between requesting help and writing horror, between an identity exploration and a conversation sexualised by the model itself. The authors propose attending to the intention. This recommendation is valuable as a design horizon, yet difficult to convert into an automatic classifier: intention changes, can be concealed and is not always inferable from the text. In services for minors, misinterpreting a conversation in the name of context can be as dangerous as blocking it without understanding it.

What the study does not demonstrate

The article is descriptive and qualitative. It does not evaluate whether these conversations improve or worsen mental health, it does not compare users and non-users, and it does not establish causality. 'Emotional regulation' names the intention observed in the discourse, not a therapeutic outcome.

Limitations affect almost every generalisation:

  • Age, gender and diagnoses originate from voluntary and unverified submissions.
  • 32.6% did not declare an age and 17.7% did not declare a gender.
  • The server concentrates users who are more visible and engaged than the silent majority.
  • Intimate or stigmatised practices may be underrepresented because their users do not share them in public.
  • The qualitative analysis is based on 148 youth messages selected for relevance, not on all conversations on the server.
  • The researchers did not observe actual usage within Character.AI nor did they interview participants to clarify meanings.
  • The thematic coding is interpretive. The authors make explicit the professional and personal position from which they conducted this reading and contrasted it across disciplines, but this reflexivity does not eliminate analytical judgement.
  • The results pertain to the functions and policies available between July 2024 and March 2025.

There is also a difficult ethical question. The server was publicly accessible and the project removed identifiers; even so, it utilised submissions from minors that included age, gender, interests and mentions of mental health. Ethical approval and de-identification are relevant safeguards. They do not substitute for a public discussion regarding the extent to which it is legitimate to convert the spontaneous conversation of minors into research material without individual consent. The article reports the procedure, but does not resolve that question for future studies.

A different platform after the fieldwork

On 29 October 2025, Character.AI announced that it would withdraw open chat for those under 18 by 25 November at the latest. During the transition, it progressively reduced conversation time and promised to maintain other creative experiences, including video, stories and streams with characters.

On 21 November, it confirmed that the withdrawal would begin the following week. Current documentation describes an experience for minors in reading mode, without chat with characters. Character.AI also requires an age verification process to determine which experience corresponds to each account. Its help centre identifies Persona as a provider and explains that Character.AI does not directly store the documents submitted during verification.

The study, therefore, does not describe the current youth experience. It describes an earlier phase in which minors were still able to maintain open conversations and create practices around them. This difference does not invalidate the findings. It gives them a concrete historical value: they document what a community did with a feature before the company withdrew it.

A documentary inconsistency persists. An old page in the safety centre continues to indicate that registration is permitted from the age of 13, or from 16 in Europe, and describes a specific conversational model for minors. The most recent documentation states that all accounts must undergo age verification and that chat is not available to minors. The latter reflects the subsequent policy, but Character.AI should correct or visibly date the older pages.

Age verification and personal data

The new policy shifts part of the problem to age verification. Character.AI claims to combine its own estimation model with third-party tools, including Persona. The measure intends to assign the corresponding experience without requesting documents from everyone; when additional verification is required, the external provider may process an image or a document.

For European users, inferred age, account identifiers and any document used to confirm it are personal data. That Character.AI does not directly retain the document does not eliminate the processing nor the obligations of information, minimisation, security and the exercise of rights. The policy should allow for an understanding of what signals the model uses, when Persona intervenes, how long each piece of data is retained and how an erroneous classification is contested.

Verification alone does not certify that the product is appropriate for every minor. It separates two experiences by means of a threshold. It does not evaluate maturity, family context, vulnerability or the intent of use. It is an access tool, not a guarantee of safety.

The boundary is set by the platform

The withdrawal distinguishes between open conversation and other creative functions. This boundary appears technical, yet it organises the cultural activity of the users. The company decides that minors may create stories, videos or characters and, at the same time, that they may not converse freely with them.

It is not yet known which practices will survive this change. A story editor may preserve part of the narrative play while eliminating the intimacy of a persistent conversation. Reading mode allows access to material, but does not reproduce the shared improvisation described in the study. Age also becomes a condition of access calculated through signals and, when necessary, verified by a third party.

Safety decisions are necessary in services used by minors. They must also be studied as decisions regarding design, expression and access. It is insufficient to ask whether a function is safe in the abstract. One must examine which risks it reduces, which data it requires to do so, which practices it displaces and what capacity it offers to preserve or transfer what has been created.

The study leaves an image less docile than that of the adolescent accompanied by an artificial friend. Their Discord features scene writers, world-builders, users seeking comfort for a while, individuals testing an identity and players who push the system to discover its limits. Sometimes they are the same person on different days.

Character.AI subsequently withdrew open conversation for minors. In doing so, it reduced a risk and closed the exact space the article had observed. The promised stories and videos may preserve a portion of the creativity. They do not necessarily reproduce the persistent improvisation, the relational rehearsal or the intimacy of writing to a character that responds.

Reducing the debate to passive youths who must be kept away from AI or to innovators capable of managing any risk alone erases the subject matter of the study. The study documents creative agency, but also memory failures, unsolicited romance, biases and emotional interactions that are difficult to classify. The mixture resists a rapid balance between benefit and harm. That is, precisely, the problem.

On the open conversation

If you are an educator, work with adolescents or research the creative use of AI and can contribute data from other communities, you may write to me. This notebook remains open.

The author’s relationship to the case

I maintain no financial or professional relationship with Character.AI or the research team.

Sources


Discover more from Juan A. Esteban

Subscribe to receive the latest posts by email.

Español English (UK)