How Kids Decide Which AI Chatbots to Trust: A Design and Policy Checklist

When 115 learners aged 8–18 built 119 chatbots in a workshop-based co-design study (How Children Design and Reason about Trustworthy AI Chatbots, posted to arXiv in September 2026), the 10–13-year-olds set significantly higher confidence for their bots than the 14–18-year-olds did, and not one of the findings pointed to a disclosure label as the thing that moves a child’s trust. That is the practical finding for anyone reviewing a chatbot reachable by under-13s: the evidence we have says children calibrate trust on conversational and design cues, so a review workflow that checks whether the bot says “I am an AI” and stops there is inspecting a surface the child-side research does not identify as the operative one.

Two cautions belong up front, because they shape everything below. First, every number from the anchoring study is author-reported from a single paper, How Children Design and Reason about Trustworthy AI Chatbots, with no replication in the studies reviewed here. Second, the claim that labels do not matter is an inference, not a measured result: no study in this evidence base directly tests whether disclosure text shifts children’s trust. What the evidence does support is a checklist that tests interaction design, and that is what this article builds.

What kids actually do when they design a “trustworthy” chatbot

The primary study is a mixed-methods co-design project: in the authors’ words, “We conducted mixed-methods study with 115 learners (ages 8–18) who made 119 chatbots” (arXiv:2609.25244). Co-design flips the usual research posture. Instead of asking children whether they trust a finished product, it watches what they build when told to make a chatbot someone should trust, including how confidently they set the bot to speak.

That posture matters for product reviewers because it surfaces what children treat as the trust-bearing parts of an agent. The strongest developmental finding in the author-reported results is about confidence calibration: “Younger students (age 10-13) set significantly higher confidence than older students (age 14-18).” The full results widen that band rather than narrow it to the middle years: “Elementary school (M=2.68, SD=0.47) and middle school students (M=2.62, SD=0.62) both set significantly higher Confidence than high school students (M=2.29, SD=0.59; both p_BH<.05), with no difference between elementary school and middle school students (p_BH=.88).” A chatbot that always sounds certain is not a neutral default for this audience; it lands differently depending on the child’s age, and the students who set it highest are the younger ones, roughly ages 8–13.

Before you put that finding in a requirements document, hold it next to the sample behind it, which the paper describes as “115 students–76 students at Grades 2-3 (ages 8-9), 24 students at Grades 6-7 (ages 10-13), 15 students at Grades 9-12 (ages 14-18).” Two thirds of the participants are eight- and nine-year-olds, and the three-group confidence comparison rests on a younger side of 76 elementary plus 24 middle schoolers whose 44 and 58 chatbots carried the trait analysis, against an older side of 15 teenagers (17 chatbots), according to the same paper. The abstract’s narrower 10-13 versus 14-18 contrast is the one that pairs 24 students against 15. The paper also flags the younger half’s data quality: “elementary student data came primarily from teacher reflections and platform logs rather than direct interviews with children,” and the authors include the elementary trait comparisons “for exploratory purposes only.” The age-band effect is the most actionable result in the studies reviewed here and also the most fragile. Treat it as a hypothesis your own testing should confirm per age band, not as a settled developmental fact.

The cues children appear to calibrate on

The wider child-agent literature points away from badges and toward relationship. In the ConvoBlocks work on teaching children about conversational agents, the authors report that an earlier workshop study “did not find any significant differences in students’ perceptions of agents’ trustworthiness through the workshops; however, they did find correlations between perceptions of trustworthiness, safeness and friendliness.” Read that twice. An instructional intervention did not move perceived trustworthiness, but the bots children perceived as safer and friendlier were the ones they perceived as more trustworthy. The trust surface, for this population, looks relational and conversational rather than declarative. (The ConvoBlocks study itself is small: 49 participants completed at least one survey, 27 children with an average age of 13.96 and 19 parents, according to the authors’ report.)

Two further cue families come from adjacent evidence, and both arrive with caveats a careful reviewer should keep attached.

Explainability and challengeability. In a general-population study of explainability and trust, “Users frequently equated ‘understandable’ decisions with ‘fair’ ones. Even when exposed to the less accurate (‘Bad AI’) system, participants expressed greater acceptance if they could interpret its reasoning or challenge it through interaction.” That is a mechanism with direct design implications: an agent that can explain an answer, and that a child can push back on, earns acceptance even when its accuracy is worse. But this was not a child sample, and the authors describe their results as exploratory given a moderate sample and self-reported trust measures. It tells you what to test with children, not what children do.

Transparency about function, limits, and data use. The children’s AI UI/UX best-practices literature frames trust as built through disclosure of how the system works, not just what it is: “Providing clear feedback and interactive guidance promotes independent learning. Building trust through transparency ensures that users are fully informed about how AI functions, its limitations, and how their data are utilised,” per that best-practices paper. Note what this is and is not. It is an expert synthesis, not a measurement of children, and the transparency it calls for is a running property of the interaction (feedback, guidance, explanation of limits) rather than a one-time label.

There is also an expert-consensus route from principles to design requirements. A Delphi-method study took existing AI ethics guidelines and adapted them “to this particular system and population… to help CAs developers improve their design towards trustworthiness and children,” producing specific guidelines for child-facing conversational agents. Only the abstract is quoted here, so the guideline contents are not established; the methodology, though, is the closest existing analogue to the policy-to-design mapping a compliance review needs.

