TuringTest: Scoring Recommendations by Human–Machine Preference Divergence

uring's (1950) original test asks whether a machine's outputs are behaviourally indistinguishable from a human's. Applied to recommendation: a system passes the recommender Turing Test if its lists are indistinguishable from those a thoughtful human curator would produce. The paper formalises this by decomposing recommendation into two fundamentally different reasoning modes — human-analogy (UserCF: memory-based, local similarity, as humans reason) and machine (SVD: global algebraic decomposition, no analogical reasoning). A (u, i) = ReLU (r̂ʰumanᵤi − r̂ᵐachineᵤi) (Turing advantage — human > machine signal) TT (u, i;α) = r̂ʰumanᵤi · (1 + α · MinMax (A (u, i) ) ) The Turing bonus amplifies items where human-like reasoning dominates algorithmic reasoning. Theorem 1 proves this captures exactly the items in the orthogonal complement of the SVD factor space: items that matrix factorisation cannot represent but that neighbourhood reasoning correctly surfaces. Proposition 1 and its Corollary identify items in the SVD null space as the prime beneficiaries — items with too few interactions to acquire reliable factor representations, but with strong local neighbourhood signals. T3 is the only method in the benchmark that consistently matches or exceeds UserCF across all three datasets. On ML-100K it achieves P@10 = 0. 0976, strictly above UserCF (0. 0965). On Book-Crossing it achieves P@10 = 0. 0073, joint-best with ANP. The Turing bonus adds genuine predictive value — not just noise on top of UserCF.