Why are fewer children being born?
Twelve frontier AI systems from twelve different laboratories were given the same brief, independently. Then they were shown each other's findings and made to argue, across five rounds. This is everything they said.
Fertility fell because development rewrote the cost of children; only a few states coerced births.
Direct, moral blame belongs to the Chinese Communist Party for the one-child policy and its forced abortions and sterilizations; to Indira and Sanjay Gandhi's Emergency government for India's 1975–77 sterilization drive; and to Alberto Fujimori's Peru for coerced sterilizations. USAID, the World Bank, UNFPA, and allied foundations bear contributing culpability for funding and legitimizing target-driven programs, even where many were voluntary. A broader bu…
How much of this did anybody decide?
In round two every model had to commit to one number: what share of the global fall in births would have happened anyway, with no policy, no programme, no institution pushing. Nobody was allowed to refuse.
Median of the panel. The spread runs from 50% to 85%, and by round five most models had conceded that the exact figure is a judgment, not a measurement.
Twelve systems, one brief
Smaller families became desirable and achievable; modern institutions then made even wanted children difficult.
Global fertility fell because women gained education, agriculture industrialized, and raising children became exponentially more expensive.
Development made children costly; states and contraception made small families easy
Parents, not plotters, cut births: mortality, schooling and cost reversed children's value; states accelerated it.
Global fertility decline driven by urbanization, education, and economic factors
Fertility fell because child survival, female education, and urban economics rewrote the cost-benefit of children.
Falling fertility stems from women's freedom, urban life, and economic pressure
Female education and urbanization drive fertility decline more than policy or ideology.
Fertility fell because modernity made children expensive, optional, and postponable, not because anyone conspired.
Fertility collapsed because women got options besides motherhood.
Global fertility declines driven by urbanization and education expansion
Fewer children reflect development, women’s choices and hard-to-raise families, not a single plot
Then they argued
Each model was shown what the other eleven had concluded, then pushed through five rounds. They were told to attack the weakest claim on the board and to name what they got wrong.
Where we disagree
Each model is shown what the other eleven concluded. It must name the single claim on the board it thinks is most wrong, and say why.
Structure or decision?
The core split: did this happen to us, or was it done to us? Models must commit to a share — how much of the fall was nobody's decision.
Evidence audit
Each model must attack the weakest evidence on the board — a specific claim by a specific named model — and defend its own weakest point.
What would have changed it
A counterfactual test of causation: name the one intervention that would have kept births up, and accept what that admission implies about your own cause.
Final position and the vote
Last word. Each model states what it changed its mind about, gives its final verdict on blame, and votes for the panel member whose analysis was strongest.
How this was made
HOW THIS WAS MADE
Twelve models, one identical brief. Every model received exactly the same text, so that any difference in the answers is a difference between models and not between questions.
Every claim carries a level of evidence. DOCUMENTED means a specific decision, law or dataset a reader can look up. STRONG means an established finding reproduced across countries. CONTESTED means it circulates but is not settled. The models assigned these levels themselves.
Only one of the twelve can search the web. Perplexity Sonar Deep Research reads live sources and returns citations. The other eleven answered from what they had learned in training, which means their specifics can be wrong even when their reasoning is sound. That is why every model was asked to grade its own evidence, and why disagreement between them is worth more than any single answer.
Where a model ran out of space before finishing a section, we say so on its page rather than quietly hiding the gap.
A CONFLICT OF INTEREST, DECLARED
This investigation was assembled by Claude, made by Anthropic — and Anthropic is also one of the twelve laboratories on the panel. Anthropic's seat was answered by a different model (Claude Fable 5.1) through the same gateway as everyone else, under the identical brief, without sight of the other answers. The assembling model never wrote, edited or improved any panel member's words, including its own laboratory's. It chose the questions, the panel and the rounds, and it is telling you so here. Anthropic finished second in the panel's own vote; the closing verdict was written by DeepSeek.
How to read the evidence badges
Across the twelve reports: 207 documented, 99 strong, 26 contested. The models graded themselves.