Imagine an artificial intelligence system analyzing a process deviation. It reconstructs the sequence of events, compares the data with previous cases, identifies correlations, and suggests a possible root cause. Within seconds, it produces a plausible, well-reasoned analysis, perhaps even more structured than the one a team of experts might have prepared.
But that is when the most interesting questions arise.
What data did the system rely on? How representative are those data? Which version of the model processed them? Could an update lead to a different answer? What error rate is acceptable for that specific use? Who verifies the output? And, above all, what must we be able to demonstrate months later if that output contributed to a decision with significant implications for quality?
This is where the debate around artificial intelligence takes a different turn. Model performance matters. So do the safeguards built into the technology by its developer. But when AI becomes part of a regulated process, reliability is no longer an abstract property of the technology itself. It becomes a characteristic of the system in which that technology is deployed.
Superintelligence, Before the Definition
On September 29, 2026, the White House offered a curious example of how terminology can precede even the definitions behind it.
With the Executive Order Inaugurating the Era of Super Intelligence, the Trump administration directed the US executive branch to use the terms “Super Intelligence” and “SI” instead of “Artificial Intelligence” and “AI” in non-statutory documents and communications. For now, however, the legal substance remains unchanged: the order itself specifies that “Super Intelligence” refers to technologies already covered by the federal definition of artificial intelligence. A proposed new statutory definition is to be submitted within sixty days.
In other words, the new name comes before the new category.
That same day, executives from Google, Anthropic, Meta, OpenAI, xAI, and Nvidia signed the White House Accord on Super Intelligence, a voluntary commitment to managing the risks associated with the most advanced systems. The framework operates at several levels: internal controls, teams responsible for verifying their implementation, assessments conducted by independent external parties, and board-level oversight. The accord, which Trump described as “morally binding”, does not, however, establish an enforcement or sanctions framework comparable to that of binding regulation.
The issue cannot simply be reduced to a contrast between American self-regulation and European regulation. That would be a convenient interpretation, but it would explain little about what actually happens when an AI system is used in the research, development, or manufacturing of a medicinal product.
Indeed, the FDA and EMA are working together to establish a far more specific framework and shared terminology.
The Regulatory Question Is a Different One
In January 2026, the two agencies jointly published the Guiding Principles of Good AI Practice in Drug Development, ten principles addressing the use of artificial intelligence in generating and analyzing evidence throughout the medicinal product lifecycle, including non-clinical and clinical development, post-marketing activities, and manufacturing.
The sequence is revealing.
Human-centric design. A risk-based approach. Adherence to standards. A clear context of use. Multidisciplinary expertise. Data governance and documentation. Good model design and development practices. Risk-based performance assessment. Life-cycle management. Clear information for system users.
This is where the focus shifts.
Knowing that a model developer has internal controls, safety procedures, and independent assessors in place is important. But that alone does not tell us whether the model is suitable for analyzing specific process data, supporting a deviation investigation, classifying a pharmacovigilance signal, or contributing to a regulatory decision.
To establish that, we need to understand exactly what we are asking the system to do.
The context of use therefore becomes much more than a regulatory term. It defines the relationship between the system, the data, the user, and the resulting decision. The same model may be entirely suitable for summarizing technical documentation, yet require a very different level of supporting evidence when used to generate an output that influences a GxP decision.
The Model Alone Is Not Enough
One aspect of the FDA-EMA principles deserves particular attention from industry.
Data provenance, data transformations, and analytical decisions must be documented in a detailed, traceable, and verifiable manner, consistent with GxP requirements. Performance assessment must address the system as a whole, including human-AI interactions. Oversight must also continue over time, through scheduled monitoring and periodic reassessment, including measures to detect phenomena such as data drift.
This approach had already been outlined by the EMA in its reflection paper on the use of AI throughout the medicinal product lifecycle, adopted in 2024. The document highlighted the risks associated with complex, poorly transparent architectures and the need for measures to protect patient safety and the integrity of results.
Let us return, then, to our initial process deviation.
The system analyzes batch records, process data, and historical cases, and suggests a root cause. Verification cannot be limited to asking, “Does the answer look correct?” We need to establish whether the dataset is appropriate, what information may have been excluded from the analysis, how consistent the system’s performance is, how the model version is recorded, and what role the operator played in the final decision.
This also changes what we mean by human oversight.
A person who automatically approves an output simply because it appears reasonable adds little to the control process. Human oversight becomes meaningful when those responsible have the expertise, information, and authority needed to challenge the system, request further verification, or reject its output.
This is probably one of the most difficult challenges for organizations beginning to integrate AI into regulated processes: designing human oversight as an integral part of the process, rather than adding it at the end as a signature.
The AI Act Is Only Part of the Picture
In Europe, additional provisions of the AI Act became applicable on August 2, 2026, together with the associated enforcement powers. Providers of general-purpose AI models are subject to documentation and transparency obligations, while models presenting systemic risk must meet additional requirements relating to risk assessment and mitigation, cybersecurity, and incident management.
For pharmaceutical companies, however, compliance with the AI Act does not address the entire challenge.
The European AI regulation and pharmaceutical legislation examine AI systems from perspectives that may overlap without fully coinciding. A model may be subject to specific obligations as an AI system or model, while its use within the medicinal product lifecycle must also be assessed in terms of the quality and reliability of the evidence it helps generate.
It is particularly significant that the FDA and EMA, operating within different regulatory systems, have chosen to develop the principles of Good AI Practice jointly. This suggests that, in the life sciences sector, the most meaningful distinction may lie less between the United States and Europe than between AI considered as a general-purpose technology and AI deployed within a specific regulated process.
Governing What Changes
There is one final characteristic that sets AI apart from many of the digital tools the industry has learned to qualify and validate: it can change.
Models evolve, data change, and configurations are modified. Even behaviors that were not the intended target of an update may be affected. When an external system is updated by its provider, the regulatory question is not simply what has improved. It is also whether the conditions on which our confidence was originally based still hold.
This is why life-cycle management, as outlined by the FDA and EMA, is probably one of the principles that deserves the closest attention. Validation risks becoming a snapshot in time if the system we have assessed continues to evolve.
Organizations therefore need criteria to determine when a change is significant, which performance characteristics must be reassessed, how versions and configurations should be managed, and which events require a new risk assessment. This calls for a governance framework involving Quality, IT, data science, regulatory affairs, and process owners. It is a responsibility that can hardly be assigned to a single corporate function.
The debate surrounding “Super Intelligence” may seem far removed from these practical concerns. Yet it highlights a question the pharmaceutical industry knows well: who do we trust, and on the basis of what evidence?
When a medicinal product is involved, the answer cannot rest solely on the characteristics declared by the technology provider. We need to be able to trace the entire path from a model to its specific application, from that application to an output, and from the output to a decision.
It is along this path that AI becomes part of the pharmaceutical quality system. And it is there that we determine how much we can reasonably entrust to it.
Vuoi trovare più facilmente MakingPharmaIndustry su Google?
Aggiungici alle tue fonti preferite
