Survival of the fittest tool: choosing between open and closed artificial intelligence in pharmaceutical science

Three years ago, this journal published the editorial Technological Darwinism , urging pharmaceutical scientists to adapt to artificial intelligence (AI) rather than resist it. Since then, the question has shifted from whether to use AI to which AI to use, under what circumstances, and with what evidence. This review addresses that question by mapping current AI tools across the five canonical stages of drug development, from discovery to post-market pharmacovigilance. It is a narrative review, not a systematic one, and it treats openness as a spectrum rather than a binary. It compares two broad categories: open systems free, often open source or open weight, inspectable, and locally deployable and closed systems, which are proprietary, commercial, and delivered as a service. The central conclusion is pragmatic. For routine laboratory, educational, and research tasks, capable open tools now exist for almost every application, including manuscript drafting, literature review, reference management, statistical analysis, pharmacokinetic and pharmacodynamic modelling, calibration and standard-curve fitting, chemical structure drawing and SMILES handling, image and figure generation, and prediction of physicochemical, ADMET (absorption, distribution, metabolism, excretion, and toxicity), formulation, and excipient compatibility properties. Free web services, however, may require confidential structures to be uploaded to a third party, so openness is distinguished here from private deployment, and independent head to head evidence against commercial tools is uneven. In contrast, integrated end to end drug discovery and validated regulatory workflows remain dominated by proprietary platforms. During the period covered by this review, the first generative AI-designed molecule targeting a generative AI-derived protein reported positive Phase IIa results and entered Phase III in July 2026; protein structure prediction was recognized with the 2024 Nobel Prize in Chemistry, shared by Hassabis, Jumper and Baker and became partly open; an autonomous frontier language model improved a longstanding medicinal chemistry reaction through thousands of wet-laboratory experiments; and regulatory agencies issued their first guidance on AI credibility. The review argues that the most enduring competency is not mastery of any single AI platform but the ability to select appropriate tools, critically validate their outputs, and document their credibility in proportion to the decisions they inform. Accordingly, open AI systems should be the default for routine, low-risk, and educational applications, whereas proprietary platforms currently retain advantages for integrated discovery and regulated development. Credibility is defined operationally, human oversight is treated as a condition of use, and the review states where comparative evidence is lacking.