Using L1 regularization in a predictive model always results in a model with fewer non-zero parameters compared to L2 regularization, regardless of the dataset.

The claim is not always true; L1 regularization encourages sparsity, but whether it produces fewer non‑zero coefficients than L2 depends on the data, the chosen penalty strength, and the model’s structure. With a large regularization parameter, both L1 and L2 can shrink many weights toward zero, but L1 can set some exactly to zero while L2 only shrinks them toward zero without eliminating them. For example, if a dataset has highly correlated features, an L1 penalty might drive one weight to zero while the other stays non‑zero, whereas an L2 penalty would keep both small but non‑zero. Thus, L1 can yield a sparser model in many cases, but it is not guaranteed to have fewer non‑zero parameters in every scenario.