4 AI Models Caught 4 Data Leaks in Real Datasets as Accuracy Hit 99%
Updated
Updated · KDnuggets · Nov 9
4 AI Models Caught 4 Data Leaks in Real Datasets as Accuracy Hit 99%
3 articles · Updated · KDnuggets · Nov 9
Summary
Four Claude models identified all four planted leakage patterns in real-world ML pipelines before execution, with the strongest model also surfacing extra flaws, baselines and realistic performance expectations.
The experiment paired each leaked pipeline with a clean version and found leakage could massively inflate results: Amazon stock-direction accuracy jumped to 97.0% from 51.6%, while a loan model's AUC rose to 0.986 from 0.637.
On delivery-time prediction, all four models caught a target-derived store-average feature, but only three also flagged a bad random split on timestamped data; the leaked setup improved RMSE to 862.7 seconds from 949.9.
In fraud detection, all four spotted that random row-level splitting let 1,950 of 2,246 users appear in both train and test sets, lifting AUC to 0.960 from 0.898 by partly recognizing known users.
The results suggest AI can serve as a cheap pre-deployment audit for common leakage patterns, though detection depth still varied by model and by how much of the pipeline reviewers could inspect.