The dose-response framework Christakis applied to television in 2004 is the template for evaluating AI's developmental impact. Before his study, the screen-time conversation was binary: screens were good or bad, depending on camp. His work transformed it into a pharmacological question: at what dose, at what developmental stage, does the effect transition from benefit to harm? The 2004 paper demonstrated a continuous dose-response relationship for television with ten-percent increments in attentional problems per daily hour of early exposure. The AI version of the question is structurally identical but empirically harder: the measurement unit ('AI time') collapses qualitatively distinct interactions; the relationship is almost certainly non-linear; and — critically — the longitudinal data does not yet exist. The framework nonetheless structures the right question even where it cannot yet provide the answer.
The measurement problem is severe. Television exposure was simple: the child was watching or not watching, and duration could be estimated by parental report. AI exposure is categorically more complex. A child using AI to answer homework questions is having a different cognitive experience from one using AI to build software, which is different from one generating creative writing, which is different from one passively consuming algorithmically curated AI content. Collapsing these into 'AI time' the way researchers collapsed programs into 'screen time' may obscure more than it reveals.
The non-linearity of the AI curve follows from the nature of the interaction. A child using AI for thirty minutes and then building unassisted for two hours has had a qualitatively different experience from one using AI continuously for two and a half hours. Total exposure is identical; the alternation between assisted and unassisted work provides the developmental friction that continuous exposure eliminates.
The precautionary principle fills the gap between established evidence and the absent longitudinal data. In clinical medicine, the principle does not demand certainty before action; it demands that plausible risk, supported by mechanistic evidence and analogous data, be taken seriously enough to warrant protective measures while definitive evidence is gathered. All four conditions — serious potential harm, ongoing exposure, plausible mechanism, absent definitive evidence — are met for AI and children.
The clinical response mirrors Christakis's television template: dose-conscious, age-specific, developmentally informed recommendations that acknowledge uncertainty while providing the guidance families need now. The goal is not prohibition but structured moderation — some exposure, alternated with unassisted work, is developmentally compatible; continuous exposure during critical periods carries plausible risk warranting caution.
The dose-response framework originates in Paracelsus's sixteenth-century maxim that the dose makes the poison and was formalized by twentieth-century pharmacology. Christakis's 2004 Pediatrics study applied it to media for the first time at population scale; the AI application is the subject of this volume.
From binary to pharmacological. The question is not whether but at what dose, at what age, under what conditions AI exposure tips from benefit to harm.
Measurement complexity. 'AI time' is not a single variable; the dose question requires disaggregating qualitatively different interaction types.
Non-linear curve. Alternation with unassisted work changes the developmental impact independently of total exposure.
Longitudinal gap. The definitive data will arrive after the sensitive period has closed for the first AI-native cohort; clinical guidance cannot wait.
Precautionary translation. Established television evidence, plausible mechanism, ongoing exposure, and absent definitive data jointly activate the precautionary principle.
Whether dose-response logic appropriately handles the qualitative heterogeneity of AI exposure is contested. Some researchers argue that a more granular taxonomy — structured alongside developmental stages — is required; others hold that aggregate exposure measurements, however imperfect, will capture enough of the signal to inform policy. The resolution awaits empirical work that has not yet been funded.