Text this: Reducing risk for assistive reinforcement learning policies with diffusion models