Can AI Do Formal Demography? Why Vaccinating the Oldest Saves Both the Most Lives and the Most Years of Life

Joshua R. Goldstein, Kenneth W. WachterarXiv 2026 · arXiv

TL;DR

Researchers successfully collaborated with an AI model (Claude) to simplify and generalize a complex mathematical proof in formal demography. The result confirms that for most common mortality patterns, vaccinating the oldest individuals saves both the most lives and the most years of life, debunking the common assumption of a trade-off between the two.

What Problem Does This Paper Address?

In public health, there is often a perceived tension between saving the most lives (which usually favors the elderly) and saving the most years of life (which usually favors the young). During the COVID-19 pandemic, authors Goldstein and Wachter previously proved that for this specific disease, these two goals actually align: vaccinating the oldest is the optimal strategy for both metrics.

However, their original 2021 proof was mathematically dense and indirect. This paper explores whether an AI can help "do" formal demography—the mathematical study of population dynamics—by finding a more intuitive, elegant, and general proof for this "Key Relationship."

Methods

The authors engaged in a multi-session human-AI collaboration with Claude Opus 5.5. The process was iterative:

  1. Explanation: The AI was asked to explain the original complex proof.
  2. Simplification: The researchers "pressed" the AI for a simpler version, specifically one built on the reciprocal of the hazard rate ($1/\mu$).
  3. The Tangent-Line Proof: Claude proposed a new three-step proof using the tangent-line inequality. This method relies on the "frozen-hazard life expectancy"—the years a person would live if their current risk of death never changed.
  4. Literature Synthesis: The AI searched for historical precedents, connecting the new proof to existing identities in reliability theory and actuarial science.

Datasets

As a theoretical paper in formal demography, it does not use large-scale empirical datasets for training. Instead, it evaluates the proof against standard parametric mortality models:

  • Gompertz Model: Where the risk of death rises exponentially with age.
  • Gompertz–Makeham Model: Which adds a constant "background" risk of death to the senescent component.
  • Weibull and Logistic Models: Used to test the generality of the AI’s proposed conditions.

Key Insights and Ideas

  • The "Frozen-Hazard" Intuition: The AI framed the problem around $1/\mu(x)$. If this value is convex (curves upward), then the product of hazard and life expectancy ($\mu(x)e(x)$) must rise with age.
  • Convexity is Key: The AI identified that the "oldest first" strategy works whenever life expectancy itself is convex in age ($e''(x) \ge 0$).
  • A New Lower Bound: The collaboration yielded a previously unstated lower bound for life expectancy: $e(x) \ge 1 / (\mu(x) + \text{rate of ageing})$. This provides a quick, accurate approximation for life expectancy at older ages.
  • AI as a Mathematical Partner: The paper demonstrates that AI is not just for data crunching; it can revisit established theory, find clearer arguments for why things work, and suggest valuable generalizations that human experts might overlook.

Key Results

The AI-generated proof is shorter and more general than the original. It shows that the "years saved per person" rises with age as long as the rate of ageing does not increase faster than its own square.

Figure 1: (a) Gompertz and Gompertz–Makeham hazards (α = 5 × 10−5, β = 0.1). Dots mark the age log(c/α)/β above which 1/µ is convex. (b) The product µ(x)e(x). It rises at all ages under Gompertz mortality, and at all ages above the dot under Gompertz– Makeham mortality. All three curves approach 1 at the oldest ages.

As seen in Figure 1, while the relationship holds universally for the Gompertz model, it can fail at younger ages in models with high background mortality (like the Gompertz-Makeham model). However, even in those cases, the relationship eventually "recovers" at older ages, supporting the prioritization of the elderly.

Limitations and Future Work

  • Human Steering: The authors emphasize that the "division of labor" still requires humans to ask the right follow-up questions and judge the value of the AI's output. The AI can derive and draft, but humans provide the framing and structure.
  • Checkable Conditions: While the AI provided an "exact" condition for when the relationship holds, that condition involves person-years lived at future ages, meaning it cannot be checked from the current hazard rate alone. Convexity of $1/\mu$ remains the most "practical" test.
  • Verification: The authors had to manually verify citations, as the AI occasionally referenced works it hadn't actually "read" or provided incorrect publisher details.
  • Future Directions: The authors suggest that as models improve, they may begin to generate their own follow-up questions, potentially leading to even deeper autonomous discoveries in formal theory.

Summary updated

Abstract

Can AI contribute to formal demography? We report an experiment in which an AI model, Claude, was asked to explain and then simplify the proof of an earlier result: that vaccinating the oldest against COVID-19 saves both the most lives and the most years of life, because the years saved per person treated rise with age. The model proposed a new proof that is shorter, more intuitive, and somewhat more general than the original, built from a known life-table identity and the definition of convexity. It states the condition for the result as the convexity of the reciprocal of the hazard, which can be checked from the hazard alone and holds for most familiar parametric models, and it yields a lower bound on life expectancy whenever that condition is met. Prompted by a further question, the model also gave an exact form of the condition, showing how far convexity can be relaxed. The model traced these identities and related results in earlier literature. An edited record of the sessions is provided, and complete transcripts are available on request. The exercise suggests that AI can revisit established results, make them easier to understand, and prove valuable generalizations.
BibTeX
@misc{goldstein2026can,
  author = {Joshua R. Goldstein and Kenneth W. Wachter},
  title = {{Can AI Do Formal Demography? Why Vaccinating the Oldest Saves Both the Most Lives and the Most Years of Life}},
  year = {2026},
  eprint = {2611.03500},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.03500},
}
Figure