FROM COUNTERFACTUAL EXPLANATIONS TO CAUSAL ALGORITHMIC RECOURSE: A REVIEW OF METHODS UNDER IMPERFECT CAUSAL KNOWLEDGE

Counterfactual explanations and algorithmic recourse aim to tell individuals how to change an unfavourable automated decision, but recommendations are useful only if they are feasible, actionable, and causally valid. This review examines how recent methods meet these requirements and where they fall short. The literature was searched mainly in Google Scholar for papers published between 2018 and 2026, and 20 studies on causal counterfactual explanations and algorithmic recourse were reviewed. The analysis identified three gaps. First, under real-world constraints, methods such as DiCE and FACE produce diverse, feasible, and actionable counterfactuals, but without modelling causal relationships, and later work shows that recourse can be fragile or invalidated over time. Second, many methods depend on a fully known structural causal model (SCM) or causal graph; recent work relaxes this through probabilistic, learned, and graph-free approaches, yet these still rely on assumptions about the available causal information. Third, hidden confounding and non-identifiability make causal effects hard to estimate; bounding approaches and multi-dataset methods address this, but depend on known confounding structure or compatible data, and learned SCMs offer limited guarantees of reliable counterfactual inference. Overall, the field has moved from single counterfactual explanations toward methods that operate under partial causal knowledge, yet no reviewed method resolves all three gaps at once. Future research should address feasibility, confounding, and incomplete causal knowledge together.