Trustworthy artificial intelligence in advanced manufacturing

Additive manufacturing is relying more and more on artificial intelligence for prediction, control, and real-time decisions. Deployment, however, is still difficult. The main reason is that the underlying thermal, flow, and microstructural behavior can shift as the process changes, which makes reliable model performance hard to maintain. This review examines how trustworthy artificial intelligence (TAI) can be used to close that gap in digital twins for additive manufacturing across both metallic and ceramic systems. It frames trustworthiness as a technical condition for deployment and translates it into an AM-specific operational framework. From that point onward, it examines how current workflows handle reliability, uncertainty quantification, interpretability, explainability, robustness, generalizability, fairness, and transparency. A structured evaluation of sixty studies in both metals and ceramics additive manufacturing systems reveals a distinct trend. Robustness and generalizability receive much more attention than transparency, explainability, and fairness. The coverage is also stronger in metallic systems than in ceramic ones. To move beyond binary classification, the review further applies a structured evidence-depth rubric that distinguishes absent coverage, limited or implicit evidence, method-level evidence, and operational evidence for each TAI attribute. We further show that trustworthy deployment depends on more than predictive accuracy. It also depends on physically consistent models as well as updates that account for uncertainty. Decision support has to remain interpretable as well. Model transfer across machines and materials should be treated as controlled adaptation, where reusable physics, sensing, and validation elements are separated from material-specific features that require recalibration. The validation path must also support qualification and industrial deployment, leading to a more cohesive digital twin framework. It brings together physics-based models, reduced-order models, learning-based surrogates, multimodal sensing, and trust-aware evaluation. The main finding is that trustworthiness is the missing technical link between digital twins that can be confidently used in advanced manufacturing and promising AI-enabled demonstrations.