4.1 New medicines
+8 years
Modeled U.S. female median gain in 2037 vs. baselineOur 20× forecast combines roughly 10× as many new medicines with twice the mortality benefit per drug.
- Historically
- 35%
- Future estimate
- 57%
1 · A data-driven research scenario
Can AI bring median lifespan to 100 by 2037?
Dario Amodei argues that AI could help defeat most diseases within 5–10 years of powerful AI arriving.
How reasonable is this? What would have to change in science, healthcare, and mortality itself to get there?
2
The pace of progress
Reaching our goal requires roughly 20× faster mortality improvement.
AI could bring larger longevity gains over the next decade than in any recent period.
To reach EOD 2037’s target of a median lifespan of 100, our model requires roughly a 20× acceleration in mortality improvement. That would mean approving many more effective medicines, improving how well they work, and getting them to more people.
5× · United States≈ 89.1median lifespan in 2037
10× · United States≈ 94.6median lifespan in 2037
20× · United States≥ 100.0median lifespan in 2037
United States median lifespan
1×25×
Median lifespan years
ObservedModeled from 2025COVID outliers
1× · historical pace83.9 yearsmedian lifespan in 2037
5× · selected pace89.1 yearsmedian lifespan in 2037
Data & calculation
Average of male and female period medians. Observed through 2023; 2025 is modeled, not observed.
Each projection starts from United States's latest observed age-specific mortality rates. Country-specific fits from 2003 onward supply annual log-hazard declines; the slider multiplies these declines from 2025. Fitted increases are capped at zero. COVID years 2020–2022 are shown but excluded from fitting. Sex-specific medians are capped at age 100; ≥ marks a lower bound when either sex reaches that limit. This is a period-life-table scenario, not a forecast of individual lifetimes.
3
Current causes of mortality
Three disease groups account for 55% of all deaths. Addressing them will be the source of most mortality gains.
Heart disease, cancer, and neurological diseases & dementias are the major causes, with a long tail of other sources of mortality.
Estimated annual risk of death by cause
Heart disease 0.00089%Cancer 0.0019%Neurological diseases & dementias 0.0013%Other causes 0.01%
Data and calculation
P(death) is an estimated one-year risk using constant, competing cause-specific hazards within each age band. It is not the probability of dying across the whole five-year band. All-age estimates are population-weighted.
Age-band curves start at 1–4; all-age totals still include infants. Bar length is relative to the largest displayed cause group, not percent of all deaths. Cause counts exclude the 57 deaths with unknown age. Totals stay fixed at 2024, all ages, both sexes combined.
Final U.S. resident underlying-cause deaths from CDC/NCHS NVSS, 1999–2024, with NCI SEER population estimates. Historical sex-specific NVSS counts are distributed by the WHO Mortality Database; Original NVSS microdata are used for 2019, 2020, and 2022–2024 to preserve CDC cardiac-arrest classification. Earlier WHO releases have small coding and revision differences. One cause per death; these categories do not overlap. Unintentional injuries includes accidental drug poisoning, falls, and other accidents. Other causes includes all remaining causes. Age bands end at 90+.
Within a band: P(cause in one year) = (D_cause / D_all) × [1 − exp(−D_all / population)]. Unknown-age records are excluded from all chart calculations. Combined totals include two unknown-sex deaths in 2009. Deaths and population denominators combine both sexes. This section is independent of the lifespan model’s country and sex settings.
