QWE AI Academy Two ways people handle the headline that universal health coverage could save $1T and 114k lives a year: study. Path A: screenshot the Yale number, argue in the comments, move on. Path B: feed the preprint into an LLM, force base case apart from sensitivity, and leave knowing which assumptions carry the result. Path B is about twenty minutes. Harder to spin afterward.
The July 2026 Yale School of Public Health preprint (still un-peer-reviewed) tore through Hacker News – hundreds of points, long threads on hospital margins and primary-care capacity. Same $1.04 trillion and 114,000 lives kept circulating without the fine print. I ran the paper through ChatGPT, then Claude, like a dataset. Below is the workflow that held up.
What the model is actually doing
Net claim first: against 2024 National Health Expenditure, the single-payer run (Medicare for All Act-style) drops spending from $5,278.6B to $4,237.4B. That is a $1,041.2B cut – 19.7% – after new care for the uninsured and underinsured is added back in. Alison Galvani and coauthors (Yale SPH + University of Maryland) published the methods on medRxiv; the Yale announcement matches the headline figures.
Gross cuts land near $1,345.2B. Drugs via international reference pricing (~$377.5B, about a 51% pharma step in the base case). Medicare-level provider rates (~$295.6B). Admin (~$286.3B). Fraud (~$285.7B). Avoidable ED/inpatient (~$100B). Offsets: +$304.0B (utilization $197.7B, dental $54.7B, unpaid care $51.7B). Readers who only hear “$1T less” miss that expanded access already eats roughly one-fifth of the gross. Full decomposition sits in the medRxiv preprint.
Mortality is not one lump. Base block vs today’s system: 62,863 lives (33,232 uninsured at HR 1.40 + 29,631 underinsured at a modeled γ=1.25). Another ~51,311 from reversing 2025 coverage retractions → 114,174 total. Underinsured deaths are interpolated – authors state direct excess-mortality estimates for that group are unavailable – so γ sensitivity 1.00-1.40 swings the current-system component from 33,232 to 79,367. Headlines almost never print that range.
Static one-year books also skip transition costs, administrative job losses, and how providers behave if paid at Medicare rates. Conservative runs (VA-style pharma or no pharma cut; no fraud cut) still show ≥$663.3B (12.6%) savings. Hospital-margin pushback on HN is sharper: thin operating margins (about 0.5-4%) make large rate cuts look like layoff pressure, even when the national spreadsheet stays green. Primary-care supply absorbing fewer avoidable ED/hospital visits under lower fees is simply unmodeled – an open gap, not a hidden rebuttal.
Compared with their earlier 2020 Lancet analysis (~$450B and 68k+ lives), this pass adds underinsurance, dental, a wider commercial-vs-Medicare gap, and the 2025 retraction layer.
Step-by-step: interrogate the study with an LLM
Paste the abstract and a short methods/results chunk first. Whole PDF on turn one fills the window and the model goes vague.
- Lock the source. “You are analyzing only the July 2026 medRxiv preprint by Pandey, Wells, Ye, Fitzpatrick, Galvani on US universal healthcare. Do not invent numbers. If a figure isn’t in the text I provide, say ‘not stated.’ Here is the abstract and Results section: [paste].”
- Force the decomposition. “List every savings line and every offsetting cost with the exact dollar figures from the paper. Show gross reduction, offsets, and net.” Target recovery: ~$1,345B – $304B = $1,041B.
- Surface the sensitivity headlines skip. “Extract the underinsured hazard-ratio sensitivity and the expenditure scenarios that drop pharma or fraud savings. What is the lowest net savings they still report?” You want the $663.3B floor and the γ band called out.
- Mortality split. “Break 114,174 into (a) current uninsured, (b) current underinsured, (c) post-2025 retraction reversal. Mark observed vs modeled.”
- Adversarial pass. Drop a hard HN-style critique (Medicare rates vs 0.5-4% hospital margins; PCP capacity). Ask which concerns the paper answers, which it lists as limitations (transition costs, job losses, provider response), and which sit outside the model entirely.
- Personal translate (optional). “I am a [age] [insured/underinsured] adult in [state] with [high deductible / chronic meds]. Using only mechanisms in this paper, which channels most change my out-of-pocket risk? No medical advice.”
Prompt starter you can copy:
Role: critical research assistant.
Source constraint: only the 2026 Galvani et al. medRxiv preprint + the text I paste.
Task 1: Net savings table (gross buckets vs offsets).
Task 2: Lives table with HR assumptions and γ range 1.00-1.40.
Task 3: List author-stated limitations in their own framing.
Task 4: Flag any popular-news claim that overstates certainty.
If unknown, write UNKNOWN - do not fill gaps.When the tables return, open the PDF for the two or three lines the model leaned on. That five-minute check is the difference between analysis and vibes.
Pro tip: Demand a “kill shot” table – base case | authors’ conservative bound | one harsher assumption you add (e.g., half the avoidable-ED line). If direction holds, the claim is not a knife-edge.
Common pitfalls when you AI the numbers
Models round 114,174 to “over 100,000” and then treat the round number as exact. Keep the paper’s digits until you choose to round.
They also smear the 62,863 “vs today’s system” block into the full 114k that includes reversing recent retractions. Different counterfactuals. Pin one baseline per prompt.
Garbage-in still wins: a news recap is not the preprint. Use medRxiv full text or the Yale write-up that quotes the five buckets and the limitations paragraph – not a viral card.
How this beats the usual alternatives
| Approach | Speed | What you learn | Main failure mode |
|---|---|---|---|
| Headline + Twitter/HN fight | Fast | Claim + vibe | No sensitivity, no offsets |
| Full PDF cold | Slow | Everything, if you finish | Most people bounce at methods |
| News explainer only | Medium | Polished narrative | Soft-pedals modeled HRs and omissions |
| LLM + source-locked prompts (this guide) | Medium-fast | Decomposition, ranges, limitations | Hallucination if you skip PDF spot-checks |
The AI path is not magic. It is a forcing function: you see the $304B offsets and the γ range before you pick a side. PDF-close reading stays the citation gold standard; the model just stops jargon from winning on pass one.
Does a static accounting exercise ever fully settle a fight this large? Probably not – and that is fine. You leave with the authors’ published numbers, not the ones that flatter your prior.
FAQ
Is the $1T figure peer-reviewed yet?
No. July 2026 medRxiv preprint. Serious model, not final gospel.
Why do some write-ups say ~68k lives and others 114k?
Different papers, different baselines. The 2020 Lancet-linked estimate (~68.5k) used that era’s system. The 2026 run is a new stack – today’s uninsured/underinsured block plus a 2025-retraction layer. If the chat merges them, wipe context and paste only the 2026 methods.
Can ChatGPT tell me whether Medicare-for-All would help my family specifically?
It can map mechanisms in the Act-style model (no cost-sharing in the write-up, drug reference pricing, dental add-on) onto a situation you describe, and it can flag costs that look like the underinsurance channel the authors stress. It cannot forecast your claims, post-transition networks, or tax incidence – the paper is national and static. Use it to sharpen questions for benefits counselors and primary sources, not as a household calculator.
Open the LLM. Paste the starter. Drop the medRxiv abstract. Run sensitivity before the hot take.