![[2602.20580] Personal Information Parroting in Language Models](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.20580] Personal Information Parroting in Language Models
Summary
This article examines the issue of personal information memorization in language models, highlighting the risks and proposing a detection suite to mitigate these concerns.
Why It Matters
As language models increasingly incorporate vast amounts of data, understanding their propensity to memorize personal information is crucial for privacy and ethical AI development. This research provides insights into how model size and training duration affect memorization, emphasizing the need for better data handling practices.
Key Takeaways
- Language models can memorize personal information, posing privacy risks.
- The study introduces a regex-based detector suite that outperforms existing solutions.
- Memorization rates increase with model size and training duration.
- Even smaller models can parrot personal information verbatim.
- Recommendations include filtering and anonymizing pretraining datasets.
Computer Science > Computation and Language arXiv:2602.20580 (cs) [Submitted on 24 Feb 2026] Title:Personal Information Parroting in Language Models Authors:Nishant Subramani, Kshitish Ghate, Mona Diab View a PDF of the paper titled Personal Information Parroting in Language Models, by Nishant Subramani and 2 other authors View PDF HTML (experimental) Abstract:Modern language models (LM) are trained on large scrapes of the Web, containing millions of personal information (PI) instances, many of which LMs memorize, increasing privacy risks. In this work, we develop the regexes and rules (R&R) detector suite to detect email addresses, phone numbers, and IP addresses, which outperforms the best regex-based PI detectors. On a manually curated set of 483 instances of PI, we measure memorization: finding that 13.6% are parroted verbatim by the Pythia-6.9b model, i.e., when the model is prompted with the tokens that precede the PI in the original document, greedy decoding generates the entire PI span exactly. We expand this analysis to study models of varying sizes (160M-6.9B) and pretraining time steps (70k-143k iterations) in the Pythia model suite and find that both model size and amount of pretraining are positively correlated with memorization. Even the smallest model, Pythia-160m, parrots 2.7% of the instances exactly. Consequently, we strongly recommend that pretraining datasets be aggressively filtered and anonymized to minimize PI parroting. Comments: Subjects: Computati...