One more axis deserves a place on the checklist even though the evidence is thin: who else is in the room. In a neuroimaging study of young children’s anthropomorphism of an AI chatbot, “In the right dmPFC, higher perceptive scores were associated with greater activation during the AI-only condition and with lower activation during the AI+Parent condition,” according to the abstract. A parent sitting beside the child appears to change how the child processes the agent socially. That is a single abstract-only finding, but it is enough to justify testing the co-presence condition rather than assuming the solo-interaction review covers it.

The label problem, stated precisely

The angle of this article is sometimes summarized as “labels don’t work.” That overstates the evidence. Here is the accurate version:

• No study in this evidence base measures whether an “I am an AI” disclosure shifts children’s trust in either direction. The effect is untested in the studies reviewed here, not refuted.
• What is tested points elsewhere: in child samples, trust covaries with perceived safeness and friendliness; in a general-population study, interpretability and challengeability raised acceptance; and at least one structured intervention with children showed no significant shift in trust perceptions.
• The design guidance that does exist locates transparency in ongoing interaction properties, not in a static declaration.

So the defensible claim is narrower and more useful: a review that checks only disclosure text is unverified against the cue surfaces the child-side literature actually identifies, and a label requirement, on current evidence, cannot be assumed to do the work it is often assigned. I would not remove the label; I would stop treating it as the control.

When trust stuck to both good and bad agents

The strongest counter-evidence among the studies reviewed here should keep any checklist author humble. In a pandemic-era study designing conversational robots with children, the researchers deliberately built trustworthy and untrustworthy conditions and got this result: “Although we tried to manipulate the robot to be either very untrustworthy or very trustworthy, we noticed high trust in both of them.” Their own diagnosis: “Although even small behavioral changes in a robot is reported to potentially influence a child’s perception of the robot (Peters et al., 2017), our manipulation might not have been strong enough to achieve the impact we were aiming for.”

Pair that with the ConvoBlocks null result on workshops and a pattern emerges. Children’s trust in agents is sticky, and designed cues in either direction may move it less than adults assume. This cuts against overconfidence in both camps: against the assumption that a warning label will defuse trust, and against the assumption that a friendly persona or a confidence slider will reliably tune it. It is why the checklist below prescribes testing cues with your actual users rather than certifying them from the literature. The child-robot field’s own authors model the right posture: “We acknowledge that our findings are preliminary and exploratory, and a study with a larger sample size is necessary to draw generalizable conclusions,” as one related child-robot study puts it. Every child-side number in your review documentation should carry the same label.

The review checklist

Each row pairs a trust cue with its evidence basis and a concrete review activity. Evidence strength is graded honestly: nothing here is a replicated, child-measured effect.

Two things this table deliberately does not do. It does not assign pass/fail thresholds, because the evidence does not support any. And it does not treat co-design choices as proof of behavior toward deployed products: what children build when asked for a trustworthy bot and what they trust in an app they did not make are different measurements, and no source here bridges them.

The regulatory half of this checklist is unwritten

A reviewer running these design tests also has a legal half to map for a product reachable by under-13s. On that half, honesty requires a blunt statement: no source in the studies reviewed here covers any legal instrument. Nothing here can tell you whether COPPA applies to an under-13 chatbot or what it requires, what the UK Age Appropriate Design Code says about transparency, or whether any state companion-chatbot statute applies in your jurisdictions and what it mandates. Any article or internal document that attaches this design checklist to a statute citation, effective date, or compliance mapping needs those obligations verified against primary legal texts first.

What the design evidence does contribute to the policy conversation is a structural point. If a rule’s operative requirement is a disclosure label, the child-side literature suggests the rule is checking a surface whose effect on children is unmeasured, while leaving the surfaces children appear to calibrate on (confidence calibration, persona warmth, challengeability) unexamined. That is a reason for a trust-and-safety lead to run interaction tests beyond what the letter of a label rule demands, and a reason for anyone drafting or commenting on such rules to ask whether the mandated artifact matches the measured mechanism. It is not, on this evidence, a reason to conclude labels are useless.

The decision

Shift under-13 chatbot review from disclosure-text checks to interaction-design tests. The concrete minimum: confidence calibration measured per age band (with the 8–13 over-setting pattern as the hypothesis to confirm, since the younger side combines 76 elementary and 24 middle school students against 15 teenagers), answers that explain themselves and survive a child’s challenge, transparency about function, limits, and data use delivered in the flow of interaction, and a test condition with a parent present. Keep the AI disclosure label, but stop crediting it as a trust control, because no study in this evidence base measures its effect on children and the correlational evidence points elsewhere.

Run that checklist with the labeling discipline this literature models: every child-side number is author-reported, single-study, or both; the counter-evidence shows children’s trust can stay high even under deliberately degraded cues; and the regulatory mapping does not exist in these sources at all. A review process built on those terms is more work than checking for a label. It is also the only version the evidence actually supports.

Frequently Asked Questions

What specific design cues do children use to calibrate trust in AI chatbots?

The design guidance that does exist locates transparency in ongoing interaction properties, not in a static declaration.

Does the evidence show that an ‘I am an AI’ disclosure label affects children’s trust?

No study in this evidence base measures whether an “I am an AI” disclosure shifts children’s trust in either direction. The effect is untested in the studies reviewed here, not refuted.

How should reviewers test the impact of a parent being present during a child’s interaction with an AI chatbot?

Test the interaction with a parent present; check whether trust cues land differently