Artificial intelligence (AI) is increasingly advocated as a solution to the growing volume, complexity, and heterogeneity of pharmacovigilance data, given its ability to process spontaneous reports, electronic health records, narratives, and digital media, which has transformed expectations for drug safety surveillance. Yet, questions remain about whether AI-generated safety signals are valid, unbiased, causally meaningful, and acceptable for regulatory decision-making. This critical review evaluates the validity of such signals from 2017 to 2026, focusing on four key dimensions: algorithmic bias, causal inference, signal detection methodology, and regulatory acceptance. Using a critical narrative review approach, the literature on AI, machine learning, natural language processing, deep learning, and large language model applications in pharmacovigilance was synthesized, prioritizing studies addressing signal detection, case processing, causality assessment, validation, explainability, bias, and regulatory use, with evidence interpreted analytically rather than pooled quantitatively due to methodological heterogeneity. Findings indicate that AI can accelerate adverse event processing, extract safety information from unstructured data, and support earlier signal prioritization, but recurring concerns persist regarding retrospective validation, database-specific learning, unmeasured confounding, weak causal reasoning, and limited assessment of demographic fairness. Regulatory acceptance remains cautious, as many AI-generated signals lack transparent evidence chains and clinically adjudicated confirmation. Therefore, AI-generated safety signals should not be treated as self-validating merely because of computational sophistication; their credibility depends on bias correction, causal augmentation, external validation, interpretability, and prospective assessment in routine pharmacovigilance workflows, making AI best understood as an adjunct to, rather than a replacement for, expert pharmacovigilance judgment.