Glossary Updates16 new terms added to the glossaries · November 12, 2026, 21:23 CET
AI TechDocKnowledge
  • English (US)
  • English (UK)
  • Deutsch
All context cards

Context card

Documenting the limits of a model: what the IFU of an AI-enabled device must say

How should the limitations of an AI model be documented for the users of a medical device?

By knowledge.aitechdoc.world

Reviewed

Review log and changes

The short answer

As information for safety, tied to evidence: the population and conditions the model was validated for, the declared performance with its uncertainty, known failure modes and inputs it was not built for, and what the user must check or do. The AI Act requires capabilities, limitations and accuracy in the instructions for use (Article 13), the MDR requires residual risks and contra-indications in the IFU, and the 2024 transparency principles of FDA, Health Canada and the MHRA add who needs which information when.

For: Technical writers, usability engineers and regulatory affairs of AI-enabled medical devices

Key points

  • A limitation is a statement about a version, backed by validation evidence.
  • Name the population and conditions of validation, not only the target population.
  • Declare performance with its uncertainty and the metric that was measured.
  • Describe inputs the model was not built for and what the device does with them.
  • Tell the user what to check and when not to rely on the output — that is human oversight in the IFU.

The context

Where limitations go

  • The AI Act instructions for use (Article 13(3)) include the characteristics, capabilities and limitations of performance, the level of accuracy and its metrics, known circumstances that may lead to risks, the performance for specific groups where appropriate, and the human oversight measures.
  • The MDR labeling and IFU requirements (Annex I Chapter III) include residual risks, contra-indications, warnings and precautions and the performance characteristics.
  • Transparency for Machine Learning-Enabled Medical Devices: Guiding Principles (FDA, Health Canada, MHRA, June 13, 2024) asks who needs the information, why, what, where, when and how — from the label to the user interface.

For a device that is a high-risk AI system, these requirements land in one IFU.

A pattern for one limitation

  1. Scope of validation — "Validated on chest X-rays of adults (18 years and older) from fixed digital radiography systems."
  2. Declared performance — metric, value, confidence interval, the data set it was measured on.
  3. Known failure modes — inputs outside scope (children, portable systems, images with implants), with what the device does: warns, refuses, or gives an output that must not be relied on.
  4. What the user does — check, confirm, override; who decides.
  5. Link to the evidence — the validation report and the risk control the statement implements.

What a limitation is not

A disclaimer that the output "may be inaccurate" is not a limitation; it tells the user nothing to act on. A model card is a useful source, but written for developers; the IFU translates it for the clinical user and the deploying organization.

When limits change

A model update can change the limits. The IFU of each device version describes the limits of that version (see model, software and device versions); drift observed after the market can make a limitation necessary that the validation did not show.

Questions readers ask next

Is a general disclaimer enough?
No. Users need concrete limits they can act on: where the model was validated, what it does with other inputs and what they must check.
Do limitations belong in the user interface too?
Often yes. The 2024 transparency principles name the user interface as a place for information that affects patient risk; the IFU and the interface must say the same.

Sources

  1. Regulation (EU) 2024/1689 (Artificial Intelligence Act) — EUR-Lex, July 12, 2024Law in force: Binding legislation that applies as of the review date.
  2. Regulation (EU) 2017/745 on medical devices (MDR) — EUR-Lex, May 5, 2017Law in force: Binding legislation that applies as of the review date.
  3. Transparency for Machine Learning-Enabled Medical Devices: Guiding Principles — FDA, Health Canada and MHRA, June 13, 2024Final guidance: Guidance published in its final form by an authority or an official body. Guidance explains the law; it is not law itself.
  4. Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations (draft guidance) — U.S. Food and Drug Administration, January 7, 2025Draft guidance: A draft for consultation. It may still change and must not be treated as the final text.

Review log and changes

Every context card is checked against its sources before it is published, and again whenever it changes; the date under the byline is the last review. Corrections (something was wrong) and additions (something was missing) are logged below with date and time (Berlin time). Typos, formatting and link fixes are not listed.

Reviewed

No corrections or additions since publication.