AI Control

Today I met a colleague, and we ended up talking about the use of AI in analytical method development, among other things. At the same time, quite a few posts on the topic appeared here as well. About two years ago, I participated in a conference where my answer on this subject left the audience somewhat disappointed — both then and now. Since that day, I’ve been asking myself what exactly wasn’t communicated clearly enough. Maybe I was too excited about the topic, and they weren’t fully ready to hear the simple truth: There is nothing new under the sun. As long as we’re dealing with a computerized system that supports a scientific process, the control requirements remain exactly the same as those we’ve known for many years. So what does this actually mean? 🔸️Intended Use – what the tool is meant to do, and what it cannot replace. 🔸️Risk Assessment (CSA/CSV) – what could be impacted. 🔸️ALCOA+ and reliable data – the model must rely on high-quality, consistent, trustworthy information. 🔸️Human or automated controls 🔸️Stability and consistency checks. 🔸️Usage limitations – what the model may or may not generate. 🔸️Verification of critical points – every AI-based idea still goes through testing like any new method. 🔸️Documentation & Audit Trail (Part 11 / Annex 11) – full traceability of inputs, model versions, and decisions. And then comes the classic question: How can we comply with “Attributable” if AI is the one producing the answer? The answer is simple: AI is a tool — not a role. “Attributable” refers not to who typed the output, but to whoever initiated the action, entered the data, requested the calculation, evaluated the result, and approved its use. All these steps are human — and fully aligned with ALCOA+. In reality, AI doesn’t replace the expert. It simply accelerates thinking and expands possibilities. Within the proper controls, it’s another high-quality scientific tool. And this applies not only to analytics — but to every field. At the end of the day, AI is just another information system. Powerful, evolving, impressive — but still a system. And I say this as someone who lives and breathes AI 24/7 and can’t imagine going backwards. (By the way, the outage today — 18.11.25 — isn’t related.) We dealt with Excel in the past, and we’ll deal with this too. The method is similar — only the environment changed. AI has been part of our tools for years; today it’s just stronger and more visible. The rules, the preparation, and the required processes remain unchanged. And the most important part: If we’re investing in training the tool, the training must rely on reliable information. Any decision based on data must rely on trustworthy data — in analytics, in quality, and in life. Those who consistently maintained strong Data Integrity principles now enjoy simpler, more confident decision-making. Those who didn’t may find things more challenging. What do you think?