Your R&D pipeline never trains anyone’s model

01

The duty has a name, and it is yours

A key principle, in the EMA’s own words: the trial sponsor, the applicant or holder of the authorisation, or the manufacturer must ensure every algorithm, model, dataset and data pipeline is fit for purpose. It warns the standard can be stricter than usual data-science practice.

02

A black box has to earn its place

Black-box models are acceptable where you substantiate that interpretable ones perform or hold up worse. Then you owe the architecture, the hyperparameter tuning, training metrics, validation and test results and a monitoring plan, and the EMA suggests discussing the case with it in advance.

03

The weights can carry the data

Large models are at particular risk of memorising what they were trained on, and the conclusion is blunt: if the data are not fit for sharing, anonymise before moving the model to a less secure environment. In that sentence, where a model runs becomes a data-transfer decision.

04

Generated text needs a reviewer

Whatever a model drafts, edits or translates for product information goes under close human supervision, because generative models produce plausible but wrong or incomplete text, and a quality review has to confirm it before a regulator sees it. That is a process, and nobody sells it to you.