Pith. sign in

REVIEW 4 major objections 6 minor 12 references

How Private are Language Models in Abstractive Summarization?

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read Language models routinely leak personal identifiers when summarizing medical and legal documents, while expert human summaries are much more protective.

desk verdict Solid empirical study with a reusable dataset; the headline leakage numbers rest on an unvalidated detector that misses the most common PII type, so the quantitative claims need another pass. read the letter →

arxiv 2412.12040 v2 pith:C64IBYYO submitted 2024-12-16 cs.CL

classification cs.CL
keywords privacy-preservingsummarizationpersonallyidentifiableinformationabstractiveinstructionfine-tuningpseudonymizationmedicalNLPlegallargelanguagemodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper asks whether language models can be trusted to summarize medical records and legal case files without copying out personal details. By taking real anonymized records, reinserting realistic synthetic personal information, and asking models to summarize them, the authors find that models regularly leak names, dates, locations, ages, and especially gender pronouns into their summaries. Two medical doctors who wrote private summaries of the same documents leaked almost nothing, apart from occasional pronouns. The authors also show that instruction fine-tuning on their pseudonymized data makes smaller open-weight models about as private as large closed-source models, while improving summary quality. If correct, the result means current summarization systems are not safe for direct deployment in privacy-sensitive settings without dedicated anonymization steps.

What carries the argument

The evaluation machinery is a three-part pipeline. First, source documents are pseudonymized: a synthetic-profile generator creates realistic profiles with name, age, gender, race, birth date, birth location, and residence, and GPT-4o reinserts those values into redacted clinical and legal texts, with a BLEU threshold plus manual checks used as quality control. Second, leakage is quantified by three metrics: PTR, the share of private tokens from the source that appear in the summary; LDR, the share of summaries containing at least one leaked private token; and TPR, the share of source PII spans that reappear in the summary, all computed with a GPT-4o-based PII detector prompted in the style of prior de-identification work. Third, the same pseudonymized data is used to instruction fine-tune open-weight models, with separate medical and legal models, so that privacy-preserving behavior is learned rather than prompted. The gold-standard comparison comes from two medical doctors who wrote private summaries of 74 documents and annotated leaked spans, providing the expert baseline.

What would settle it

Re-annotate a random sample of machine-generated and doctor-written summaries with trained human annotators using the same five PII categories, then recompute PTR, LDR, and TPR from the human spans. If humans find models are no leakier than the doctors, the central gap claim fails; if humans find even more leakage, especially of gender pronouns, the claim is strengthened.

Watch

Extended reading notes

Core claim

The paper claims that language models, when prompted to summarize pseudonymized medical and legal source documents, frequently copy personally identifiable information from the source into their summaries, whereas expert-written privacy-preserving summaries leak far less. The authors construct this comparison by reinserting synthetic PII into two publicly released anonymized datasets, recruiting medical doctors to write gold-standard private summaries and to annotate leaked spans, and evaluating six models under six prompting strategies plus instruction fine-tuning. Their headline result is that leakage is common and uneven: gender pronouns are the most frequently leaked category, chain-of-thought prompting increases leakage relative to dedicated anonymization steps, and instruction fine-tuning on the new pseudonymized data lets open-weight models such as Llama-3.3-70B reach privacy protection comparable to larger closed-source models while improving summary quality. A human evaluation of 100 document-summary pairs found 95 summaries free of source-linked PII, with slight leakage attributed to the fine-tuned open model and none to the closed-source model, and participants preferred the open model's summaries.

Load-bearing premise

The comparison depends on the automatic GPT-4o PII detector catching every leaked identifier in the summaries, and the paper's own examples show that detector misses gender pronouns, the category it finds most often leaked, so the measured leakage rates may understate the true rates.

