Few-shot AI approach helps decode vaccine hesitancy narratives on social media

Vaccine hesitancy is increasingly unfolding as a narrative problem rather than a simple information deficit, with online users turning personal experience, selective scientific claims, distrust of institutions and policy resistance into arguments against vaccination. In a study published in Big Data and Cognitive Computing, researchers show how artificial intelligence (AI) can help public health teams detect and interpret these narratives at scale, using South Carolina as a case example of a wider global challenge.

The study, titled Teaching AI to Decode Vaccine Hesitancy Narratives: A Few-Shot Learning and Topic Modeling Approach, analyzed 298,356 COVID-19 vaccine-related X posts geolocated to South Carolina between June 2021 and May 2022, applying zero-shot and few-shot learning with large language models and topic modeling to classify vaccine stance and identify dominant narratives behind hesitancy.

AI offers a faster way to read vaccine hesitancy at scale

Vaccine debates now move faster than traditional monitoring systems can track. Online conversations unfold across fragmented platforms, often using sarcasm, anecdotes, selective evidence, political language and emotionally charged claims. Manual coding can capture some of this complexity, but it is slow, costly and difficult to scale during a public health crisis.

To overcome that barrier, the researchers tested whether LLMS could classify vaccine-related social media posts with limited labeled examples. The study used zero-shot and few-shot learning. Zero-shot classification asked models to classify content without task-specific examples, while few-shot learning provided a small set of examples and rules inside the prompt to guide the model.

The research tested instruction-tuned models including Mistral-7B, Meta-Llama-3.1 and DeepSeek-7B, along with a DeBERTa-based zero-shot classifier. The classification task sorted posts into three categories: pro-vaccine, vaccine hesitant and irrelevant. A manually classified sample of 1,000 posts served as the validation benchmark.

Meta-Llama-3.1 performed best for the study's central task, reaching 93 percent agreement with human annotators for vaccine-hesitant posts. The model was then used to classify the full dataset, identifying 179,876 vaccine-hesitant posts. Those posts were later analyzed through Latent Dirichlet Allocation topic modeling to uncover recurring narratives.

The South Carolina dataset was used as a local case, but the paper frames the problem in broader terms. Vaccine hesitancy remains a global public health challenge, and online discourse often links local grievances to national or international debates. South Carolina provided a focused setting in which researchers could test a scalable method and examine how broader vaccine skepticism appeared in a defined geographic context.

The method matters because vaccine hesitancy is often too complex for simple keyword tracking. A post may mention vaccines, mandates or public health agencies without being hesitant. Another may express skepticism indirectly through irony, comparison or claims about institutional trust. Few-shot prompting allowed the researchers to write clearer decision rules and use examples to reduce misclassification.

The study does not stop at identifying whether a post is hesitant. By combining AI classification with topic modeling and qualitative review, the researchers examined what kinds of arguments, emotions and storylines made up vaccine skepticism. That step is important for public health officials because knowing that hesitancy exists is less useful than knowing how people justify it.

The final model identified five main narrative themes: skepticism of vaccine efficacy, comparative framing, scientific justification, disapproval of regulations and distrust. Skepticism of vaccine efficacy dominated the dataset, appearing in 73,122 posts. Comparative framing appeared in 8,554 posts, scientific justification in 7,823, disapproval of regulations in 5,012 and distrust in 4,438.

The large gap between the leading theme and the others suggests that vaccine effectiveness was the central issue in the analyzed discourse. But the smaller themes show that efficacy doubts were not isolated. They were tied to broader concerns about policy, authority, science and institutional credibility.

Hesitancy narratives use experience, comparison and selective science

One of the key findings is that vaccine hesitancy should not be treated only as misinformation. Many hesitant posts did not simply make false claims. They assembled arguments through personal stories, selective evidence, distrust and comparisons that made skepticism appear rational to the users expressing it.

The dominant theme, skepticism of vaccine efficacy, included claims about breakthrough infections, alleged side effects, vaccine injury, illness after vaccination and doubts about whether vaccination reduced transmission. Users often relied on personal or nearby experience, such as stories involving friends, family members or community members. These accounts were used to challenge broader public health messaging.

Lived experience can outweigh statistical evidence in personal risk decisions. A person who knows someone who fell ill after vaccination may treat that case as more persuasive than population-level data. Public health messaging that ignores this emotional and experiential logic may fail to reach hesitant audiences.

