<p>ChatGPT’s growth highlights the importance of Large Language Models (LLMs) in AI innovation, driving a surge in AI-generated content (AIGC). Federated Learning (FL), valued for privacy and cooperation, is drawing interest for using distributed data while ensuring privacy, critical for AI scalability. Yet, incorporating FL into AIGC is challenging, particularly in building trust across LLM training. This paper introduces GuardChain, a novel framework to fortify trustworthiness in every segment of FL-empowered AIGC training processes. We focus on three key designs: (1) Integrating the Nostr protocol into the FL paradigm to facilitate trustworthy cross-data validation. (2) The deployment of a dual-layered (on-chain and off-chain) mechanism for rigorous consistency checks and the identification of adversarial adapters. (3) The enhancement of model aggregation trustworthiness via a dynamic rotating election protocol for the selection of credible aggregators, augmented by a ‘Sandwich’ smart contract verification methodology. Our empirical studies reveal that GuardChain markedly surpasses existing baseline methods in mitigating diverse adversarial attacks and expediting trust verification processes. Comparative performance experiments reveal that GuardChain efficiently manages various workloads and hostile threats, with blockchain-based verification completed within 1–10 seconds, concurrently improving BLEU and ROUGE scores by 0.14 to 0.17 on the standard public dataset, thereby maintaining its reliability.</p>