Editorial extensions

If this is right

  • Unconstrained zero-shot summarization leaks PII in nearly all documents, with leaked-document ratios close to 100 percent in the medical task.
  • Explicit privacy instructions reduce leakage but do not eliminate it, and adding in-context examples can increase leakage for some closed-source models.
  • Anonymizing after summarizing is more effective at suppressing PII than anonymizing before summarizing.
  • Chain-of-thought prompting improves summary quality but is less privacy-safe than the two-stage anonymization approaches.
  • Instruction fine-tuning on the pseudonymized data gives open-weight models privacy metrics comparable to or better than frontier closed-source models while raising ROUGE and BERTScore.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the paper's own qualitative analysis shows the automatic detector and an off-the-shelf tagger miss gender pronouns, the reported leakage figures likely undercount the most common leakage category; a benchmark built on human-annotated spans would give a sterner test and may widen the human-model gap.
  • The order effect the paper observes — summarizing before anonymizing beats anonymizing before summarizing — suggests a practical pipeline design: let the model summarize the original text, then run a dedicated redaction pass, rather than relying on the model to redact first.
  • The pseudonymization recipe of synthetic profiles inserted by GPT-4o with a BLEU sanity threshold is a reusable way to create privacy-preserving summarization training data in other sensitive domains, but its realism has only been spot-checked, so transfer should be validated by domain experts.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. This paper asks whether language models leak personally identifiable information when summarizing medical and legal documents. The authors construct pseudonymized versions of the Discharge Me! and AsyLex datasets by injecting synthetic profiles with GPT-4o, then evaluate two closed-source and four open-weight models under six prompting methods and instruction fine-tuning. They measure leakage with a GPT-4o-based PII detector using PTR, LDR, and TPR, and supplement this with a human evaluation of two selected models. The central claims are that LMs frequently leak PII, that expert doctors produce substantially more private summaries, and that instruction fine-tuning on the pseudonymized data brings open-weight models close to frontier closed-source models.

Significance. If these claims hold, the paper makes a useful contribution: it releases new pseudonymized datasets, expert-written private summaries, and annotations; it provides the first systematic human-expert comparison for privacy in summarization; and it proposes a practical fine-tuning recipe for open-weight models. The human evaluation supports the qualitative direction of the findings. However, the headline quantitative comparisons rest on an unvalidated automatic PII detector that the paper itself shows misses the most frequently leaked PII type, and several text/table inconsistencies undermine the strength of the stated conclusions. The contribution is significant enough to warrant revision rather than rejection, provided the measurement and reporting issues are addressed.

