AI in regulatory work: grounded in sources, not a text generator

Disclosure: The operator of this site provides engineering services in regulatory affairs for medical devices.

What happened: Two recent research papers look at how large language models (LLMs) can be used in regulatory work on medical devices. Both are arXiv preprints, so they have not yet been peer reviewed. Neither publishes code.

The problem: A language model can draft a section of an ISO 14971 risk file or IEC 62304 software documentation in seconds. But when a notified body or regulator asks where a statement comes from, freely generated text is no help. Without traceability to a source, you gain an audit risk rather than saving time.

The approach in the first paper separates suggesting from deciding:

1. The device documents sit in a controlled knowledge base.
2. The model suggests hazards or safety statements, each with a reference to where it found them.
3. A review step assesses the suggestions and flags uncertainty.
4. A qualified person decides, and the decision is recorded.

The second paper shows a variant built on a knowledge graph, a structured database of terms and their relationships. A locally run open model (Mistral 7B Instruct) generates queries against it to classify software as a medical device under US law. The authors report less manual review effort but give no figures in the abstract.

What RA teams can take from this:

• Use AI only as an assistant, and require every statement to point to a source in your own records.
• Approval stays with a person and is documented in a traceable way.
• Validate the tool itself as QMS software, because it affects quality records.
• Local or open models are an option when confidential device data should not leave the company.

Context: Both papers are concepts without published results. Vendor claims such as “up to 90 % less documentation effort” are not supported by them. The value lies in the principle, not in a finished product.

Source: Gangopadhyay et al.: From Safety Documentation to Safety Knowledge Support: An Evidence-Grounded LLM Framework for Medical Devices (arXiv), of 12 Aug 2026