| Age | Heart disease | Cancer | Neurological diseases & dementias | Other causes |
|---|---|---|---|---|
| 1–4 | 0.00089% | 0.0019% | 0.0013% | 0.01% |
| 5–9 | 0.00040% | 0.0020% | 0.00098% | 0.0049% |
| 10–14 | 0.00048% | 0.0021% | 0.0014% | 0.0053% |
| 15–19 | 0.0013% | 0.0028% | 0.0020% | 0.02% |
| 20–24 | 0.0024% | 0.0035% | 0.0029% | 0.02% |
| 25–29 | 0.0051% | 0.0055% | 0.0031% | 0.03% |
| 30–34 | 0.0099% | 0.01% | 0.0036% | 0.03% |
| 35–39 | 0.02% | 0.02% | 0.0044% | 0.04% |
| 40–44 | 0.03% | 0.03% | 0.0060% | 0.06% |
| 45–49 | 0.05% | 0.06% | 0.0084% | 0.07% |
| 50–54 | 0.09% | 0.10% | 0.01% | 0.09% |
| 55–59 | 0.14% | 0.17% | 0.02% | 0.13% |
| 60–64 | 0.22% | 0.29% | 0.04% | 0.19% |
| 65–69 | 0.32% | 0.44% | 0.07% | 0.27% |
| 70–74 | 0.46% | 0.59% | 0.17% | 0.39% |
| 75–79 | 0.71% | 0.81% | 0.41% | 0.60% |
| 80–84 | 1.3% | 1.1% | 0.95% | 1.0% |
| 85–89 | 2.4% | 1.5% | 2.0% | 1.7% |
| 90+ | 5.3% | 1.6% | 4.3% | 3.2% |
4
How AI could change the curve
New medicines and wider access to care will lead to the biggest gains.
Medicines could deliver the largest gain; public health could spread the longevity already enjoyed by the wealthy; other medical advances could add a smaller contribution.
4.2 Public health
+5 years
Potential life-expectancy gain from income–longevity equalizationAI could narrow the longevity gap between typical Americans and the wealthy by making effective prevention and care widely available.
- Historically
- 44%
- Future estimate
- 36%
4.3 Other medical care
+1 year
Author’s assumption for additional longevity gainsBetter diagnosis, procedures, and care decisions make a smaller contribution in our forecast. Detecting disease earlier helps only when it changes outcomes.
- Historically
- 13%
- Future estimate
- 7%
Historically: estimated shares of U.S. life-expectancy gains, 1990–2015, from Buxbaum et al. (2020).
How to interpret the estimates and contribution weights
The future estimates express the relative sizes of the 8-, 5-, and 1-year assumptions: 8/14, 5/14, and 1/14, rounded to 57%, 36%, and 7%. They sum to 100%. The denominator is an arithmetic total for this comparison, not a predicted 14-year combined lifespan gain. These descriptive shares do not replace the drug calculator’s 0.44 mortality-improvement weight or form a fitted joint model. The historical figures describe a different period and are not a complete allocation.
The year estimates have different meanings. The 8 years is the existing model’s U.S. female period-median gain in 2037 at 500 new drugs per year and 2× benefit per drug. The ~5 years is a rounded income–longevity comparison in expected lifespan conditional on reaching age 40, not a demonstrated median gain. The ~1 year is a judgment for other care, not an output of the screening studies.
These estimates overlap and should not be added as independent causal effects. The drug result retains the model’s immediate, sustained acceleration from 2025; it does not simulate a gradual ten-year ramp to the predicted pace. The estimates are reference cases, not gains for every country or sex.
4.1New medicines
+8 years of lifespan57% future estimate
Our forecast: 20× faster drug progress over the next 10 years.
The target combines roughly 10× as many new medicines with twice the mortality benefit per drug. In the U.S. female reference case, the model translates that pace into about 8 additional years of median lifespan by 2037, relative to baseline.
4.1.1 · Prediction & calculation
More effective medicines could change the economics of clinical trials.
Drug contribution500 new drugs per year · 2× benefit per drug
50 is based on approximately 47 annual FDA CDER novel approvals in 2015–2024. FDA annual report.
- 2× effect size → roughly ¼ the participants
For a simple trial with otherwise fixed statistical assumptions, a larger treatment effect can be detected with a smaller sample.
- 2× faster trials
This assumption rests on AI and useful biomarkers shortening recruitment, measurement, and follow-up where the endpoint permits it.
