Substantial modification
A substantial modification is a change to a high-risk AI system — to its intended purpose or one affecting its compliance — that was not foreseen in the original conformity assessment, triggering a fresh assessment.
Under Article 25, an operator that substantially modifies a high-risk system already on the market, or changes the intended purpose of a system so that it becomes high-risk, is treated as the provider for that system and inherits the full provider obligation set. The Act deliberately excludes from the definition changes that were pre-planned and assessed: a system designed to continue learning within predetermined parameters set out in the technical documentation does not undergo a substantial modification simply by learning as intended.
- Art. 3(23)
- Art. 25
- Art. 43(4)
Substantial modification: why it matters
This is where MLOps meets regulation. Every high-risk system will be retrained, tuned and extended, and the question of which of those changes restarts conformity assessment determines whether releases take days or quarters. The Act's answer rewards foresight: changes anticipated and documented in the original technical file are not substantial, so the cheapest compliance investment is a technical file that describes the intended change envelope generously and precisely rather than narrowly.
Substantial modification in practice
A credit-scoring provider's Annex IV file states that the model is retrained quarterly on new repayment data within a defined performance and feature envelope. Those retrains are not substantial modifications. Adding a new input feature drawn from social media, or extending the model from consumer lending to SME lending, sits outside the assessed envelope and triggers a fresh conformity assessment before the change can ship.
Substantial modification: the mistake people make
Assuming any model update is a substantial modification, and freezing releases as a result. Continued learning and updates that fall inside the pre-determined, documented change envelope are expressly not substantial — the boundary is what the original assessment covered, not whether the weights changed.
Substantial modification: common questions
- Does retraining a model count as a substantial modification?
- Not if the continued learning was pre-determined and assessed in the original technical documentation. Retraining that changes the intended purpose, or that takes the system outside the envelope the conformity assessment covered, does count.
- Who becomes responsible after a substantial modification?
- The operator that made the modification is considered the provider of that high-risk system under Article 25 and takes on the full provider obligations, including a new conformity assessment.
- How do we minimise re-assessment cost?
- Define the intended change envelope explicitly in the Annex IV documentation at the outset — the retraining cadence, the feature set boundaries, the performance thresholds — so foreseen changes stay inside the original assessment instead of triggering a new one.
Substantial modification: terms that come up alongside it
Run the free 2-minute EU AI Act checker — no sign-up.