Methods · How we read the day
Two frameworks. One scan.
CAIHL frames the system. CLAIM equips the patient.
Every item on the Daily Scan is annotated through CAIHL — Critical AI Health Literacy — a four-dimension lens that asks who the AI serves and which way patient agency moves. Alongside it, CLAIM — Contextual Literacy for AI in Medicine — defines the patient-side competencies required to use AI health outputs safely. Scholarly anchors additionally carry a source-integrity verdict (ASSAY) on the underlying paper.
Framework · Every item
CAIHL — Critical AI Health Literacy
CAIHL was articulated by Hugo Campos and Liz Salmi in a 2025 National Academy of Medicine paper to give patients, clinicians, and policymakers a shared vocabulary for evaluating AI in health. It is a lens, not a verdict. Every item on the scan is annotated on four dimensions; the dimensions compose into an agency direction for the item: expanding, constraining, or neutral.
primaryUser
- patient
- clinician
- mixed
- na
Who is the human directly interacting with the AI? The patient at home with ChatGPT is a different primaryUser from the clinician using an EHR copilot, and the same model deployed in both contexts is structurally a different artifact.
hosting
- institutional
- public
- government
- na
Where does the AI live? A health-system chatbot behind a login is institutional; a consumer chatbot on the open web is public; a state-deployed triage line is government. Hosting determines the governance surface and the consent envelope.
interests
- patient-aligned
- patient-directed
- patient-facing
- institutional
- commercial
- mixed
Whose interests does the AI advance? Patient-facing is not the same as patient-aligned. A commercial chatbot can be patient-facing without being patient-aligned; an institutional governance framework can be patient-aligned without being patient-directed.
agency
- expanding
- constraining
- neutral
Does the AI expand or constrain patient agency? An aid-in-dying tracker expands agency (more states, more options). A chatbot that poses as a doctor constrains it (the patient cannot evaluate the impersonation). The most consequential dimension is the one CAIHL was built to surface.
How CAIHL reads in the scan
Each item shows a one-line CAIHL chip directly under its summary — for example,
patient · public · commercial · constraining.
Clicking the Full CAIHL read link on any item opens a deep page laying out all four dimensions and the
editor's note for that specific artifact.
Framework · Patient-side competency
CLAIM — Contextual Literacy for AI in Medicine
CLAIM is a competency framework for patients engaging with AI health outputs. Its core premise: AI output requires patient context to be meaningful, and patient context requires structured competency to be applied correctly. CLAIM operates at the intersection of four domains that must be addressed from the patient perspective.
Contextual Interpretation
Training patients to filter AI outputs through their own clinical history, comorbidities, medications, and priorities.
Source and Confidence Evaluation
Teaching patients to assess the basis, limitations, and uncertainty of AI health outputs.
Interaction Quality Audit
Enabling patients to recognize when AI outputs lack sufficient context to be actionable.
Equity Awareness
Building awareness of how AI bias manifests differently in patient experience than in population-level metrics.
CLAIM does not evaluate AI systems. It evaluates the human receiving AI output.
Audit · Scholarly anchors only
ASSAY — source-integrity verdict on the underlying paper
When a scan item is a peer-reviewed journal article or a preprint, we additionally run a structured source-integrity audit on the paper itself. ASSAY surfaces the paper's central claim, counts the supporting claims, and assesses methodology, citation network behavior, generalizability, and the gap between data and framing. The output is a one-word verdict and an integrity note.
Scope. ASSAY runs only on scholarly sources (medRxiv, bioRxiv, arXiv, JAMA, NEJM, BMJ, npj Digital Medicine, JMIR, and similar). News coverage of a study is not the study; opinion pieces, government press releases, and product launches are not appropriate ASSAY targets. When a news article references an underlying paper, ASSAY runs on the paper and the news item links to that deep page.
Verdicts
sound
Methodology is appropriate to the question, the key claim is supported by the data presented, citation network is non-pathological, no significant inheritance of un-audited claims. Limits are noted but do not undermine the central finding.
mixed
Central finding is credible but the integrity assessment surfaces material caveats — generalizability gap, selection confounding, single-site sample, observational design where causal language appears. Treat as hypothesis-generating rather than prevalence- or effect-establishing.
flagged
One or more integrity issues identified that materially weaken the central claim — methodology mismatch, key claim outruns the data, citation cascade pattern, evidence sparse relative to framing strength. Use with explicit caution.
- problematic
- cascade
The strongest negative verdicts. Problematic: the claim does not survive integrity assessment. Cascade: the claim's authority is manufactured by citation dynamics rather than supported by primary data.
How ASSAY reads in the scan
ASSAY-eligible items carry a verdict card directly under the CAIHL chip — for example,
ASSAY sound · 9 claims.
The deep page opens with the key claim, the count of supporting claims, and the integrity note for that paper.
Why both
CAIHL frames the system. CLAIM equips the patient. They answer different questions.
CAIHL asks who the AI is serving and which way patient agency is moving — a lens on the system. CLAIM asks what competencies the patient needs to use AI health output safely — a curriculum for the patient. The same scan item can be CAIHL-constraining (the chatbot narrows patient agency) and still be a CLAIM-actionable artifact (the patient who applies Contextual Interpretation can use it usefully anyway). The two frameworks are not redundant; they are orthogonal. ASSAY runs underneath both, on the scholarly anchors.