The premise
Over 1,000 AI-related regulations across 70+ countries — and no single place that compares how regulators govern them.Global AI regulation for medicines is fragmented. AICURIS reads the public output of major regulators, classifies the AI-relevant material with a hybrid semantic-and-keyword model, and lays it side by side so a reviewer can see alignment and divergence at a glance.
The register
Ranked by AI-relevance. FDA documents in the paper's classified set carry the AI-CLASSIFIED stamp.Federated view
One theme, four regulators. Click a theme to filter the register.The point of AICURIS is comparison. For each recurring theme in AI medicines regulation, here is how many documents each authority has produced — a map of where the world's regulators are converging, and where they are silent.
Method & performance
A dual-scoring OR-ensemble — semantic similarity for implicit AI, keyword similarity for explicit terminology.Regulatory text discusses AI in two ways: explicitly (naming machine learning, LLMs) and implicitly (describing data-driven methods without the vocabulary). One signal alone misses half of it.
Each document is scored two ways — semantic similarity to a curated reference set of AI-regulatory documents, and keyword similarity across AI, legal and biomedical term banks. A document is flagged AI-relevant if either signal clears its threshold.
The two methods correlate only moderately (≈0.41–0.53), which is exactly why combining them helps — they fail on different documents. Thresholds were chosen by grid search to hold recall high while keeping the review burden manageable.
⚑ Honest limitations
- Recall-only. Precision was not estimated — labelling 99,070 documents by hand was not feasible.
- Scope: English-language documents; FDA, EMA, WHO (MHRA added here for breadth), 2019–2025.
- Proof-of-concept. It surfaces and structures documents to inform human review; it issues no regulatory judgement.
- This demo reproduces the paper's exact classification for the FDA set; other agencies are shown ranked by relevance (doc-content re-scoring pending).
Recall vs. review burden
Each curve is one scoring signal: as the threshold relaxes, recall rises but more documents need review. The ensemble lets us sit high-left — high recall, low burden.
AI-related regulatory output, by year
Indexed documents per authority, 2019–2025. The slope is the story.
About & data
Open access. Cite the paper. Download the corpus.AICURIS is the public companion to a peer-reviewed study. It exists to make the paper's claim testable: that an AI-enabled system can federate AI-related drug regulation across jurisdictions and keep it current.
The full indexed corpus — every document, agency, date and score — is available below for independent analysis, in keeping with the paper's data-availability commitment.
Engineered by Mukesh Pareek (software & validation author). Part of Singh, Pareek, Prokle, Sanchez, Hohgrawe & Auclair, Frontiers in Drug Safety and Regulation, 2026.