Who—or What—Should We Trust in the LLM Era?

Large language models (LLMs) and autonomous agents can synthesise scientific arguments and construct computational models rapidly, but fluency is not scientific validity. Neither expert status, human oversight, multi-agent consensus, publication nor patent protection substitutes for independent empirical testing. Experiments themselves can conflict because of bias, measurement error or genuine biological heterogeneity. We propose an evidence-to-decision trust architecture for drug research and model-informed drug development (MIDD) that links verifiable primary observations, provenance and uncertainty to competing mechanistic hypotheses, adversarial AI review, discriminating experiments and accountable decisions. Credibility is evaluated for a specified context of use, not assigned globally to people or algorithms. The framework spans target validation, translational pharmacology, dosing, clinical development and regulatory assessment, illustrated by a proposed IQANOVA ATLAS AI–quantitative systems pharmacology (AI-QSP) workflow. We extend the framework to regulatory agencies using AI agents: scientific assessment and model credibility are distinct from the legally binding act of medicine authorisation. Inspired by Juvenal's question of who checks the guardians and the historical Rabkrin inspection debate, we propose auditable agent roles, explicit legal decision rights, independent scrutiny and routes for challenge. A synthetic SBML/SED-ML/COMBINE package with provenance ontologies illustrates reproducible reporting without claiming biological validation. A prospective benchmark is proposed to test scientific quality, regulatory decision loss, agent independence and appeal outcomes. The final question—whether, when and under whose authority an AI could make binding decisions—remains unresolved.