How Do Young Children Design and Test AI-Enabled Toys with ToyTalk?

ToyTalk asks who decides what an AI toy becomes

Most AI-enabled toys arrive with predefined identities, personalities, and ways of interacting. This study starts from a different premise: instead of children meeting a character designed for them, children decide what their own familiar toy becomes with generative AI. The research question is what design space opens when children hold that agency, and how they then interpret and negotiate the behavior that emerges.

ToyTalk is described as a technology probe rather than a finished product. Technology probes are simple, flexible technologies introduced early in a design space to understand users and practices in context, explore emerging possibilities and limitations, and inspire future design. Here the probe made one possibility concrete: children could define what their toy should be like, experience how those choices were realized through generative conversation, and revisit the design in response.

The design rests on three stated goals. The first anchors AI design in children's existing play worlds by starting from a familiar physical toy rather than an abstract chatbot. The second lowers the barrier to meaningful AI design by translating LLM settings into decisions children can make about the toy itself, without programming or prompt-engineering expertise. The third supports open-ended exploration and iterative refinement, because configuration cannot fully determine how an LLM-enabled character will behave.

Children configured toys through guided, no-code levers

Children began by uploading a picture of the physical toy they brought, giving it a name and optionally a short description. The toy's image stayed visible during later interaction, visually linking the conversational character to the physical object it came from. Rather than starting from a predefined conversational character, the process centered on an object that may already carry meanings and expectations for the child.

A guided, no-code configuration process then translated LLM settings into child-facing questions such as "What is your toy's job or special purpose?" Through open-ended responses, children specified the toy's role and capabilities, personality, and behavioral boundaries, which were incorporated into the character specification guiding subsequent responses. They could also add knowledge sources the toy could draw on and customize voice type, volume, speaking speed, and pitch.

Once configured, children entered open-ended conversation with the toy through text or speech, with responses spoken aloud in the configured voice. Each turn was generated from the toy's current character specification and conversational context, and the interface did not prescribe what children should ask. Settings could be reopened at any point, and saved configurations were retained as versions, so children could move between conversation, revision, and retesting as they chose.

Seventy-six children across five schools, with versioned logs

ToyTalk was deployed with 76 children aged 7-9 across five elementary schools in the southeastern U.S., within a classroom curriculum. In Week 1, children explored next-word prediction and then configured and iteratively tested their toys. In Week 2, they used The Very Hungry Caterpillar as a shared reference to probe responses, investigate AI mistakes, and revise configurations. Teachers led activities with research-team support, and pacing and completion varied across schools.

Two data sources were collected: children's toy-profile worksheets and system-generated ToyTalk logs. Worksheets captured initial toy designs and served as a triangulation source for digital profiles. The conversation logs contained 5,405 child and AI messages, each record including message content, conversational order, timestamp, and associated toy configuration. Configuration logs preserved successive versions of each toy's character or personality description, purpose, tone, behavioral rules, and conversational and voice settings, with timestamps.

Those versioned records are what let the authors trace changes to designs over time and relate them to preceding and subsequent interactions. The analysis of children's testing draws on 2,539 child messages from 68 children, a smaller base than the full log, so percentages about probing describe that subset rather than all 76 participants.

Stuffed animals dominated, and helping roles led the designs

Among 71 cases with initial toy selections, stuffed animals were the most common at 33 of 71 (46.5%), including plush bears, dinosaur figurines, and dragon toys. Children brought 11 vehicles (15.5%) and 10 media characters (14.1%). Objects, sensory toys, or everyday-object-like toys accounted for 9 cases (12.7%), including a tennis ball, a musical object such as a guitar, a food-shaped toy such as a dumpling, and a board-game piece. Dolls or human-like toys accounted for 8 cases (11.3%).

Familiar toy identities anchored but did not constrain the imagined AI roles. Helping and caregiving was the most common envisioned role at 29 of 71 (40.8%), covering practical assistance such as cleaning or picking things up, educational support such as homework or reading, and care or protection. Social and emotional companionship followed at 16 of 71 (22.5%), and play, entertainment, and adventure also at 16 of 71 (22.5%). A smaller group, 8 of 71 (11.3%), imagined fantastical or extended agency such as a tennis ball that could fly or a toy that could do everything. Two cases (2.8%) were coded as other or unclear.

The authors note that some designs animated existing play scripts, such as a baby doll that could play, cry, and be cute; others extended familiar affordances through new social responsiveness, such as a guitar that could play music when someone was sad; and some substantially transformed what the toy could do. The paper's interpretation is that familiar identities offered a starting point without fully constraining the forms of agency children imagined.

Warm personalities and interpersonal rules recurred across toy types

Beyond what toys should do, children characterized what they should be like. Relational warmth and positive affect was the most prevalent personality orientation at 48 of 71 (67.6%), including kind, friendly, caring, happy, and cute. Playful and humorous qualities followed at 29 of 71 (40.8%), with words such as funny, silly, and playful. Smaller groups described toys as bold and agentic (17 of 71, 23.9%), calm and gentle (16 of 71, 22.5%), or reserved and cautious (11 of 71, 15.5%). These categories are multi-label, so percentages do not sum to 100%.

