Artificial Intelligence (AI) holds significant promise for healthcare but often struggles to transition from development to clinical integration. This paper argues that Learning Health Systems (LHS)-socio-technical ecosystems designed for continuous data-driven improvement-provide a potential "glide path" for safe, sustainable AI deployment. Just as modern aviation depends on instrument landing systems, the safe and effective integration of AI into healthcare requires the socio-technical infrastructure of LHSs, that enable iterative development and monitoring of AI tools, integrating clinical, technical, and ethical considerations through stakeholder collaboration. They address key challenges in AI implementation, including model generalizability, workflow integration, and transparency, by embedding co-creation, real-world evaluation, and continuous learning into care processes. Unlike static deployments, LHSs support the dynamic evolution of AI systems, incorporating feedback and recalibration to mitigate performance drift and bias. Moreover, they embed governance and regulatory functions-clarifying accountability, supporting data and model provenance, and upholding FAIR (Findable, Accessible, Interoperable, Reusable) principles. LHSs also promote "human-in-the-loop" safety through structured studies of human-AI interaction and shared decision-making. The paper outlines practical steps to align AI with LHS frameworks, including investment in data infrastructure, continuous model monitoring, and fostering a learning culture. Embedding AI in LHSs transforms implementation from a one-time event into a sustained, evidence-based learning process that aligns innovation with clinical realities, ultimately advancing patient care, health equity, and system resilience. The arguments build on insights from an international workshop hosted in 2025, offering a strategic vision for the future of AI in healthcare.

Deep latent variable models, trained using variational autoencoders or generative adversarial networks, are now a key technique for representation learning of continuous structures. However, applying similar methods to discrete structures, such as text sequences or discretized images, has proven to be more challenging. In this work, we propose a more flexible method for training deep latent variable models of discrete structures. Our approach is based on the recently proposed Wasserstein Autoencoder (WAE) which formalizes adversarial autoencoders as an optimal transport problem. We first extend this framework to model discrete sequences, and then further explore different learned priors targeting a controllable representation. Unlike many other latent variable generative models for text, this adversarially regularized autoencoder (ARAE) allows us to generate fluent textual outputs as well as perform manipulations in the latent space to induce change in the output space. Finally we show that the latent representation can be trained to perform unaligned textual style transfer, giving improvements both in automatic measures and human evaluation.