AI improved discourse

How cheap models and cheap verification change what an argument is worth

J

AI can now do two things public arguments have always needed: it can put a mathematical model into a debate, and it can validate the model and the citations another party brings. That is the case for saying this technology can improve public discourse rather than merely flood it. This essay tries to demonstrate the point instead of asserting it: it builds a working model of a live policy, publishes the code, and then turns the same capability on the model and the sources the policy's opponents put in the room.

The worked case is Alberta's self-referred, private-pay diagnostic imaging regime. It is small, concrete, and contested, which makes it a fair test. The model decomposes the screened population into findings, referrals, cancers, and false alarms; prices the follow-up through the public fee schedule; and reports the capacity it consumes — family-doctor visits, emergency-department encounters, laboratory tests, and further imaging — and the per-capita cost to everyone else. The model's code is downloadable and every figure below is recomputed from it.

The law and its stated rationale

From 2026-07-31, Alberta residents can buy their own MRI, CT, ultrasound, and X-ray scans without a physician's referral, and the public system reimburses only a new cancer diagnosis — "if no cancer is diagnosed, the cost of the test is not reimbursed" (Alberta Health, 2026-07-31). The government's stated rationale is on the same page: a claimed 62% of Albertans want more private supply, the private sector is said to build capacity that "doesn't require any tax dollars," and private clinics are said to "ease pressure on the public system, reducing waiting times." Those are three testable claims. The first is a poll; the other two are arithmetic, and arithmetic is exactly what a cheap model can now check.

The Alberta Medical Association's objection is also concrete. Its statement names the "full body scan" and warns that incidental benign findings will drag a wave of follow-up through the public system (AMA statement). Its environmental scan puts numbers on the fear: a claimed cost of $784.15 per test, an assumed ~30% positive-finding rate, 3,315 scans/yr (13 units, about one a day, 255 days), and a resulting ~$911,000/yr, with fee codes 03.03A, 03.04G, 03.08A, and 03.03F and a 78.2%/21.8% GP/ER split (AMA environmental scan). CBC reported the same objection on 2026-07-27 (updated 07-28), with a provider estimating private-test volume could reach 50,000–60,000 a year against roughly 7 million total exams (CBC, 2026-07-27, updated 07-28).

False positives and false negatives, defined by the law

A screening test only means something once "positive" is defined, and here the law defines it for us. A test is positive when it leads to a reimbursed new cancer diagnosis. A false positive is therefore a finding that triggers work-up but is not cancer, and a false negative is a cancer the scan misses and that surfaces later. The pooled whole-body MRI literature reports the false-positive-for-cancer rate directly: an actionable non-malignant finding in 14.8% of screened people (95% CI 6.6–30.0), against a malignancy rate of 1.4% (95% CI 0.9–2.2) (J Gen Intern Med, 2026). The phrase actionable non-malignant is the study's own term, and it is quoted here on purpose: these are real findings that required real follow-up; they are simply not cancer, so under this law they are false positives.

What the programme finds: the layered yield

The headline number in the debate is that whole-body MRI "finds something" in 78.0% of people (95% CI 44.9–93.9). That is not a result. It is a finding reported — the study's own conclusion is that most such findings need no follow-up. The yield is layered: 78.0% any finding → about 16.2% referred for work-up (14.8% false positive for cancer + 1.4% cancer) → 1.4% cancer and 14.8% false positive. Of the false positives, 58.2% (95% CI 36.0–77.5) went on to further labs or imaging and 20.8% (95% CI 9.7–39.2) to a procedure (same meta-analysis).

Set that against the expected true yield. The background cancer rate in Alberta is about 24,600 projected new cancer diagnoses in 2026, and 254,100 in Canada (Canadian Cancer Society / CMAJ, 2026-04-13). Against Alberta's population of 5,101,050 (Statistics Canada Table 17-10-0009-01, 2026-07-01), that is an expected 0.48% a year. The observed pooled detection is 1.4% — about three times the annual background rate. A screening exam plausibly catches prevalent cancers as well as incident ones, so the comparison is a sanity check on the yield, not a verdict on it. A second pooled estimate puts cancer detection at 1.57% (Eur Radiol, 2026). No trial yet connects screening to outcomes — the Cochrane review is still a protocol (Cochrane protocol). The model below therefore measures cost and capacity, never mortality benefit.

What the false positives are, and what they cost to chase

The common actionable non-malignant findings after whole-body imaging are ordinary and familiar: thyroid nodules, benign liver and renal cysts, disc degeneration, and sinus disease. Their standard follow-up is a repeat ultrasound, a targeted CT, occasionally a biopsy, and a specialist referral. The meta-analysis's full text is behind a bot wall, so this enumeration is a labelled reasoned estimate drawn from the whole-body MRI screening literature rather than a quotation from the paper; the arithmetic does not depend on it, only the specificity of the prose does. The pooled rate — 14.8% of everyone screened — is what the model uses, and it is sourced.