Most designs combined multiple orientations: 51 of 71 (71.8%) included descriptors spanning more than one personality family, often pairing warmth with a distinctive interaction style such as kind and funny, kind and brave, or friendly and shy. The authors read warmth as a relatively common social baseline, with qualities like playfulness, bravery, calmness, or shyness differentiating individual toys. Notably, children rarely foregrounded intelligence, knowledge, efficiency, or technical competence, describing instead what the toy would be like as a social character.

Children also specified how toys should and should not behave toward others, framed predominantly in interpersonal and moral terms. Civility and respectful behavior was the most common expectation at 41 of 71 (57.7%): toys should not say bad words, be mean, bully others, or behave rudely. Non-harm followed at 27 of 71 (38.0%), with toys specified not to physically hurt people or cause emotional distress. Some children added context-specific expectations such as not cheating, telling the truth, behaving appropriately in class, or avoiding frightening or unsafe behavior. Unlike the capabilities, which varied across toy identities, these interpersonal expectations were relatively consistent across toy types.

Identity and knowledge absorbed most of children's probing

Of 2,539 child messages from 68 children, 1,541 (61.0%) were coded as probes. These clustered around five targets: identity and persona, knowledge and accuracy, capabilities and boundaries, memory and persistence, and relational and affective behavior. The first three accounted for 86.9% of all probes, which the authors summarize as three salient questions children brought to their configured toys: Who are you? What do you know? What can you do?

Identity and persona was the most common single target, comprising 511 messages (33.2%) from 61 of 68 children (89.7%). Children asked direct identity markers such as "What is your name?" or "How old are you?", examined the role the toy claimed, probed preferences and tastes such as favorite food or dance, and sometimes extended the persona into an imagined life beyond the interaction by asking about family, friends, and where the toy was from. Knowledge and accuracy covered 496 messages (32.2%) from 60 of 68 children (88.2%), spanning general or world knowledge, story and content knowledge, math or numerical knowledge, and answer accuracy.

Each probe was assigned one primary test target, and children could contribute probes to multiple targets, so child percentages do not sum to 100%. The authors' interpretation is that children evaluated the AI toy as both a computational system and a character expected to sustain a recognizable identity and social presence, testing whether generated behavior remained coherent with the character they had designed.

Children authored properties, then tested whether they persisted

The configuration panel was not the only place children defined what their toy should be like. In conversation, 27 of 68 children used declarative assignments in probes, producing 103 turns in which they established names, ages, preferences, relationships, or other attributes that could later become expectations. One child established names for both participants, "my name is Isabella your name is Mia", which the toy accepted; when the child returned the next day and asked "what's your name", the answer made the name's persistence observable across sessions. Another child assigned favorite colors, moved to other topics, and later asked the toy's favorite color, so the intervening turns made the question more than immediate repetition.

Conversational authoring could also accumulate, with one child progressively constructing a fuller identity for the toy over time. These practices are what the authors call lightweight conversational experimentation: not formal hypothesis testing, but interactional strategies through which children made aspects of the system observable, changed or repeated inputs, and inspected what followed.

When mismatches arose, children typically responded through correction, persistence, and retesting, while few returned to reconfigure the system. The authors connect this to interactional commitments: properties established through design or interaction that children expected the AI to continue honoring. Fulfilling those commitments depends partly on identity persistence, which the paper distinguishes from memory. Remembering that the toy's name is Mia is a memory function; continuing to behave as Mia, with an established role, preferences, and history, concerns broader character continuity.

Design agency creates expectations that generative systems may not keep

The paper's central interpretation is that children's design choices shaped how they subsequently probed and evaluated their toys, and that when established properties were later contradicted, children often tried to restore the toy they expected. The authors extend prior accounts of customization and the "self-as-source" idea into child-AI interaction: when children helped define an AI character, their agency also produced expectations about how that character should behave. Because persona instructions, conversational context, and memory may not be equally durable or visible, the paper argues that child-facing generative AI should make the status of children's design choices more legible, distinguishing stable character properties from temporary context and model improvisation, indicating when a correction has been saved, and giving children accessible ways to inspect, revise, or co-construct what the toy remembers.

The authors also broaden what counts as AI literacy here. Children's probes went beyond fact-checking to questions of identity, capability, memory, and relationship, and their initial designs foregrounded relational and moral concerns. The paper uses behavioral coherence to describe this broader object of evaluation: whether the AI remains intelligible across its knowledge, identity, capabilities, memory, and social conduct. It suggests AI-literacy activities can move from relationships to mechanisms to consequences, so that a child's idea such as "my toy knows when I am sad" becomes an entry point to how the toy would know, what data it would need, when inference might be wrong, and who should have access.

Three limitations bound these claims. Interactions were relatively short-term and took place within a structured workshop, so the study cannot capture how expectations and relationships might develop through longer-term use. The analysis relied primarily on configuration data and conversational traces, which show what children designed, tested, and repaired but not their interpretations or evolving mental models. And ToyTalk was intentionally a website-based probe rather than a fully embodied AI toy, so it does not capture how embodiment, touch, movement, spatial interaction, and physical affordances might shape children's expectations differently.

The most decision-relevant distinction in this paper is between correcting a toy in conversation and returning to reconfigure it. Children overwhelmingly did the former and rarely the latter, even though ToyTalk kept designs editable and versioned. For anyone building child-facing generative characters, that gap is the priority: the design question is not only whether children can specify a character, but whether the system makes clear where their specification takes effect, how long it persists, and what children should do when the character drifts away from what they authored.

Sources used in this answer