NOT BECAUSE OF AI, home

Measured AI benefits · as of

+7%TB cases found per camp with AI-picked sites, not significant
receipt
Camp Positivity Yield, was 7% higher in the intervention arm relative to the control (adjusted RR: 1.07, 95% CI: 0.94 – 1.22Results, intention-to-treat analysismedRxiv preprint: Geographical targeting of active case finding for tuberculosis in Pakistan using artificial intelligence software (SPOT-TB): a pragmatic stepped wedge cluster randomized control trial (Mahfooz et al.) · medrxiv.org

Sending Pakistan's mobile TB X-ray camps to sites picked by AI did not find significantly more tuberculosis per camp than sites picked by field staff, across 3,936 camps and 269,254 people screened, in the trial's primary analysis.

receipt

A stepped-wedge cluster RCT (medRxiv preprint May 2026, published in The Lancet Global Health in November 2026) moved 30 mobile chest X-ray van teams in 68 Pakistani districts, in random order, from staff-chosen camp sites to sites recommended by MATCH-AI, a Bayesian TB-hotspot model, between August 2023 and September 2024; in the intention-to-treat analysis, bacteriologically confirmed TB found per camp did not differ significantly, and no secondary outcome did either.

Sources

Editor's notes

Filed under not-because on the registered primary analysis (intention to treat). The per-protocol result is real but chosen after randomization by adherence, so it cannot carry a "because" headline; the paper's own interpretation leads with it. The unadjusted means run the other way (0.85 vs 0.94 per camp); the adjusted 1.07 accounts for time trends in a stepped-wedge design. Both are shown. Evidence A argued: registered on ClinicalTrials.gov on 30 August 2023, before the randomized switches began in October 2023, with yield of TB cases per camp as the aim. The peer-reviewed version (Lancet Global Health, November 2026; abstract read) reports the same ITT and per-protocol figures and the same camp period; the numbers here are still quoted from the preprint's full text. Window pre: camps ran August 2023 to September 2024. verify: whether the Centre for Global Public Health Pakistan or its staff developed MATCH-AI (the paper calls it "commercially available software" and cites its model description without naming the developer).

Related

History

  • Updated: the trial is now published in The Lancet Global Health (12 November 2026); its abstract reports the same ITT RR 1.07 (0.94-1.22), per-protocol 1.32, camp period and 269,254 screened. Added as a source, the not-peer-reviewed caveat dropped, Global Fund funding added; no value changed.
  • Added. Weekly run; read in full from the medRxiv preprint and checked against the ClinicalTrials.gov record.

Verification

Drafted by
Anthropic claude-opus-5-5
Verified by
OpenAI gpt-6.1-sol · pass
How sure
Strict
Record
data/verifications/spot-tb-ai-camp-targeting-null/2026-11-15.json · The sources support the reported screening totals and a nonsignificant 7% increase in the randomized intention-to-treat analysis. The 32% increase concerns selected adherent camps and does not establish an intention-to-treat benefit.

Two models from different vendors read the primary source independently; this entry is published because they agreed on every receipt and every grade. The models grading an AI registry have an interest in the answer. Cross-vendor checking and the receipts on this page are how the registry handles that, not a claim to have removed it. How the gate works · Dispute this entry

Cite this entry

Permalink https://becauseofai.org/entry/spot-tb-ai-camp-targeting-null“Sending Pakistan's mobile TB X-ray camps to sites picked by AI did not find significantly more tuberculosis per camp than sites picked by field staff, across 3,936 camps and 269,254 people screened, in the trial's primary analysis.” Because of AI, updated Nov 15, 2026. https://becauseofai.org/entry/spot-tb-ai-camp-targeting-null