AI can be used at different stages of developing a medicine, from nonclinical research to clinical evidence and manufacturing. But using a model is not the same as proving that a drug is safe or effective. In draft guidance, the U.S. Food and Drug Administration proposed a risk-based way for sponsors to assess whether an AI model is credible for a specific regulatory use. The document is draft guidance, not a report of faster approvals.
What must be tested
The FDA and European Medicines Agency's joint principles say AI should support reliable evidence across the drug life cycle while maintaining quality, efficacy and safety requirements. That means a useful model needs a defined purpose and an evaluation that matches the decision it informs.
The original article offered no identified clinical trial, comparison group or measured time saving. A molecular prediction may help choose what to investigate next, but patients only benefit after the candidate has passed the required testing and regulatory review. Claims of saved years or lower cost need a named study and a clear measure, neither of which the earlier story supplied.
Different uses call for different evidence
Drug development involves many decisions, from selecting a candidate for further study to assessing information about safety, effectiveness and quality. A model may be designed for one specific task within that process. Its usefulness depends on what it is intended to do, the data available and how its output will influence a decision.
The FDA’s January 2025 draft guidance discusses AI used to produce information intended to support regulatory decisions about drugs and biological products. It proposes a risk-based approach to assessing a model’s credibility for a particular context of use. The FDA page labels the document draft, nonbinding and not for implementation; it is not a final rule or a report that AI has accelerated every trial.
Context of use makes a broad statement such as “the model is accurate” more specific. What question is the model answering? What input does it receive? How will its output be used? A test that is relevant to one purpose may not establish performance for a different decision or population.
From a prediction to an experiment
A prediction can help researchers decide what to investigate. For an illustrative example, a model might suggest that one candidate is worth examining in a laboratory. That suggestion is not the same as a demonstrated effect in people, an approved treatment or proof that the candidate is safe. The example describes a possible sequence, not a named product or study.
Evidence needs to match the claim being made. If a report says a model identified a promising candidate, readers should look for what was actually tested and what the results showed. If it claims a treatment works in patients, evidence from the relevant clinical research is needed. A result at one stage should not be described as the outcome of a later stage.
It is also important to distinguish the model’s technical performance from the clinical question. A model may perform well against a defined test set yet still need evaluation for the setting in which its output will be used. The consequences of a wrong output can differ depending on the decision it informs.
What the joint principles emphasize
The FDA says its drug and biologics centres collaborated with the European Medicines Agency on ten guiding principles for AI in drug development. The listed themes include a clear context of use, human-centred design, a risk-based approach, multidisciplinary expertise, data governance, documentation and performance assessment.
These are principles for consideration, not evidence that a particular model or medicine has passed a test. They also do not provide a universal estimate of time or money saved. Their relevance is that the purpose, evidence and management of a model must be considered together.
Documentation can help reviewers understand where information came from and how a model was developed and evaluated. Data governance addresses questions such as whether the data are suitable for the intended task and how they are managed. The exact evidence needed for a specific use depends on the model and the decision being supported.
How to assess an acceleration claim
A claim that AI shortened development should identify the starting and ending points being compared. A reduction in the time needed to rank candidates is not automatically a reduction in the time required for laboratory studies, clinical trials or regulatory review. Those stages answer different questions and may involve different evidence.
A useful comparison would identify the candidate, the work performed with and without the tool, the period measured and the method used to calculate any difference. It should also explain whether quality or failure rates changed. A before-and-after comparison can be affected by other changes in the project, so attributing the entire difference to AI requires care.
Cost claims require similar detail. A figure for one computational task may not include all the costs of developing and evaluating a medicine. The earlier article’s estimate of a 40% reduction was unsupported; the FDA documents cited here do not establish such a saving.
What patients and readers should look for
A headline about AI and medicines should name the treatment or candidate, the research stage and the source of the result. An early prediction, a laboratory experiment and a clinical trial provide different types of information. The word “discovery” does not mean a treatment is available or suitable for a patient.
For a reported trial result, readers can look for the study design, the participants, the outcomes measured and the limitations described by the researchers. A promising finding may justify further study without resolving every question about safety or effectiveness. Regulatory status should be checked for the specific product and jurisdiction.
AI may help researchers work with complex information, but the model’s role should be described precisely. The FDA’s draft framework and the joint principles focus on the credibility and management of evidence. Neither source supports a general promise that AI has compressed years of development into months.
People considering a treatment should discuss their individual circumstances with an appropriate health professional. A general article about an AI model cannot establish whether a medicine is suitable for a particular person. In reporting, separating the model’s research role from the medicine’s clinical evidence helps readers understand both the potential application and the questions that remain open.
