The study investigates self‑attribution in language models—whether models explicitly signal when their responses depend on supplied information. Evaluating eleven frontier models, the authors observed that most rarely attribute truly influential information, while the remaining models tend to attribute non‑influential cues. They also found that inserting acknowledgment layers during reasoning does not propagate to the final answer, leaving attribution unchanged. To address this, the paper introduces CREDIT, a counterfactual training scheme that directly supervises final‑answer self‑attribution using automatically generated counterfactual labels, eliminating the need for human annotation. CREDIT is applied to four open‑source models, including two distilled reasoning variants, and yields substantial gains in attribution accuracy, outperforming both reasoning‑level fine‑tuning and prompting baselines. The method successfully discriminates between answers altered by a user suggestion and those the model would have produced independently, and these improvements partially generalize to unseen information types and prompts that present conflicting source evidence. By focusing supervision on the final answer with counterfactual supervision, CREDIT provides a scalable pathway to more explainable and accountable language models.