The follow-up is where the policy's claimed savings go to work in the other direction. At the AMA's own anchor of 3,315 scans a year, the pooled decomposition means about 490.62 false positives, of which about 285.54 pursue further labs or imaging and about 102.05 a procedure. That is the true-positive/false-positive split the law creates, computed rather than assumed.

Capacity, not only money

The load-bearing claim of the policy is that private supply "eases pressure on the public system." Whether it does is a capacity question, and the honest answer is that the programme adds demand to the same constrained system it claims to relieve. The baselines matter because the added demand lands on top of them:

• Family doctors. 16% of Albertans have no family doctor (down from 18% in 2025), 7% are actively searching for one, and Canada fields 119 family physicians per 100,000 residents (AMA State of Health Care, 2026-06; CIHI physicians profile, 2025-10-30). Every false positive routed to a family doctor competes for a panel slot that may not exist.
• Emergency departments. Alberta records roughly two million ED visits a year (Alberta Health Services) against 15.6 million unscheduled visits nationally in 2025–26 (CIHI NACRS, 2026-09-17). The ER share of the index visits is small, but it lands on the most saturated venue.
• Laboratories. No clean public per-year Alberta lab-volume figure surfaced. This is a labelled estimate: tens of millions of tests a year (population × roughly 10–20 tests per capita). It is context only — the model's load-bearing laboratory output is the added test count, not the provincial total.
• Other imaging. The CIHI MRI wait-time indicator tracks 50th- and 90th-percentile days in reporting provinces including Alberta (CIHI, data updated 2026-06). A private scan that finds something turns into a public scan that also has to be scheduled.

At the anchor, the model's added demand is small in absolute terms but pointed: about 807.55 family-physician visits, 117.07 emergency-department encounters, 280.46 physician-hours, 571.08 laboratory tests, and 119.93 additional imaging studies a year. The essay reports these as added demand; the wait-time effect is a bounded estimate, never a point claim, because the queue length depends on baseline capacity the model does not have. The point is structural: the money is the smaller half of the story. The impact is capacity, not only money.

Who uses it, and how many

The people who buy a private scan at these prices are the higher-income and the worried well — the programme is not means-tested, and the public system carries the follow-up. How many people take it up is the largest uncertainty in the whole analysis, so it is reported as three take-up scenarios reconciled to the same formula, low / base / high:

• Low — 3,315 a year. The AMA's own one-a-day, 13-unit scenario (AMA environmental scan).
• Base — about 15,000 a year. A labelled reasoned estimate: a low single-digit fraction of adults, working default 0.3% of the provincial population, chosen as the midpoint of a 0.1%–0.5% band that straddles the AMA low.
• High — 50,000–60,000 a year. The CBC's provider estimate against roughly 7 million total exams (CBC, 2026-07-27, updated 07-28).

A reproducible model, built bottom-up

Cost is composed from the decomposition, not assumed as a whole. For N scans at a referred rate p (default 16.2%):

referred = N·p · falsePos = referred·(0.148/0.162) · truePos = referred·(0.014/0.162) · labArm = falsePos·0.582 · procArm = falsePos·0.208
C_index = 0.782·feeGP + 0.218·feeER · C_labs = fBlood·cBlood + fScan·cScan · C_proc = cProc
annualCost = referred·C_index + labArm·C_labs + procArm·C_proc

Every unit price with a public fee-schedule value is sourced to the AHCIP Schedule of Medical Benefits (SOMB, 2026-04-01), corroborated per code through the AMA Fee Navigator (Fee Navigator): family-physician visit 03.03A at $25.09, emergency visit 03.04G at $102.50, laboratory code E1 at $18.53, and generic minor procedure 13.99BB at $65.44. The bloodwork/scans split of the follow-up arm and the additional-scan price have no public source — the SOMB defers radiology to a section with no CT/MRI fee — so both are labelled modelled estimates: the additional-scan price is ranged, while the follow-up split is a point estimate. The procedure arm is a further sensitivity: it is priced at one generic minor-procedure fee (13.99BB), while the source's "pursued a subsequent procedure" arm spans biopsies, endoscopies, and excisions that can cost far more, so the reported annual range covers only the additional-scan unit and not the real procedure mix. The parameter table states the kind and basis of every value.

Auditing the AMA's own model

The AMA's two headline figures do not reconcile under the scan's own framing: $784.15 × 3,315 ≈ $2.6M ≠ $911,000, and $784.15 × (3,315 × 0.30) ≈ $780k ≠ $911,000. The per-test figure and the annual figure are not the same quantity, and the scan does not say so. That is the citation-and-model check the essay is arguing for: two numbers, one arithmetic, and a gap the reader would otherwise absorb as authority.

