The importance of convolutional neural networks is increasing in applications that require safety and mission-criticalness. Using accelerators for CNNs requires balancing various design parameters and reliability. While CNNs exhibit some fault-tolerant and error-resilient properties, their dependability must be evaluated considering the features of the hardware accelerator, especially for critical applications. A novel reliability technique for reducing soft errors in an AI computing core’s combinational logic is presented in this paper. The key contributions are: (1) A new approximate and adaptive multiplier is designed specifically for CNN accelerators, and it features an adaptive adder that optimizes adder resources and detects and mitigates faults, (2) the implementation and verification of the multiplier, and (3) Analyzing the multiplier’s reliability and comparing it to precise and approximate multipliers design with the help of standard CNN benchmarks. The proposed multiplier is able to achieve a reliability level that is similar to that of TMR based multipliers, while reducing its area by 66.42% and having a 39.84% lower power-delay product than the precise multiplier, as shown in the results.