Abstract The philosophical notion of possible worlds has long been debated as a metaphysical abstraction. Machine learning (ML) can present an empirically grounded form of this plurality when multiple models trained on identical data achieve equivalent performance while inferring incompatible internal representations. This paper develops an algorithmic lens on this phenomenon by treating such cases as evidence-compatible (epistemic) possible worlds. We demonstrate how competing ML and causal algorithms can identify mutually incompatible causal graphs, each statistically credible yet diverging in mechanistic, interventional, and counterfactual commitments. We also show how this phenomenon extends to large language models (LLMs), where hallucinations can be interpreted not only as errors but as internally coherent alternative textual world-states sampled from the same parameterized distribution, with varying alignment to our reference world. By framing models as evidence-compatible worlds, we provide an empirical grounding on underdetermination (a modernized Duhem–Quine thesis) and argue that accuracy alone cannot adjudicate between incompatible explanations. We conclude by outlining the ethical implications of privileging one algorithmic narrative over rivals that remain comparably supported by the data.
M.Z. Naser (Tue,) studied this question.