AI Transparency

How Our AI Works

Transparency about MediWing's AI system, data flows, and limitations

Last Updated: June 3, 2026

1The AI Model

MediWing uses Claude Sonnet 5, developed by Anthropic, PBC. Claude is a large language model (LLM) trained on broad text data including medical literature. For fast tasks like document text extraction, we use Claude Haiku 4.5.

LLMs generate text by predicting likely next tokens based on their training. They are highly capable at language understanding and summarization, but they can produce plausible-sounding errors — a phenomenon called "hallucination." This is why MediWing's outputs are educational only and must always be verified with a healthcare professional.

2What Data Is Sent to AI

When you upload a document, the following is sent to Anthropic's API:

The extracted text content of your document
Your selected language preference
A structured analysis prompt authored by MediWing

We never send Anthropic your name, email address, account ID, or any other personally identifying information. Anthropic does not use API-submitted data to train its models. Anthropic's API operates under commercial terms with no-training guarantees.

3Citations & Evidence

MediWing integrates with PubMed — the U.S. National Library of Medicine's database of 35M+ peer-reviewed biomedical publications. After AI analysis, relevant citations are fetched and linked to findings.

Citations serve two purposes: (1) they allow you to verify the basis for any interpretation against published research; (2) they help meet the transparency standard required of AI health tools — that reasoning should be grounded in verifiable sources.

Not every finding will have a citation. The absence of a citation does not mean a finding is wrong, but you should discuss uncited interpretations with your healthcare provider.

4Confidence Scores

MediWing displays a confidence score (0–100%) for each analysis. This score reflects the AI's self-assessed certainty about the interpretation, not a clinical accuracy guarantee.

When confidence is below 85%, MediWing displays an explanatory note about the uncertainty. A high confidence score does not mean the interpretation is correct — AI confidence and factual accuracy are not the same thing. Always verify with your healthcare provider.

5Drug Interaction Data Sources

The drug interaction feature uses three external data sources:

OpenFDA Drug Labels — FDA-approved prescribing information and labeled interactions
RxNorm — the NIH's standard drug naming system used for drug normalization
FDA FAERS — the FDA Adverse Event Reporting System, containing voluntary post-market safety reports

These sources are authoritative databases but are not exhaustive. New interactions may not yet be in these databases. Always have a pharmacist review your full medication list.

6What AI Cannot Do

✗Access your complete medical history, other test results, or clinical context
✗Account for rare individual genetic variations or drug metabolism differences
✗Provide real-time or up-to-the-moment medical knowledge (training has a cutoff date)
✗Replace clinical judgment that integrates physical examination, history, and patient context
✗Guarantee accuracy — AI outputs should always be treated as a starting point for conversation with your provider, not a conclusion

7Regulatory Classification

MediWing is classified as an educational wellness tool, not a Software as a Medical Device (SaMD) under FDA guidelines. We are not FDA-cleared or FDA-approved. Our outputs are not intended to drive clinical decisions.

The FDA's 2026 Clinical Decision Support guidance established that tools providing educational health information with verifiable, cited reasoning are generally outside medical device classification when they support — rather than replace — clinical judgment. MediWing is designed in alignment with these principles.

← Terms of Service ← Privacy Policy