Efficient multiple-input multiple-output (MIMO) detection algorithms with satisfactory performance and low complexity are critical for future multi-antenna systems to meet the high throughput and ultra-low latency requirements in 5G and beyond communications. In this paper, we propose a low complexity graph neural network (GNN) enhanced approximate message passing (AMP) algorithm, AMP-GNN, for MIMO detection. The structure of the neural network is customized by unfolding the AMP algorithm and introducing the GNN module to address the inaccuracy of the Gaussian approximation for multiuser interference cancellation. Numerical results will show that the proposed AMP-GNN significantly improves the performance of the AMP detector and achieves comparable performance as the state-of-the-art deep learning-based MIMO detectors but with reduced computational complexity.