Should children’s clinical AI follow different rules?
The American Academy of Pediatrics says generative-AI tools used in children’s care need pediatric-specific evidence, family-centred design, disclosure and continuous monitoring. The policy is influential guidance—not a trial, regulation or finding that any tool improves outcomes.
The short version
- What happened
- The AAP calls for evidence of safety and effectiveness in children and adolescents specifically, with validation across age, development, language, disability, race, ethnicity, gender and socioeconomic circumstances.
- Who it affects
- Families should be told in understandable language when generative AI materially shapes clinical care and who remains accountable for the decision.
- What remains unproven
- This is a professional policy statement, not a clinical trial, systematic review with reproducible methods, product approval or binding law.
Keep in mind
This is a professional policy statement, not a clinical trial, systematic review with reproducible methods, product approval or binding law.
The statement provides no new patient sample, comparator, outcome denominator or effect estimate and therefore cannot show that any recommendation improves care.
Its legal and privacy discussion is centred on the United States; implementation requirements differ across jurisdictions.
The scope excludes AI used directly by children and families and non-clinical applications affecting children.
Author financial relationships are disclosed and managed through the AAP process; readers should still consider them when assessing commercially relevant recommendations.
Source analysis · 1 linked source. Primary analysed source date: 3 October 2026.
Read the full report · 10 min The evidence, methods, implications and limitations
The policy draws a line between adult and pediatric evidence
The American Academy of Pediatrics published a policy statement on 3 October setting out how generative-AI tools should be developed, bought, introduced and monitored when they support children’s clinical care. Its central premise is simple: evidence from adults, general-purpose benchmarks or simulated medical questions should not automatically be treated as evidence for children.
Children and adolescents differ in physiology, development, communication, consent and family involvement. Neonatal, childhood and adolescent care also span very different risk profiles. A model that produces a plausible answer for an adult encounter may fail when doses, developmental milestones, rare disorders, safeguarding concerns or a parent-child-clinician relationship change the context.
The AAP says real-world validation remains limited and notes that much of the literature tests models on examinations or simulated scenarios rather than care delivered to patients. It also cites evidence that large language models can perform worse in pediatrics than in some adult specialties. The policy therefore asks developers and regulators to demonstrate safety and efficacy in pediatric populations specifically.
That is a policy judgement informed by a literature review, not a new performance study. The statement supplies no new patient cohort, intervention group, comparator, outcome denominator or effect estimate. Its contribution is to define what evidence and governance should be demanded before institutions rely on these tools.[1]
Developers are asked to design around children and families
The first group of recommendations targets researchers and developers. Training data should represent pediatric diversity across age, developmental stage, race and ethnicity, gender, sexuality, disability and socioeconomic status. Validation should report performance across relevant subgroups rather than compressing children into one average score.
The policy also treats family-centred care as a design requirement. Pediatric decisions often involve a child or young person, one or more caregivers and clinicians with different legal and ethical responsibilities. A generative system that addresses only the clinician’s efficiency could weaken communication, obscure whose preferences were considered or introduce advice that is developmentally inappropriate.
For identifiable patient data, the AAP calls for consent and authorisation consistent with law and ethics, transparent explanations of use and commercialisation, and safeguards reflecting institutions’ custodial duties toward children. It highlights re-identification risk in rare pediatric genetic conditions, where removing obvious identifiers may not make a record practically anonymous.
It further recommends accessible processes for requesting removal of data, while acknowledging that deletion may be limited after a model has been trained or deployed. This is an important qualification: a data-removal promise should explain what can actually be changed in source records, training sets, model weights, logs and downstream products.[1]
Hospitals are told to verify more than a vendor’s headline score
For procurement, the policy tells health-care institutions to examine whether a developer validated the tool across diverse pediatric populations, understand conflicts of interest, complete privacy and security due diligence, minimise data collection and inspect the mechanisms for monitoring and reporting errors.
This shifts responsibility away from a one-time demonstration. A model’s behaviour may change when its underlying service, prompt, retrieval source or clinical workflow changes. Performance can also differ by hospital, documentation style, specialty and patient mix. Institutions therefore need local acceptance criteria and continuing surveillance rather than assuming that a published benchmark remains valid after deployment.
The statement refers specifically to United States HIPAA obligations and business associate agreements. Those rules are relevant to American institutions but are not a complete global privacy framework. UK organisations would need to assess UK GDPR, the Data Protection Act, NHS information-governance requirements and applicable medical-device rules; other countries have their own bases for processing children’s data and their own thresholds for regulated software.