- One Trial, Start to Finish
The envisioned pathway runs from first patients to approval in one seamless clinical trial, after preclinical work.
Trial assumptions and the link to mortality
Our output target: 10× medicines × 2× benefit = 20× drug progress.
The sample-size argument uses the approximate inverse-square relationship between effect size and sample size at fixed power, significance, and variance; see NIST’s two-sample test formulation. It does not establish a fourfold reduction in total trial cost, duration, or required safety follow-up. A trial endpoint’s effect size is also not automatically the population mortality benefit used in the calculator.
Even if participant requirements and duration translated directly into capacity, 4 × 2 would give 8×, not 10×. A seamless trial could reduce repeated setup and handoffs, but we assign it no separate numerical multiplier. Thus 10× approved-drug output remains our forecast assumption, rather than a result proved by multiplying these factors.
Our single-trial pathway is a forecast of regulatory change, not a universal FDA rule. FDA guidance describes circumstances where one pivotal trial plus confirmatory evidence can support effectiveness. That does not eliminate early clinical development, safety follow-up, or manufacturing requirements.
In the mortality model, 20× drug progress gives 1 + 0.44 × (20 − 1) = 9.36× overall mortality improvement. It is distinct from the roughly 20× overall pace needed for the age-100 target. The controls let you test the prediction’s output and benefit assumptions.
United StatesCountry-specific trend
Median lifespan in 203794.0 years
+8.0 yearsvs. baseline · 86.0
HMD observations through 2023 · projections to 2037
Data & calculation
HMD period life tables, existing 2025 data snapshot. Country follows the selection at the top of the page. Country-specific trends use 2003 onward, excluding 2020–2022. From 2025, the annual log-hazard decline is multiplied by 1 + 0.44 × (drug contribution − 1), with other progress unchanged. Median lifespan is the age where survival reaches 50%, assuming constant hazard within each age band, exactly as in the main lifespan model. Historical points are reconstructed from the observed HMD survivor curves. Shading shows the difference between scenarios, not uncertainty. The chart's upper limit is fixed at 100 years; ≥ marks a lower bound when the median reaches the unmodeled 100+ ages, not a biological ceiling. These scenarios assume immediate, sustained proportional benefits; they are not calibrated forecasts.
4.1.2 · Breakthrough timeline
Cardiometabolic progress will see the biggest gains, followed by cancer and then neurodegeneration.
- 2026–203001
Whole-system cardiometabolic treatment
Shared risks. Coordinated prevention.
Advances in obesity, diabetes, and cardiovascular care could reduce metabolic risk and prevent heart attacks and strokes.
- 2028–203202
Personalized cancer treatments
Different biology, different treatment.
Tumor-specific vaccines and better-matched therapies could make cancer treatment more personal and durable.
- 2032–2037+03
Fundamental breakthroughs in neurodegeneration
Translation remains a challenge.
Broader breakthroughs may take longer. Model organisms capture only parts of diseases such as Alzheimer’s, making results harder to translate into effective treatments.
The evidence behind this order
Historical precedent · United States, 2000–2019. These are observed changes in age-adjusted mortality, not our predictions. The combined stroke-and-heart-disease rate fell 37.7% (318.5 → 198.5 per 100,000). We sum the two rates in each year, then calculate the percentage change; we do not average their percentage declines. Heart-disease mortality fell 37.3% (257.6 → 161.5 per 100,000), stroke 39.2% (60.9 → 37.0), cancer 26.8% (199.6 → 146.2). Alzheimer’s mortality rose 64.6% (18.1 → 29.8). These are changes calculated from published 2000 NCHS rates and 2019 NCHS rates. The common pre-pandemic window avoids COVID disruption; it is not a measure of the latest momentum.
