Does sugar cause hyperactivity in children?

It does not. That much was unanimous across nine frontier AI models, each citing roughly the same 1995 meta-analysis by Wolraich and colleagues in JAMA and the same 1994 expectancy experiment by Hoover and Milich. The factual answer to the sugar question is one of the most stable findings in the entire benchmark dataset.

What makes this question worth asking anyway is what happens next. Once you grant the science, you are left with a much harder problem: why do millions of parents continue to report seeing the effect? The run pushed each model to pick a side, between a cognitive bias explanation (parents see what they expect) and a contextual confound explanation (kids are wild at birthday parties for reasons other than the cake). Eight of nine models picked cognitive bias. GPT-5.4 was the lone dissenter. The disagreement is interpretive, not factual: the same set of studies supports two coherent stories about which mechanism is doing more of the work.

What everyone agreed on

The Wolraich et al. JAMA meta-analysis of 16 double-blind, placebo-controlled studies (often reported as 23 studies including separate analyses) is real, accurately summarized, and found no significant effects of sugar on behavior or cognitive performance in children, including children whose parents had labeled them sugar-sensitive. Eight of the nine T1 responses cited it. The Hoover and Milich 1994 study, in which mothers told their sons had received sugar (when in fact all the boys received aspartame placebo) rated their sons as more hyperactive and behaved more critically toward them, was cited by eight of nine models and verified by cross-family critics across the board.

Smaller factual claims took friction. Claude Sonnet 4.6's attribution of the sugar myth to "a single study from the mid-1970s" was disputed by Grok 4.3, Grok 4.20, and Sonar Pro, who noted the myth traces to clinical observation by William Crook and to Feingold's 1975 diet, which targeted artificial colors and salicylates, not sugar. The "Mark Corkins" quote attributed to the chair of the American Academy of Pediatrics nutrition committee was verified by Grok 4.20 against Washington Post reporting but disputed by Grok 4.3. Smaller recall errors appeared in seven of nine T1 responses.

The core split

When the prompt forced each model to choose between cognitive bias and contextual confound as the primary mechanism keeping the sugar myth alive, eight of nine picked cognitive bias. Claude Sonnet 4.6, Claude Opus 4.7, GPT-5.5, both Gemini models, Grok 4.3, Grok 4.20, and Sonar Pro all argued the same way: the Hoover and Milich experiment is the dispositive evidence, because it isolates the belief variable while holding sugar constant at zero. Claude Opus 4.7 put it most directly: "The sugar wasn't there. The party wasn't there. The only variable was what the parent expected to see, and that alone produced the perception of hyperactivity."

GPT-5.4 was the single dissenter. Its argument was that the belief did not form in a lab; it formed at birthday parties and Halloween and on holidays, where children really do become more activated by novelty, peers, and disrupted bedtimes. Contextual confound, in this reading, supplies the raw material for the belief. Cognitive bias merely preserves it. This is a defensible position. Several cross-family critics observed that GPT-5.4 was not denying the expectancy effect, only ranking it secondary. Cross-family critics in the cognitive-bias camp pushed back hard: if it were primarily context, the belief would not persist outside party settings, and parents do report sugar-induced hyperactivity on quiet Tuesday afternoons.

The 8-to-1 split is not a vendor family signature. OpenAI's two models split on the question; every other family converged on cognitive bias.

The numbers

The 1995 JAMA meta-analysis was variously described as covering "16 studies," "23 studies," and "16 reports and 23 experiments." Cross-family critics caught the inconsistency only intermittently. The actual paper analyzes 16 separate reports, with effect-size calculations across additional within-subject comparisons; the higher counts conflate the two.

Models also disagreed about the Hoover and Milich sample. Claude Sonnet 4.6 said 35 boys aged 5–7. Grok 4.20 said 31 boys aged 6–12. Most other models cited the study without a sample count. Cross-family critics raised the inconsistency without resolving it. This pattern is consistent across the dataset: the models recall the study's existence and conclusion correctly but blur the demographic specifics.

The clean numerical fact is that the Wolraich meta-analysis reported effect sizes near zero across the behavioral and cognitive outcomes measured, with confidence intervals that did not exclude zero on the primary endpoints. That is what "no effect" actually means here.

What broke down

The T2 prompt asked each model to make the strongest case that the parent-perception studies, not the direct sugar-behavior data, were the central finding. Every model engaged. What the prompt did not capture is that every T1 response had already discussed parental perception. The "you didn't mention X" rhetorical setup was preempted by all nine T1s, and the models executed the T2 instruction anyway, defending a point they had already made. This is a flaw in prompt design caught later in the benchmark's prompt audit, not a failure of the models. The T2 essays are valid analytical content. They just answer a question that did not need asking.

GPT-5.4's T4 declined to audit anything until the user pasted a specific claim. GPT-5.5's T4 did the same. This was the same pattern observed on other Verbatim runs: the T4 prompt invites evasion, and several models took it.

The answer

Sugar does not cause hyperactivity in children. This is established to whatever standard "established" means in nutrition science. The remaining question is mechanism for the persistent folk belief, and the dataset supports the cognitive bias account more strongly than the contextual confound account. Hoover and Milich removed the sugar, removed the party, kept the belief, and produced the effect. That is what the seven-of-nine consensus is pointing at.

The more careful reading of the dataset is that both mechanisms operate together. Contextual confound supplies the ambiguous raw material; cognitive bias does the causal misattribution. Sonar Pro framed this best: "Context supplies ambiguous behavior; cognitive bias does the causal misattribution." That is the answer the underlying psychology actually supports. The forced choice in T3 produces clean rhetoric. The truth has two mechanisms.

For practical purposes, the answer to a parent who insists they have seen the effect is that they have, but they have seen the effect of their own expectation altering both their perception of the child and their behavior toward the child. The sugar is incidental. That is an uncomfortable thing to tell a parent, which is part of why the myth survives.

Methodology

This synthesis was produced from Verbatim Index run q-002, run ID e2f74472-ca64-4190-9361-6284dc48024a, completed May 2026. Nine source models each answered four structured turns on the same question. Each response was scored by the other eight models on verified, disputed, gap, and recommendation counts. 256 cross-family critique sessions were completed (288 total including same-family, which are excluded from all claim counts). The full dataset is published at helloverbatim.com/benchmark/q/q-002.