Using homomorphic encryption (HE) as an underlying technology for trustworthy elections is IMO a much better application of HE versus using HE for ML/AI applications, where performance requirements are incomparably higher. While there seems to be some limited progress on the HE acceleration front, potential improvements are limited by datasets size and algorithm complexity issues. I would argue that, for secure ML/AI applications, using technologies based on microprocessor-level secure instruction codes (e.g., Intel Software Guard Extensions aka SGX enclaves) is a much better option.