The historical stroke-and-heart-disease figure covers only part of the broader cardiometabolic category; it does not include diabetes. All-cancer mortality is broader than personalized treatments; Alzheimer’s is one neurological disease, not all dementias or all neurology. Rates use underlying causes of death, both sexes and all ages, standardized to the 2000 U.S. population. Prevention, smoking trends, treatment, and care all contribute; these are not drug-only effects. Changes in diagnosis and death-certificate reporting also affect recorded Alzheimer’s mortality. We do not treat that increase as proof that every neurological treatment failed.
The larger historical declines for stroke and heart disease provide historical context for the cardiovascular part of our first stage. Recent treatment evidence also informs our ordering. Neither historical percentage declines nor trial results establish the forecast dates.
Cardiovascular prevention: demonstrated outcomes at scale. In SELECT, 17,604 adults with established cardiovascular disease and overweight or obesity, without diabetes, were randomized to semaglutide or placebo. Cardiovascular death, nonfatal heart attack, or nonfatal stroke occurred in 6.5% versus 8.0% during follow-up: a 1.5-percentage-point absolute difference and hazard ratio of 0.80. This is a composite cardiovascular outcome, not a 20% reduction in all deaths or a measured lifespan gain. The result supported an FDA-approved cardiovascular indication.
Personalized cancer vaccines: promising evidence, further validation. The published phase 2b KEYNOTE-942 study randomized 157 patients with resected high-risk melanoma. At 18 months, recurrence-free survival was 79% with V940 plus pembrolizumab versus 62% with pembrolizumab alone. The recurrence-or-death hazard ratio was 0.561 (95% CI 0.309–1.017), so uncertainty remained substantial. V940 is listed as investigational in Merck’s August 2026 pipeline, with phase 3 development underway. This example concerns personalized vaccines, not the maturity of all cancer medicines.
Alzheimer’s: a translation challenge. Model organisms are useful, but do not reproduce all stages and features of human Alzheimer’s disease. This limits how confidently a result in mice predicts benefit in people; see the NIA workshop on modeling brain aging and Alzheimer’s disease. This is one reason for our later forecast window, not a claim that progress is impossible.
We infer near-term readiness from evidence maturity and regulatory status; these trials have different populations and endpoints, so their effect sizes cannot rank the two fields. They do not measure AI’s contribution. The dates remain our judgment, not fitted predictions. Early Alzheimer’s treatments already exist, while broader Alzheimer’s breakthroughs remain a forecast. These categories do not allocate the model’s eight-year gain.
4.2Public health
+5 years of lifespan36% future estimate
AI-driven abundance will make better public health possible. People will demand it.
Our forecast links growing capacity with rising expectations. As AI lowers the cost of expertise, goods, and services, societies can devote more resources to health. People will increasingly demand that this abundance deliver accessible care, cleaner environments, and protection from preventable illness. Public health will become a stronger social and political priority.
4.2.1 · Prediction & calculation
Wider access to top healthcare could unlock ~5 years of life expectancy.
Public health is far from saturated. At the middle of the income distribution, women live about 4 years less than the wealthiest women, and men about 5.6 years less than the wealthiest men. AI could help close much of this gap over the next decade by extending effective prevention, treatment, and follow-up to everyone.
0% of the gap closed
Bottom 1%+0.0 gained
50th percentile+0.0 gained
Sex
Life expectancy
Lower incomeHigher income
Household income percentile at age 40ObservedAfter dissemination
How to read this chart
The tint shows years gained; gray dashed brackets show the gap remaining, not years lost. Race-adjusted estimates use U.S. tax and Social Security records. The solid curve moves each income percentile toward its sex-specific top-1% level; the dashed curve preserves the observed gradient. At the 50th income percentile, the full gaps are 3.9 years for women and 5.6 for men. Their average is approximately 4.8 years, rounded to five for our narrative estimate. These are differences in expected age at death conditional on reaching 40, not median-lifespan gains. Income is associated with many health determinants; the entire gap is not known to be removable through healthcare. See the study’s interpretation. Closing it within a decade is our assumption, and gains can overlap with those attributed to medicines.