Procurement should also distinguish use cases. Drafting a clinician’s letter, translating discharge instructions, summarising a chart and recommending a treatment do not create the same risk. Evidence, oversight and fallback procedures should be proportional to what the system can influence and how readily a human can detect an error.[1]
Disclosure and direct human oversight are the near-term standard
During implementation, the AAP calls for equitable access, explicit safety protocols, error reporting, continuous quality monitoring and disclosure of generative-AI involvement in culturally appropriate plain language. Disclosure should support shared decision-making rather than become a technical notice that families cannot understand or act on.
The policy is unambiguous about present responsibility: current clinical uses require a human in the loop, and the pediatrician remains responsible for medical decisions. More autonomous operation, it says, would need rigorous validation and regulatory approval. That position rejects the idea that a disclaimer can transfer clinical accountability to a model or its vendor.
Human oversight is necessary but not sufficient. A clinician under time pressure may accept fluent text, miss a subtle dosing error or assume the system saw data that were absent. Useful monitoring must therefore measure whether people detect and correct errors, whether workload changes, and whether families understand when AI shaped their care.
The policy does not specify one universal disclosure script or consent requirement. Local law, risk and use case will affect what is appropriate. An administrative drafting tool may need a different conversation from a system that ranks diagnoses or recommends an intervention. Institutions will need to turn the principles into testable procedures.[1]
Regulators are asked to treat changing models as a lifecycle problem
The AAP asks regulators and oversight bodies to require pediatric evidence before approving systems for children and adolescents, then conduct premarket evaluation and postmarket surveillance that account for response variability and performance drift. This lifecycle framing is especially important when a hosted model can change without the hospital installing a conventional software update.
A meaningful postmarket programme would track versions, inputs, outputs, overrides, incidents and subgroup performance. It would define when a change requires revalidation, when a tool should be restricted and who can suspend it. Families and clinicians would also need a route to report harm or confusing behaviour without navigating a vendor’s internal process.
The statement is influential professional guidance, not a new federal regulation. It does not approve or prohibit a product, create an enforcement mechanism or settle whether a particular tool is a medical device. Its recommendations may shape hospital policy, vendor requirements and advocacy, but legal force depends on regulators, legislatures, contracts and existing professional duties.
The AAP also states that its guidance is not an exclusive course of treatment or a standard of medical care. Its policies expire after five years unless reaffirmed, revised or retired. Given the pace of model development, procurement evidence and regulatory practice are likely to change well before that outer date.[1]
The policy’s evidence base and interests should stay visible
The authors say there was no external funding and that the AAP neither sought nor accepted commercial involvement in developing the publication. The statement records that one author has ownership in the MEDA Angels fund and another has advisory, board or consultancy relationships with Big Health, Glooko, MDIC/Nestcc, Rock Health, Sanofi and NCQA. The AAP says disclosures were reviewed through its conflict-management process.
Those disclosures do not invalidate the policy, but they are relevant to readers assessing recommendations that may influence a commercial clinical-AI market. The statement is produced by a professional association and committees with clinical informatics expertise; it is not an independent systematic review with a published search strategy, study-selection flow, risk-of-bias table or graded certainty for each recommendation.
Its scope is also narrower than ‘children and AI’ generally. It concerns generative-AI tools used by pediatricians to support clinical care. It explicitly does not cover children’s direct use of AI, family-facing consumer chatbots, education, social media or other non-clinical applications. Those areas need separate evidence and safeguards.
Most importantly, recommendations are not outcome data. The policy does not show that disclosure prevents harm, that a human reviewer reliably catches errors, or that pediatric-specific training improves diagnosis. Those are plausible and defensible requirements that still need empirical testing.[1]
What would show the guidance is working
Health systems adopting the framework should publish use-case-specific evaluations with clear denominators: numbers of encounters and patients, age and developmental groups, languages, disabilities, clinical settings, model versions, error types, overrides and adverse events. Comparators should include the existing workflow, not only another AI system.
Prospective studies should measure patient safety, decision quality, communication, time, clinician workload and family understanding—not merely answer accuracy. Subgroup results and confidence intervals are essential, especially where rare conditions make averages misleading. Independent replication across hospitals and countries would test whether results transfer.
Regulators and vendors should make model changes traceable and report what triggers revalidation. Institutions should test whether their disclosure is understood, whether incident reporting is used, and whether clinicians can stop or work safely without the tool. Children, adolescents and caregivers should participate in defining acceptable trade-offs.
For now, the policy’s practical message is cautious and actionable: children’s clinical AI should not borrow credibility from adult data or general benchmarks. It should earn trust through pediatric evidence, understandable disclosure, secure data practice, accountable clinicians and monitoring that continues after purchase.[1]