major comments (4)
  1. [§4.4, Figure 5, Table 3] The PTR, LDR, and TPR metrics all rely on the GPT-4o-based detector described in §4.4 and prompted in Figure 5, but this detector is never validated against human annotations. The paper's own qualitative analysis (Table 3, rows 5 and 6) shows that both GPT-4o and Presidio fail to identify gender pronouns as PII even though the detection prompt explicitly targets GENDER, and §6 with Table 4 reports gender as the most leaked category. Consequently, the reported leakage rates undercount the dominant PII type, and model rankings can be distorted because models differ in which PII types they leak. Please validate the automatic detector (e.g., precision/recall against human annotations on a random sample) or recompute the headline metrics from human annotations for all models and conditions.
  2. [§6, Table 4 versus §4.5] The gold-standard comparison of doctors' summaries with GPT-4o and Llama-3.3-70b appears to mix annotation instruments. The human evaluation described in §4.5 covers only DeepSeek-Chat and IFT+Llama-3.3-70B, so the TPR values for GPT-4o and Llama-3.3-70b in Table 4 cannot come from that protocol; they can only come from the unvalidated GPT-4o detector applied to the same 74 documents. This asymmetry undermines the headline contrast between human and LM privacy protection and the conclusion in §6 that frontier LMs are close to human performance. Please state explicitly how each TPR in Table 4 was computed and, ideally, measure humans and LMs with the same annotation instrument.
  3. [§5.2 and Tables 11–12] The text claims that 'PTR decreases across all models in the medical domain' after IFT, but the appendix shows the opposite for several model/prompt combinations (e.g., IFT-Llama-3.1 8B 0-Shot Priv Sum PTR 25.74 vs Llama-3.1 8B 17.60; IFT-Qwen2.5 14B 18.52 vs Qwen2.5 14B 6.65 in Table 11). Similarly, 'Deepseek-Chat exhibiting a 2% increase in PTR when using in-context samples' is not supported by the reported values (1.85 to 1.89 in Table 11; 1.79 to 1.91 in Table 12). Please either correct these statements or clarify which baselines and conditions are being compared, since the abstract's claim that IFT makes smaller models comparable to frontier models depends on these comparisons.
  4. [§5.1–§5.3, §6] All headline comparisons are point estimates without confidence intervals, significance tests, or effect-size measures. Given that leakage rates are low and the test sets are relatively small (e.g., 74 documents for the gold standard, 100 for human evaluation), the claim that GPT-4o is close to human performance (8% vs 4% gender TPR) is not distinguishable from noise. Please add uncertainty quantification (e.g., bootstrap confidence intervals) or appropriate statistical tests for the main privacy comparisons.
minor comments (6)
  1. [Abstract] The phrase 'expert human expert performance' contains a duplicated word; it should read 'expert human performance' or 'human expert performance.'
  2. [Figure 6] The instruction 'The data has been redacted' is duplicated in the pseudonymization prompt template; please remove the redundant sentence.
  3. [§3.2] The BLEU threshold of 20% is introduced after manual checking of 200 documents, but no justification or sensitivity analysis is provided; please clarify how this threshold was chosen and whether the main results are robust to it.
  4. [Table 4] The model name 'Llama-3.3-70b' uses inconsistent capitalization compared to 'Llama-3.3-70B' elsewhere in the paper; please standardize.
  5. [References] The reference for P˘ais, et al. appears to have a formatting artifact in the author field; please fix the bibliographic entry.
  6. [§3.3] The paper states that 'we perform our initial analysis using Presidio' but the main evaluation later switches to GPT-4o; please clarify whether the dataset stratification PII counts come from Presidio and whether the GPT-4o detector uses the same PII taxonomy.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an empirical measurement study whose claims rest on held-out evaluations, not on definitions, fitted predictions, or self-citation.

full rationale

The paper's claims are empirical measurements, not derivations. There is no equation in which a predicted quantity is constructed from its own input, no fitted parameter that is subsequently reported as a prediction, and no load-bearing self-citation: the PII-detection prompt is credited to Kim et al. (2024), an external source, and no uniqueness theorem is invoked. The closest candidate is the use of GPT-4o both to synthesize PII in Section 3.2 and to detect leaked PII in Section 4.4, but the detection step is applied to independently generated summaries and is therefore not equivalent by construction to the pseudonymization step; at worst it is a shared-instrument validity risk, which the paper partially acknowledges by reporting that GPT-4o and Presidio both miss gender pronouns (Table 3, rows 5-6), the most leaked category (Table 4 and Section 6). The IFT results in Section 5 are standard train/test evaluations on stratified held-out splits, not fits renamed as findings. The human-vs-LM comparison in Table 4 may mix annotation protocols, but instrument asymmetry is a measurement-validity concern, not circularity. Per the review standard requiring a specific reduction (Eq. X = Eq. Y by construction, or a fitted input renamed as prediction), no circular step is established.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claims rest on empirical measurements rather than a mathematical derivation. The free parameters are the data-construction thresholds (BLEU 20%, frequency 20) that shape the pseudonymized corpus, and the key loaded assumptions are the reliability of GPT-4o as a PII detector and the realism of synthetic pseudonymization, both of which the paper only partially supports.