Rebuilt bottom-up at the AMA's own anchor (N = 3,315, referred rate 16.2%), the model returns $92,617.53 a year — 537.03 people referred, of whom 490.62 are false positives, at a blended index-visit cost of $41.97 and follow-up arms of $222.04 (labs) and $65.44 (procedure). That is $818,382.47 below the AMA's ~$911,000. The model is not tuned to close the gap, because closing it would require inventing follow-up the AMA does not publish: for the AMA's figure to hold, each referred person would have to cost $1,696.37, several times what the sourced rates and labelled shares produce. Over the modelled additional-scan range the annual figure runs from $47,644.84 to $332,471.83. The decomposition counts and capacity at the anchor are recomputed from the model in the fallback block below.

Interactive: cost and capacity as the sliders move

Move the take-up N and the finding rate. The annual cost and the physician workload are computed live by the same model.mjs the tests run; the anchor point is the initial position (N = 3,315, finding rate = 0.162).

Annual cost now: — · Physician workload: — · Added lab tests: — · Added imaging: —

The per-capita cost to everyone else

Dividing the annual cost by Alberta's 5,101,050 residents gives the cost the public system carries on behalf of everyone, at each take-up scenario. It is small in dollars — this programme is a slice of the system, not a reordering of it — but it is the figure the policy's "doesn't require any tax dollars" claim has to answer to, because the follow-up is publicly paid.

Annual cost and per-capita cost follow N·p through the decomposition and the sourced unit prices; they are recomputed from model.mjs for every row.

How this was made

This essay and its model are an AI-assisted build: an AI agent drafted the model and the prose under review, and the verification rules were fixed in advance. Every unit price is either sourced to the public fee schedule or labelled an explicit assumption; every figure in the page is recomputed from the model's own code by a script; the page ships that code as a download (model.mjs); every outbound link was checked to return HTTP 200; and the model was never tuned to reproduce the AMA's anchor. Where the public sources disagree — the $784.15 and $911k figures — the disagreement is printed rather than smoothed over. The follow-up shares and the additional-scan unit are the model's weakest assumptions; they are labelled — the additional-scan unit is ranged, the follow-up split is a point estimate. The AI did the producing and the checking; the judgment about what is worth modelling, and which definitions the law actually fixes, stayed with the author.

Sources

• Alberta Health, preventative health testing (policy live 2026-07-31; stated rationale and the cancer-only reimbursement).
• Alberta Medical Association, statement on privately purchased diagnostic screening and testing services.
• Alberta Medical Association, environmental scan of self-referral private-pay lab and diagnostic imaging ($784.15/test, 3,315 scans/yr, ~30%, ~$911,000/yr, fee codes, 78.2%/21.8% GP/ER split).
• Alberta Medical Association, State of Health Care 2026 (16% without a family doctor, down from 18% in 2025; 7% actively searching).
• CBC, pay-preventative-diagnostic-tests-alberta (posted 2026-07-27, updated 07-28; 50,000–60,000 provider estimate).
• J Gen Intern Med 2026, pooled whole-body MRI findings (78.0% / 14.8% / 58.2% / 20.8% / 1.4%, with CIs).
• Eur Radiol 2026, pooled cancer detection (1.57%).
• Cochrane, whole-body MRI screening protocol (no outcomes yet).
• Canadian Cancer Society / CMAJ, 2026 cancer statistics (24,600 new Alberta diagnoses; 254,100 Canada).
• Statistics Canada, Table 17-10-0009-01 (Alberta 5,101,050 at 2026-07-01).
• CIHI, physicians in Canada (119 family physicians per 100,000, 2024).
• Alberta Health Services, emergency department visits (approximately 2 million Alberta ED visits a year).
• CIHI, NACRS emergency-department visits and lengths of stay (2026-09-17; 15.6M national 2025–26).
• CIHI, wait times for MRI scan (data updated 2026-06).
• Alberta Health, Schedule of Medical Benefits, 2026-04-01 (fee values).
• Alberta Medical Association, Fee Navigator (per-code corroboration).
• The essay hub at jokerc.ca.

Two private-pricing figures — a full-body MRI at roughly CAD 2,500–3,500 and a self-refer MRI at $600–1,200 — are reported but unverified; they shape only the modelled scan range, not a sourced claim.

Colophon

Written by J. Essay text is licensed CC BY-SA 4.0, credited to J. Plain HTML, CSS, and vanilla JavaScript; no framework, no build pipeline, no external fonts. The site sets no trackers, no cookies, and no ads. The model is served as a static download: Download model.mjs, the same file the page imports and the tests run.