The theme also included counter-narratives about transmission. Some users argued that vaccinated people were spreading COVID-19 and contributing to the pandemic. This reversed the public health framing of vaccination as a protective intervention and reinterpreted vaccination as a possible source of harm or failure.

Another strand involved skepticism about reporting accuracy. Users accused media organizations, public health agencies and platforms of emphasizing unvaccinated deaths while ignoring vaccinated people who became sick or died. This narrative made official information appear selective, strengthening distrust in institutions.

Distrust was identified as a standalone theme. Posts questioned pharmaceutical companies, regulators, government agencies and media organizations. The rhetoric often portrayed these actors as financially driven, politically compromised or corrupt. Once that distrust was established, even accurate information from official sources could be rejected as self-serving or incomplete.

Disapproval of regulations focused on mandates, masks and institutional rules. Users framed these policies as inconsistent, coercive or politically motivated. In this narrative, rising cases despite mandates were used as evidence that public health rules had failed. The criticism was not only about vaccines but about perceived overreach by authorities.

The scientific justification theme is one of the most important findings. Users did not always reject science directly. Many tried to claim scientific authority for skepticism by citing studies, preprints, natural immunity, antibodies or statistical arguments. These posts framed vaccine hesitancy as informed dissent rather than ignorance.

This finding complicates standard public health assumptions. If skeptical users see themselves as evidence-seeking and scientifically literate, fact-checking alone may not shift their views. They may reject corrections if they believe institutions are suppressing inconvenient evidence or selectively presenting science.

Comparative framing also shaped vaccine skepticism. Users compared infection rates before and after vaccine rollout, or contrasted highly vaccinated places with places experiencing case surges. These comparisons were used to question whether vaccines had delivered promised benefits. The argument often relied on visible numbers or broad comparisons without full epidemiological context, but it carried persuasive force because it appeared data-driven.

These themes collectively show that vaccine hesitancy is not one belief system, but a cluster of narratives that can overlap. A user may distrust institutions, oppose mandates, point to breakthrough infections and cite selective studies in the same broader argument. Public health responses that address only one component may miss the structure of the hesitation.

Public health messaging must respond to narratives, not just claims

The analysis found that vaccine-hesitant discourse was especially active from mid-2021 to mid-2022, with skepticism of vaccine efficacy peaking in late 2021. That timing coincided with booster campaigns and intense debates over mandates. The finding shows how online vaccine narratives respond to public events, policy shifts and changing pandemic conditions.

Other themes were less frequent but event-driven. Disapproval of regulations and comparative framing showed smaller spikes around periods of policy conflict. Distrust also grew as users linked vaccine debates to claims about pharmaceutical profit, political motives and changing public health guidance.

After the Omicron surge in early 2022, the overall volume of discourse declined. The paper suggests this may reflect pandemic fatigue and shifting attention to other public issues. But a decline in posting does not necessarily mean distrust has been resolved. It may only mean that public attention has moved elsewhere.

Vaccine hesitancy cannot be addressed only through more information. Facts remain essential, but hesitant audiences may filter facts through distrust, identity and personal experience. When users frame skepticism as scientific, independent or morally justified, simple correction may be seen as another act of institutional control.

A stronger communication strategy would engage the narratives themselves. That means acknowledging concerns about side effects without validating false claims, explaining uncertainty without appearing evasive, clarifying how evidence changes over time and showing why population-level evidence may differ from individual anecdotes. It also means rebuilding credibility with communities before a crisis forces rapid persuasion.

The study supports a glocal approach to risk communication. Vaccine hesitancy is a global issue, but it takes local forms. South Carolina served as one example of how regional politics, community trust, health access and online conversation can shape the expression of broader skepticism. Similar themes may appear elsewhere, but tone, intensity and triggers will vary by place.

The research also points to a practical role for AI in public health. Few-shot learning can help agencies monitor fast-moving discourse without waiting for large labeled datasets. Topic modeling can help identify which narratives are rising, which ones are fading and which ones may need tailored communication responses.

Furthermore, the authors note several constraints. The analysis relied only on geotagged X posts from South Carolina, which may not represent broader populations or conversations on platforms such as Facebook, Reddit, TikTok or Threads. The study also focused on text, excluding images, videos and memes, even though visual content can strongly influence vaccine discourse.

Prompt sensitivity is another limitation. Few-shot learning depends on prompt wording, examples and classification rules. The researchers used structured prompts, balanced examples and deterministic decoding to improve consistency, but LLM-based classification still requires careful validation.