free parameters (2)
  • Minimum PII frequency threshold = 20 occurrences
    PII types appearing fewer than 20 times are excluded to keep categories consistent, following prior work. This choice determines which PII classes are measured and could mask rare but sensitive leakage.
  • BLEU quality threshold for pseudonymization = 20%
    Pseudonymized documents with BLEU below 20% against the original are discarded after manual assessment of 200 documents. The threshold tunes the realism of the synthetic data.
assumptions (3)
  • domain assumption GPT-4o-based automatic detection accurately identifies leaked PII spans in summaries.
    The PTR, LDR, and TPR metrics are computed with a GPT-4o detection prompt shown in Figure 5. The paper's qualitative examples (Table 3, rows 5 and 6) demonstrate the detector misses gender pronouns, so this assumption is only partially met and directly affects the headline numbers.
  • domain assumption Reinserting synthetic PII with GPT-4o into redacted documents creates realistic analogues of private medical and legal records.
    All downstream measurements are on these synthetic documents. Real PII appears in more varied linguistic contexts and with different co-occurrence patterns, so leakage rates on genuine data may differ, a limitation the authors acknowledge.
  • domain assumption Writing summaries under an explicit private-summary instruction captures real expert privacy-preserving summarization behavior.
    Two doctors were recruited to produce private summaries of 74 pseudonymized records and then annotate leaked PII. Expert behavior in routine practice without such an explicit instruction might differ.

how reviews work

0 comments
Cite this review

Pith. "Pith review of How Private are Language Models in Abstractive Summarization?." pith.science (2026). https://pith.science/paper/C64IBYYO

@misc{pith2026241212040,
  author       = {Pith},
  title        = {Pith review of: How Private are Language Models in Abstractive Summarization?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C64IBYYO}},
  note         = {Machine review of arXiv:2412.12040}
}
read the original abstract

In sensitive domains such as medical and legal, protecting sensitive information is critical, with protective laws strictly prohibiting the disclosure of personal data. This poses challenges for sharing valuable data such as medical reports and legal cases summaries. While language models (LMs) have shown strong performance in text summarization, it is still an open question to what extent they can provide privacy-preserving summaries from non-private source documents. In this paper, we perform a comprehensive study of privacy risks in LM-based summarization across two closed- and four open-weight models of different sizes and families. We experiment with both prompting and fine-tuning strategies for privacy-preservation across a range of summarization datasets including medical and legal domains. Our quantitative and qualitative analysis, including human evaluation, shows that LMs frequently leak personally identifiable information in their summaries, in contrast to human-generated privacy-preserving summaries, which demonstrate significantly higher privacy protection levels. These findings highlight a substantial gap between current LM capabilities and expert human expert performance in privacy-sensitive summarization tasks.

Figures

Figures reproduced from arXiv: 2412.12040 by the authors.

Figure 1
Figure 1. Prompting GPT-4o to generate a private sum [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An overview of the pseudonymization process. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Prompt templates for summarization. source documents in the test set (Dt). This allows us to quantify the breadth of the privacy concerns across a given dataset. Finally, we use the True Positive Rate (TPR) to identify when a PII span appears in both the source and the summary. All metrics are averaged across the test set. Automatic PII Leakage Detection. We use GPT￾4o to automatically identify leaked PII tokens in … view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Results of the private summary experiments. Top two rows display summarization quality metrics, while [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Prompt for PII detection 15 [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: Prompt for document pseudonymization. Prompt for Chain-of-Thought method "Does the text mention a person’s race?", "Are there any full or partial calendar dates mentioned (years, months, holidays, seasons)?" "Are there any specific personal names mentioned in the text?…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 8 canonical work pages

  1. [6]

    In 2023 IEEE Sym- posium on Security and Privacy (SP), pages 346–363, San Francisco, CA, USA

    Analyzing Leakage of Personally Identifiable Information in Language Models. In 2023 IEEE Sym- posium on Security and Privacy (SP), pages 346–363, San Francisco, CA, USA. IEEE. Laura Manor and Junyi Jessy Li. 2019. Plain English Summarization of Contracts. In Proceedings of the Natural Legal Language Processing Workshop 2019, pages 1–11, Minneapolis, Minn...

  2. [7]

    Computational Linguistics , 48(4):1053–1101

    The Text Anonymization Benchmark (TAB): A Dedicated Corpus and Evaluation Framework for Text Anonymization. Computational Linguistics , 48(4):1053–1101. Vasile P˘ais, , Dan Tufis, Elena Irimia, and Verginica Barbu Mititelu. 2024. Building a corpus for the anonymization of Romanian jurisprudence. In Pro- ceedings of the 18th Linguistic Annotation Workshop ...

