While modern machine learning has enabled flexible, high-dimensional modeling, reliable causal inference remains constrained by assumptions that are often unreliable in the settings where we most want answers: observational studies with unmeasured confounding, quasi-experiments with weak instruments and imperfect compliance, and structured systems with dependence across space, time, or networks. This dissertation develops machine learning methods for reliable causal inference under unreliable assumptions: methods that remain informative when identification is fragile, nuisance components must be learned flexibly, or the data-generating process departs from idealized models. Across settings, the recurring technical themes are orthogonalization and debiasing, sharp characterization of uncertainty (often through partial identification), and the careful use of structure to recover credible causal information. We start in Part I with observational settings where average-effect estimands can miss important distributional features and where unobserved confounding can undermine causal identification. Chapter 2 introduces a model-agnostic, debiased pseudo-outcome approach for learning rich causal functionals beyond averages, including conditional distributional treatment effects. Building on this machinery, Chapter 3 derives sharp, quasi-oracle bounds on heterogeneous effects under explicit sensitivity constraints. Chapter 4 extends partial-identification ideas to sequential decision-making by characterizing sharp bounds and efficient estimators for off-policy policy value in robust Markov decision processes. Next, Part II addresses causal learning from quasi-experiments, where identification is driven by imperfect experimental variation and compliance is often sparse or heterogeneous. Chapter 5 combines weak instrumental variation with observational data to estimate heterogeneous treatment effects, using observational structure to model heterogeneity while the instrument anchors identification. Chapter 6 then studies adaptive experimentation under noncompliance, introducing sequential encouragement designs that minimize asymptotic variance and support robust estimation with anytime-valid inference for safe monitoring and early stopping. Lastly, Part III turns to structured data with spatiotemporal and network dependence, where interference and time-varying confounding are central challenges. Chapter 7 introduces a neural framework for spatiotemporal causal inference with time-varying confounding. Chapters 8 and 9 further study causal inference under spatial and network dependence, developing interference-aware deconfounding methods and partial identification guarantees when exposure mappings may be misspecified. Taken together, this dissertation advances a unified perspective: causal inference in modern applications should not rely on idealized assumptions, but instead explicitly confront their limitations. By combining machine learning with orthogonal estimation, adaptive design, and partial identification, this work provides principled tools for reliable cause-and-effect analysis in observational, quasi-experimental, and structured domains.