Wireless Hallucination in Generative AI-enabled Communications: Concepts, Issues, and Solutions

TL;DR

This work identifies wireless hallucination as GenAI outputs in wireless tasks that violate real-world constraints. It builds a three-tier mitigation framework—data augmentation/cleaning, model-level robustness via RL, attention, and adversarial training, and post-generation validation with interactive AI and mixture-of-experts—and demonstrates it with a GDM-based channel-estimation case. The integrated approach combines GAN-based data augmentation, attention-enhanced diffusion, and LLM-driven MoE gating, achieving a NMSE reduction of $0.19$ at $0$ dB SNR and stronger benefits in low-SNR scenarios. The findings offer practical pathways to deploy GenAI in wireless systems with reduced hallucinations and enhanced reliability, supported by datasets and reproducible experiments.

Abstract

Generative AI (GenAI) is driving the intelligence of wireless communications. Due to data limitations, random generation, and dynamic environments, GenAI may generate channel information or optimization strategies that violate physical laws or deviate from actual real-world requirements. We refer to this phenomenon as wireless hallucination, which results in invalid channel information, spectrum wastage, and low communication reliability but remains underexplored. To address this gap, this article provides a comprehensive concept of wireless hallucinations in GenAI-driven communications, focusing on hallucination mitigation. Specifically, we first introduce the fundamental, analyze its causes based on the GenAI workflow, and propose mitigation solutions at the data, model, and post-generation levels. Then, we systematically examines representative hallucination scenarios in GenAI-enabled communications and their corresponding solutions. Finally, we propose a novel integrated mitigation solution for GenAI-based channel estimation. At the data level, we establish a channel estimation hallucination dataset and employ generative adversarial networks (GANs)-based data augmentation. Additionally, we incorporate attention mechanisms and large language models (LLMs) to enhance both training and inference performance. Experimental results demonstrate that the proposed hybrid solutions reduce the normalized mean square error (NMSE) by 0.19, effectively reducing wireless hallucinations.

Figures (4)

• Figure 1: A summary of different types of GenAI models, potential causes and manifestations of hallucination.
• Figure 2: The principles, hallucination manifestations, and causes of hallucination in the application of GenAI to wireless communications from the perspectives of the physical layer, data link layer, and network layer.
• Figure 3: Part A illustrates the proposed integrated solution for reducing wireless hallucination in channel estimation. At the data level, a GAN enhances uneven training data. At the model level, an attention mechanism captures latent channel distribution features. At the post-generation level, an LLM-enhanced MoE selects models that align with user requirements for for channel estimation. Part B outlines the workflow: Upon receiving pilot signals, the LLM analyzes and infers user requirements, subsequently selecting the suitable expert for channel estimation.
• Figure 4: Experimental results. (A): The training loss curves of the proposed integrated mitigation strategy and other baselines over epochs. (B): NMSE performance of the proposed integrated mitigation strategy and other baselines over SNR. Note that we adopt the following baselines: Hallucination Strategy, a diffusion-based channel estimation method without any hallucination mitigation; Integrated Strategy-w/o Attention, which employs GAN-based data augmentation training but excludes the attention mechanism; Integrated Strategy-w/o LLM, which uses random expert selection during inference.