  3. [9]

    Selective Differential Privacy for Language Modeling. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies, pages 2848–2859, Seattle, United States. Association for Computational Linguistics. Abhay Shukla, Paheli Bhattacharya, Soham Poddar, Ra- jdeep Mukherjee, ...

  4. [10]

    In AMIA Annual Symposium Proceedings , volume 2023, page 1324

    Large language models for healthcare data augmentation: An example on patient-trial matching. In AMIA Annual Symposium Proceedings , volume 2023, page 1324. Xiang Yue and Shuang Zhou. 2020. PHICON: Improv- ing Generalization of Clinical Text De-identification Models via Data Augmentation. In Proceedings of the 3rd Clinical Natural Language Processing Work...

  5. [11]

    In Proceedings of the 30th ACM SIGKDD Con- ference on Knowledge Discovery and Data Mining, pages 6655–6665, Barcelona Spain

    A Survey on Safe Multi-Modal Learning Sys- tems. In Proceedings of the 30th ACM SIGKDD Con- ference on Knowledge Discovery and Data Mining, pages 6655–6665, Barcelona Spain. ACM. Linwu Zhong, Ziyi Zhong, Zinian Zhao, Siyuan Wang, Kevin D. Ashley, and Matthias Grabmair. 2019. Au- tomatic Summarization of Legal Decisions using It- erative Masking of Predict...

  6. [12]

    Does the text mention a person’s race?

    with rank and α of 16, mixed-precision (FP16/BF16), and gradient checkpointing for a single epoch with a batch size of one. AdamW (Loshchilov and Hutter, 2019) is used with a weight decay of 0.01 and a learning rate of 5e-4 using a linear learning rate scheduler. See Appendix D for full implementation details. D Implementation Details We conduct our exper...

  7. [2012]

    font- variant:small-caps

    From Research to Health Policy Impact. Health Services Research, 47(1pt2):337–343. Trevor Cohn and Mirella Lapata. 2008. Sentence Com- pression Beyond Word Deletion. In Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 137–144, Manch- ester, UK. Coling 2008 Organizing Committee. Abhisek Dash, Anurag Shandil...

  8. [2020]

    In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871–7880, Online

    BART: Denoising Sequence-to-Sequence Pre- training for Natural Language Generation, Transla- tion, and Comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871–7880, Online. Association for Computational Linguistics. Songda Li, Yunqi Zhang, Chunyuan Deng, Yake Niu, and Hui Zhao. 2024. Better Lat...

Show all 12 references
  1. [2022]

    What Does it Mean for a Language Model to Preserve Privacy? In 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 2280–2292, Seoul Republic of Korea. ACM. Nicholas Carlini, Matthew Jagielski, Chiyuan Zhang, Nicolas Papernot, Andreas Terzis, and Florian Tr...

  2. [2023]

    In Find- ings of the Association for Computational Linguis- tics: ACL 2023, pages 2992–3005, Toronto, Canada

    Automated Refugee Case Analysis: A NLP Pipeline for Supporting Legal Practitioners. In Find- ings of the Association for Computational Linguis- tics: ACL 2023, pages 2992–3005, Toronto, Canada. Association for Computational Linguistics. Regina Barzilay and Kathleen R McKeown. ...

  3. [2024]

    In Proceedings of the 62nd Annual Meet- ing of the Association for Computational Linguis- tics (Volume 1: Long Papers) , pages 8742–8757, Bangkok, Thailand

    MetaSumPerceiver: Multimodal Multi- Document Evidence Summarization for Fact- Checking. In Proceedings of the 62nd Annual Meet- ing of the Association for Computational Linguis- tics (Volume 1: Long Papers) , pages 8742–8757, Bangkok, Thailand. Association for Computational Li...

  4. [2025]

    SPY: Enhancing Privacy with Synthetic PII Detection Dataset. In Proceedings of the 2025 Con- ference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 4: Student Research Workshop), pages 236–246, Albuq...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.