Generative AI Will Not End Human Oversight in Drug Safety, FDA Scientists Argue

Generative artificial intelligence has been heralded as a technology that could automate nearly anything, including the painstaking work of monitoring medicines for safety. But a new perspective from scientists at the U.S. Food and Drug Administration argues that the arrival of large language models does not mean the end of human judgment in pharmacovigilance, the discipline devoted to detecting, assessing, and understanding adverse drug events. Instead, the authors suggest, the field is heading toward a synthesis: hybrid systems in which machines handle computational heavy lifting while experts retain control over the decisions that matter most.

The paper, published on 18 August 2026 in Frontiers in Drug Safety and Regulation by Leihong Wu, Joshua Xu, Oanh Dang, and Robert Ball, draws its provocative framing from history. The authors invoke Francis Fukuyama’s famous “end of history” thesis from the close of the Cold War, which borrowed Hegel’s dialectical process of thesis, antithesis, and synthesis. In their analogy, the historical thesis of pharmacovigilance is the human expert who reads narrative case reports and applies introspective clinical judgment. The contemporary antithesis is the claim that drug safety surveillance should be fully computable, independent of human reviewers. The emerging synthesis is the human-AI pharmacovigilance system, in which neither pole alone is sufficient.

The stakes are enormous. Pharmacovigilance revolves around individual case safety reports, or ICSRs, submitted in vast volumes by healthcare professionals, patients, and pharmaceutical companies. Each report contains structured fields such as age, sex, suspect drug, and outcome, alongside sprawling free-text narratives of variable quality. Reviewers must verify completeness, detect duplicate reports, assign priority, and integrate information across cases and external data sources to identify potential safety signals. Back in 2022, Ball and Gregory Dal Pan asked whether artificial intelligence for pharmacovigilance was “ready for prime time” and concluded that while machine learning could support parts of the workflow, full automation of case processing was not realistic. Human review remained essential, and the field lacked a “pharmacovigilance cognitive framework,” a computable description of the rules of inference that safety reviewers apply when judging whether a drug likely caused an adverse event.

The new paper asks whether large language models, trained on massive text corpora to summarize, extract, and generate language, have changed that conclusion. The authors’ answer is a framework they call “computable pharmacovigilance,” in which tasks are evaluated for their computational tractability. A task is considered computable when it can be translated into explicit, auditable computational steps with clearly defined inputs and outputs, processing details precise enough for traceable execution, and outputs that can be judged by reliable, reproducible criteria. By that standard, computability is a spectrum, shaped by input standardization, output verifiability, tolerance for error, auditability, reliance on clinical judgment, and the regulatory consequences of mistakes.

By this lens, many routine tasks are increasingly ripe for automation. Completeness checks that identify missing minimum report elements, detection of duplicate reports, and structured information extraction from narratives all have well-defined requirements and fixed parameters. Real-world results support this. The ALIGN system, a compositional large language model architecture that generates candidates and then self-evaluates them, achieves 87 to 90 percent accuracy in MedDRA medical coding. Deduplication tools using probabilistic record linkage on structured and unstructured data can reduce manual workloads by up to 60 percent, and these tools have seen relatively high adoption because their algorithms transparently mirror existing human cognitive processes.

At the other end of the spectrum sits case-level causality assessment, the judgment of whether a specific drug actually caused a specific adverse event. The authors argue this remains profoundly difficult to formalize. It demands integration of temporality, rechallenge information, and alternative etiologies against a background of pharmacological mechanism, along with abstract reasoning whose structure depends on the particular drug-event pair in question. Human adjudicators themselves reach full agreement on binary causality classifications in only about 61 percent of cases, and current benchmarks robust enough to measure AI performance against expert judgment are scarce, complicated by the nuances of individual drug-event combinations and the continual emergence of rare or novel adverse events.

The paper also weighs the choice between “big” and “small” models, a distinction that shapes architecture and governance. Large foundation models can be repurposed across many tasks through prompting or light fine-tuning, summarizing long narratives, extracting structured fields, and even translating natural-language reviewer questions into executable database queries. But they bring hallucinations, prompt sensitivity, variable outputs, and higher validation burdens, and running them across every report in a large safety database is expensive, with external APIs raising data protection questions. Small, task-specific models, by contrast, have constrained inputs and well-defined outputs, more stable failure modes, and can be trained on internal historical safety data that commercial models cannot access, making them easier to document and audit.

Illustrating the trade-off, the authors report a pilot study benchmarking a fine-tuned BioBERT named entity recognition model against a zero-shot Llama-4 large language model on safety concept extraction from roughly 800 annotated FAERS reports and an existing VAERS corpus. The fine-tuned BioBERT outperformed the large model on both datasets, achieving an overall F1-score of 0.54 versus 0.42 on FAERS narratives and 0.53 versus 0.31 on the VAERS vaccine corpus. Error analysis showed the large language model prone to hallucinated character positions, boundary drift, and label confusion, while the BERT model behaved as a conservative, schema-consistent workhorse. The authors stress the comparison illustrates the value of task-specific fine-tuning rather than settling any definitive ranking of model capability.

Their preliminary work on causality assessment reinforces caution. When an LLM was applied to a well-documented case series, the model acted too confidently, labeling ambiguous cases as “causality assessable” where human reviewers concluded no reliable determination could be made. While AI and humans agreed on clear cases at the extremes, the machine was overly optimistic in the muddled middle, accepting cases experts would reject. The authors also caution that explainable AI techniques such as SHAP and LIME, though capable of improving trust, can increase cognitive load and misalign with established clinical reasoning, so transparency alone does not guarantee that outputs are accurate, contextually relevant, and actionable.

The authors conclude that generative AI does not signal the end of history for pharmacovigilance but the end of purely manual, labor-intensive case processing. They propose treating computability as a design principle, not an endpoint: decompose workflows into steps with established inputs and outputs, match each step to an appropriate level of automation, deploy in phased and risk-based stages, and embed AI within sociotechnical systems that preserve accountability and human control. In the hybrid architecture they envision, large models act as flexible interfaces and orchestrators calling on small models, rule engines, terminologies, and retrieval tools, while experts provide strategic and ethical oversight. Trustworthiness, they argue, is itself the core of the pharmacovigilance cognitive framework, something that for now only matters to humans and only humans can judge. The future, in their view, lies in replacing AI-averse workflows with AI-enabled ones that remain accountable to human experts.

Subject of Research: Automating pharmacovigilance workflows with generative AI and large language models under human oversight

Article Title: Does generative AI mean the “end of history” for pharmacovigilance automation? towards a framework for the future of human-AI systems

Article References: Wu, L., Xu, J., Dang, O., & Ball, R. (2026). Does generative AI mean the “end of history” for pharmacovigilance automation? towards a framework for the future of human-AI systems. Frontiers in Drug Safety and Regulation, 6, Article 1846339. https://doi.org/10.3389/fdsfr.2026.1846339

Image Credits: AI Generated

DOI: 10.3389/fdsfr.2026.1846339

Keywords: pharmacovigilance, generative AI, large language models, drug safety, individual case safety reports, causality assessment, FDA, human-AI systems, medical coding, deduplication, computability, regulatory science