What AI does — and what it won't do.

The boundary in one line

AI assists with language and structure. A registered pharmacist is responsible for what's clinically true on every reviewed page. AI never signs off.

What AI does on AllMeds.ai

• Drafts prose from sourced clinical points, in plain language a non-clinician can follow.
• Structures pages — headings, FAQ formats, what to cover.
• Tightens language for readability without changing clinical meaning.
• Summarises long source documents (e.g. regulator product information) for editorial use.
• Powers interactive tools like the drug interaction checker, with responses constrained to vetted source material and clearly labelled as informational.

What AI does not do

• Decide what's clinically true. A claim must trace to a human-authored source (regulator, guideline, peer-reviewed paper).
• Diagnose, prescribe, or replace your clinician. Tools and pages are general information; they're not advice tailored to your health.
• Set dosing or classify severity unsupervised. Any severity or dose information on a reviewed page came through pharmacist review.
• Sign off on a published page. Every "Reviewed by" badge is a real review by a real pharmacist.
• Recommend stopping or starting a medication. Tools never tell you to change your prescription. They tell you when to talk to a clinician.

How our interactive tools stay safe

Tools like the drug interaction checker run on a constrained AI workflow:

1. Source-grounded answers. Responses are anchored to regulator product information, national clinical guidelines, and our internal drug interaction database — not unrestricted model output.
2. Citation by default. Each interaction call includes the source (TGA, FDA, MHRA, NICE, etc.) so you can verify the answer.
3. Severity language is conservative. When the evidence is uncertain or thin, we say so rather than overstating risk.
4. Red-flag escalation. High-severity outputs are flagged with explicit "talk to your pharmacist or doctor before continuing" guidance and emergency information where relevant.
5. Pharmacist review of the workflow. Tool output patterns are periodically reviewed by Luke to catch failure modes and tune the constraints.
6. No personal medical advice. Tools answer the question you asked using general information; they explicitly do not factor in your individual medical history, lab results, or other treatments.

Hallucination — what we do about it

Large language models can confidently generate text that isn't true. We accept that. The mitigations we use are:

• Sourced content. If a claim doesn't trace to a real source, it doesn't go on the page. Reviewer removes unsourced content rather than rewriting it.
• Constrained tool outputs. Interactive tools answer using a vetted interaction database rather than free-form model generation.
• Conservative severity language. Where the evidence is thin, we say "evidence is limited" rather than picking a confident-sounding answer.
• Correction process. When we get something wrong, we publish a dated correction and update the page. See how to report a correction.

Emergency boundaries

Transparency about model versions

Different AllMeds features use different underlying AI models from established providers (currently including Anthropic's Claude family). We update the models as they improve, with the constraints described above remaining the same: source-grounded, pharmacist-supervised, no personal medical advice. If you'd like to know what model powers a specific feature, get in touch.

Reporting an AI failure

If a tool gives you a confusing, wrong, or unsafe answer, please tell us. Email luke@allmeds.ai with the question you asked and what you saw. We use these reports to improve the constraints around the tool. Material issues are also addressed via our corrections process.