4.2.2 · Future public-health goals
AI will make futuristic public-health ventures possible.
Three ambitions for the next public-health frontier: less illness, greater independence, and cleaner environments.
End respiratory viral disease
Prevent accidents. Preserve independence.
Eliminate all pollution
Ambitions, evidence, and limits
These are future public-health goals, not quantified forecasts or an allocation of the five-year income–longevity estimate above. Their value includes less illness, disability, and disruption, alongside any mortality benefit. AI could help tailor prevention, coordinate services, and improve monitoring; reaching these goals also requires infrastructure, policy, and broad access.
Respiratory viral disease: ending all such disease is an aspirational endpoint, not a demonstrated capability or a prediction of eradication by 2037. WHO guidance on ventilation supports cleaner indoor air as part of reducing respiratory infection risk. The protection available from vaccines and treatments differs across viruses; no single intervention eliminates all transmission.
Accidents: preventing all accidents, including car crashes and falls, is an aspirational goal, not a prediction for 2037. For falls, WHO’s Step Safely guidance supports strength and balance training and safer environments. Smarter assistive devices are a future opportunity, not a guarantee that all falls can be prevented.
Pollution: eliminating all pollution is our long-term ambition, not a prediction of zero pollution by 2037. It includes particulate, noise, and light pollution as well as water and soil contamination. Electrification paired with clean generation can reduce combustion-related air pollution; see the IEA’s overview of electricity and clean-energy transitions. Electrification alone does not remove all pollution: tire and road dust, agricultural and industrial emissions, noise, excess artificial light, and water and soil contamination require additional measures. Clean power matters because fossil-fuel electricity generation can shift emissions upstream.
4.3Other medical care
+1 year of lifespan7% future estimate
Advanced diagnosis will pinpoint rare conditions and enable earlier care as therapeutics advance.
AI could make scanning, monitoring, and clinical decisions faster and more precise, shortening the search for a diagnosis, including for rare conditions, and starting care sooner. Earlier diagnosis matters when it changes treatment: screening trials show where it already helps, and where better treatments are needed for it to count. We assign this category roughly one year of lifespan, a judgment rather than a result calculated from screening trials.
Heart disease: the tools exist, the challenge is reach.
A major near-term opportunity is making existing prevention work reliably through blood-pressure and cholesterol control, risk reduction, and follow-up. Additional screening has shown modest or uncertain survival benefits so far.
Two cardiovascular screening trials
DANCAVAS · NEJM, 2022
In men aged 65–74, comprehensive screening produced 12.6% versus 13.1% all-cause mortality over median 5.6 years; the difference was not statistically significant (p = 0.06).
VIVA · Lancet, 2017
Vascular screening followed by intervention in men aged 65–74 produced 10.2% versus 10.8% mortality over median 4.4 years, a modest, significant benefit. These are screening-and-care results, not effects of diagnosis alone.
Cancer: early detection could stop it before it turns invasive.
Finding cancer sooner could allow interception, treating abnormal cells before they become invasive, and microscopic surgery that removes tumors while they are still tiny. Today, earlier detection makes a substantial difference for selected cancers when effective treatment follows. Finding more early-stage tumors does not, by itself, prove that lives are saved.
Two screening trials
NELSON · NEJM, 2020
CT screening in high-risk smokers reduced lung-cancer mortality by 24% at 10 years in the primary male population. All-cause mortality was not significantly reduced.
UKCTOCS · Lancet, 2021
Ovarian screening shifted diagnoses toward earlier stages but did not significantly reduce ovarian-cancer deaths over median 16.3 years. Surgery was available; earlier detection was not enough.
Neurodegeneration: waiting on better treatments.
Treatments can improve symptoms or slow decline, and earlier diagnosis would matter more if disease-modifying therapies improve. For common neurodegenerative diseases, it has not yet established a lifespan benefit.
