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

The retrospective descriptive study, conducted by Mikyung Kim and Dong-jun Choi of the Department of Internal Medicine and the Regional Pharmacovigilance Center for Korean Herbal Medicines, mined the shared adverse event reporting databases of the two hospitals. After removing duplicates, the researchers identified 454 reports in which an herbal medicine was recorded as the suspected agent. That figure represents just 1.2 percent of the 38,367 adverse event reports handled by the centers over the entire study period, a proportion that itself hints at how rarely herbal products are formally flagged as suspected causes of harm even in institutions where they are prescribed routinely.

The most striking finding concerns the regulatory status of the products involved. Non-licensed oral herbal medicines accounted for the largest share of suspected agents at 48.7 percent, narrowly exceeding the 47.0 percent share of oral herbal medicine products licensed by Korea’s Ministry of Food and Drug Safety. At the report level, 51.3 percent of cases involved at least one non-licensed formulation. These are preparations that fall outside the standard regulatory licensing pathway, including individualized herbal decoctions compounded for specific patients, and they are precisely the products that national databases built around licensed pharmaceuticals are least equipped to track.

The clinical profile of the reported events followed recognizable patterns. Gastrointestinal disorders were the most frequently reported system organ class, classified according to the Medical Dictionary for Regulatory Activities, followed by skin and subcutaneous tissue disorders. Among the 572 drug-event pairs with evaluable causality assessments, the overwhelming majority, 87.7 percent, were judged merely possible, while cases assessed as certain were vanishingly rare at 0.2 percent. That distribution reflects a familiar limitation of spontaneous reporting: without rechallenge or unambiguous temporal and biological evidence, causality in herbal medicine cases usually remains provisional, complicated by polypharmacy, delayed onset, and the multi-component nature of herbal formulas.

The surveillance gap emerges most clearly when the researchers compared reports that were forwarded to the national system with those that were not. Overall, 24.0 percent of the 454 reports were never submitted to national pharmacovigilance. The differences between the two groups were stark and statistically significant. Non-submitted reports far more frequently involved non-licensed herbal medicines, at 83.5 percent compared with 41.2 percent of submitted reports. They were far less likely to include documented clinical severity grading, at 7.3 percent versus 59.1 percent, and far less likely to contain a causality assessment, at 26.6 percent versus 98.8 percent, with p values below 0.001 across all three comparisons.

In other words, the reports most likely to slip through the cracks were those describing the products least visible to regulators, and those least complete in the clinical detail that pharmacovigilance assessment depends on. Severity grading in the study followed the Common Terminology Criteria for Adverse Events, the standardized scale used in oncology and increasingly across drug safety work, while causality was evaluated within the framework of individual case safety reports submitted through Korea’s adverse event reporting infrastructure. A report lacking both a severity grade and a causality judgment offers national assessors little to work with, which may partly explain why incomplete cases were not prioritized for submission.

The structural reasons behind the gap are not difficult to infer from the Korean context. Traditional Korean medicine doctors routinely prescribe individualized multi-herb decoctions that are compounded fresh for each patient. These formulations do not carry product licenses in the way manufactured drugs do, so they sit awkwardly within a reporting system whose data structures, coding conventions, and regulatory workflows were designed around licensed medicinal products. A suspected reaction to a ten-herb decoction does not map neatly onto a single product code, and the paper’s authors argue that the current pharmacovigilance framework therefore does not fully capture the range of herbal medicines actually used in real-world Korean practice.

The study’s setting gives its findings particular weight. Dongguk University Ilsan Oriental Hospital hosts a regional pharmacovigilance center specifically dedicated to Korean herbal medicines, one of the institutional mechanisms Korea has built to strengthen reporting from the traditional medicine sector. If, even at a center with dedicated reporting infrastructure, a quarter of herbal adverse event reports failed to reach the national system and half involved non-licensed formulations, the authors suggest that the true blind spots in less specialized settings are likely to be larger. Fifteen years of data also lend temporal depth, spanning a period in which Korea expanded its pharmacovigilance networks and electronic reporting systems considerably.

The methodological caveats are worth noting. This was a single-center retrospective descriptive analysis, so the absolute numbers and proportions cannot be generalized to the whole country, and spontaneous reporting data inherently capture only a fraction of adverse events that actually occur, with underreporting the norm rather than the exception across all drug classes. The study was exempted from full ethical review because it used de-identified data from previously collected reports. Still, the internal comparison between submitted and non-submitted reports is the analytically powerful part of the design, because both groups came from the same institutions over the same period, isolating the characteristics that predict whether a case enters national surveillance.

The authors’ conclusions point toward concrete reforms: broader surveillance coverage that explicitly encompasses non-licensed formulations, improved documentation of clinical severity and causality assessments at the point of reporting, and stronger reporting support for traditional Korean medicine professionals, who are often the first and only clinicians positioned to recognize a herbal adverse event. As herbal medicine use grows globally and integrative care models spread, the Korean experience offers a cautionary lesson for any country attempting to fold traditional formulations into regulatory safety systems built for licensed pharmaceuticals. A safety net with holes in exactly the places where the most-used products fall through is, the study implies, a net that needs reweaving.

Subject of Research: Herbal medicine adverse event reporting and pharmacovigilance surveillance gaps in traditional Korean medicine

Article Title: Real-world patterns of herbal medicine-related adverse event reports and surveillance gaps in national pharmacovigilance: a 15-year retrospective analysis from a single integrative medical center in Korea

Article References: Kim, M., & Choi, D.-J. (2026). Real-world patterns of herbal medicine-related adverse event reports and surveillance gaps in national pharmacovigilance: a 15-year retrospective analysis from a single integrative medical center in Korea. BMC Complementary Medicine and Therapies. https://doi.org/10.1186/s12906-026-05532-x

Image Credits: AI Generated

DOI: 10.1186/s12906-026-05532-x

Keywords: pharmacovigilance, herbal medicine, adverse event reporting, traditional Korean medicine, drug safety, non-licensed formulations, spontaneous reporting, causality assessment, real-world data, Korea, integrative medicine, surveillance gap