Just as Humans Need Vaccines, So Do Models: Model Immunization to Combat Falsehoods

TL;DR

The paper tackles the persistent challenge of AI models generating misinformation by proposing model immunization, a proactive training paradigm that treats fact-checked falsehoods as a vaccine signal. It introduces a four-stage pipeline—data curation with a quarantined falsehood repository, immunization fine-tuning with small falsehood doses, validation, and deployment—all governed by an explicit ethics-and-audit framework. A proof-of-concept with a 1.5B-parameter model demonstrates a substantial improvement in truthfulness on misinformation prompts (from ~60% to ~78%) with negligible impact on general QA accuracy, supporting the viability of targeted, negative supervision. If refined and scaled, this approach could become a practical component of AI development pipelines to enhance factual alignment, though it requires careful handling of coverage, unknown future misinformation, and ongoing governance.

Abstract

Generative AI models often learn and reproduce false information present in their training corpora. This position paper argues that, analogous to biological immunization, where controlled exposure to a weakened pathogen builds immunity, AI models should be fine tuned on small, quarantined sets of explicitly labeled falsehoods as a "vaccine" against misinformation. These curated false examples are periodically injected during finetuning, strengthening the model ability to recognize and reject misleading claims while preserving accuracy on truthful inputs. An illustrative case study shows that immunized models generate substantially less misinformation than baselines. To our knowledge, this is the first training framework that treats fact checked falsehoods themselves as a supervised vaccine, rather than relying on input perturbations or generic human feedback signals, to harden models against future misinformation. We also outline ethical safeguards and governance controls to ensure the safe use of false data. Model immunization offers a proactive paradigm for aligning AI systems with factuality.

Figures (6)

• Figure 1: Biological vaccination vs. model immunization. Controlled exposure to a weakened pathogen trains the immune system; similarly, controlled exposure to labeled falsehoods trains models to reject misinformation.
• Figure 2: Immunization fine-tuning: The model is periodically exposed to a small fraction of labeled falsehoods (orange) amidst mostly truthful data (teal), simulating a “vaccine dose.” This improves its resistance to misinformation.
• Figure 3: Overview of misinformation–defense techniques across the LLM lifecycle. Top: Timeline showing when each method applies. Bottom: Summary of technique properties.
• Figure 4: Conceptual Model Immunization Framework. Authentic true data and real-world falsehoods are collected and augmented with synthetic regulated false examples. All false items are isolated in a quarantined repository for review. During immunization fine-tuning, the model receives a 5–10 % micro-dose of these labelled falsehoods alongside clean data, yielding an immunized model. Validation then scores truthfulness, fairness, robustness, and feeds failures back for retraining. Finally, deployment enforces safety guards and continuous performance monitoring. All stages operate within an overarching governance and audit layer that supports iterative refinement.
• Figure 5: Governance workflow. Fact-checked falsehoods enter a quarantined dataset after independent reviews; audit logs at each arrow ensure traceability from source to training.