Two early-treatment trials
LEAP · NEJM, 2019
Starting levodopa 40 weeks earlier in Parkinson’s disease did not demonstrate disease modification over 80 weeks. This was not a mortality trial.
CLARITY AD · NEJM, 2023
Lecanemab modestly slowed cognitive and functional decline in early Alzheimer’s over 18 months, with important adverse effects. It did not establish longer lifespan or test population screening.
5Multimorbidity
Late-life function will become the next medical frontier.
Mortality gains have been smaller at the oldest ages, when chronic conditions are often comorbid. Unlocking high function late in life will require a paradigm change, such as organ replacement or stem-cell rejuvenation.
5.1 · Historical evidence
Mortality improvement over the past 25 years is lowest at old ages.
1999–2024
Mortality improvement % reduction in five-year probability of death
29% vs 6%
Relative reduction since 1999: ages 80–84 versus 95–99, women in Sweden.
5.2 · Cohort evidence
Chronic conditions accumulate in later life, complicating care.
Sweden · nationwide 1920–1922 birth cohorts
MenWomen40 conditions assessed
Mean chronic conditions per surviving person
274,108
People at age 70, followed through later life. Diagnoses are retained after first detection.
How to read these two studies
The left chart compares mortality at fixed age bands in the selected country’s first and latest complete observed years, starting from 1999 where available (South Korea begins in 2003). Improvement = 100 × (1 − q_last / q_first), using each age band’s period five-year death probability. Intermediate years do not enter this endpoint calculation. Date ranges vary; these are total changes, not annualized comparisons. 2026 is not observed. The right chart always describes Sweden and follows birth cohorts as they age; the country selector does not change it. They are not linked patient data. Sweden’s condition counts exclude primary care, begin with a short diagnosis lookback, and gain specialist outpatient coverage in 2001, when these cohorts were 79–81. An expanding recorded history and changing coverage can affect the age slope. The average at each age describes the people still alive, so survival selection also matters.
6Historical calibration
Is any of this possible? A historical perspective.
20× faster mortality improvement would dramatically surpass all historical records in our analysis. Compare the fastest sustained direct lifespan gains in our historical sample with 20× faster mortality improvement, starting from the United States’ observed mortality profile. For EOD 2037 to be attainable, AI must provide historically unprecedented scientific progress and growth.
6.1 · Historical benchmark
Historical records versus 20× progress.
United States · median lifespan
Historical periods, data & methods
These are period medians, not predictions of how long today’s newborns will live. “Best” refers to 17 selected countries during 1960–2019. The 20× line multiplies recent baseline mortality improvement, not the historical record for direct lifespan gain. ≥ marks a lower bound at the unmodeled ages above 100.
Direct median records: men, New Zealand 1996–2006 (+0.385 years/year); women, Finland 1968–1978 (+0.372 years/year). Japan and South Korea remain eligible; countries were not excluded because they supplied large gains.
Each historical fit uses 10 elapsed years (11 annual observations). Period median lifespan is fitted directly with OLS, and the record annual gain is applied linearly from the United States’ 2023 observed median.
The rebound screen requires five preceding years. Lifespan windows are rejected for any annual fall greater than 0.5 years over the prior five years or fitted window, or a starting value more than 0.25 years below the preceding trend. These heuristic thresholds reduce rebounds; they do not certify the absence of wars, epidemics, reporting changes, or catch-up effects. Results depend on the thresholds and selecting maxima is optimistic.
The 20× curve uses the same sex-specific model as the main lifespan projection: the United States’ latest observed age-specific hazards, trends fitted from 2003 excluding 2020–2022, and acceleration from 2025. Fitted hazard increases are clipped to zero. This chart shows each sex separately, rather than the main chart’s average of male and female medians.
Source: Human Mortality Database, via Our World in Data. COVID observations remain visible but are outside the 1960–2019 record search. Curves end at the age-100 modeling boundary; no mortality tail above age 100 is invented.