The value alignment problem in advisory AI: a systematic literature review

Abstract The rapid evolution of advisory artificial intelligence (AI) systems has intensified interest in the value alignment (VA) problem—how to ensure that AI-generated advice reflects human values, preferences, norms, and ethical standards. This systematic review, conducted according to PRISMA 2020 guidelines, synthesizes 83 peer-reviewed studies published between 2011 and 2025 that address the alignment of AI systems—particularly large language models (LLMs)—in advisory, decision-support, and recommendation contexts. Our thematic analysis identifies four dominant alignment approaches: personalized preference-based tuning, normative or principle-driven frameworks, fairness and cultural adaptation, and cognitive bias mitigation. While preference-based and normative strategies dominate the landscape, fairness- and cognition-focused methods remain underdeveloped. We find that alignment is not a static technical target but a dynamic, context-sensitive process shaped by evolving user values, cultural conditions, and domain-specific norms. Although alignment methods can enhance trust, personalization, and regulatory compliance, they also introduce risks, including hidden biases, overreliance, adversarial exploitation, and cognitive distortions. We conclude by outlining future research needs, including pluralistic alignment frameworks, standardized evaluation protocols, and interdisciplinary governance models to ensure that advisory AI remains trustworthy, fair, and socially responsive.