acceptodds
Under review as a conference paper at ICLR 2027

DocAI: A Framework for Generating and Identifying AI-Edited Documents

Abstract

Document-level detection of AI-generated content is largely unaddressed, and off-the-shelf tools do not fill the gap: across nine AI-text and AI-image detectors, and both ways of composing them—score the whole page, or segment it and score each region—no configuration exceeds AUROC . The task demands dedicated tools and the data to build them; we provide both. DocAIgen edits real documents into large-scale corpora with region-level ground truth, from which we release a -page benchmark. DocAIdet is a document-native Vision-Language Model that reads the page as a whole, jointly detecting and segmenting AI-generated regions. It reaches page-level AUROC in-generation and – on two unseen generators ( at document level), with Dice and – respectively; the strongest piecewise baseline, retrained on our own data, reaches AUROC and Dice on the harder unseen generator, and the margins are significant with non-overlapping intervals. We release the engine, the benchmark and the detector.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Document Intelligence, Forensics, AI-generated, Fake, VLM, dataset, ethical, safety

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_w4ryOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces DocAIGen, a framework for generating AI-modified documents to evaluate detection and localization methods. The authors further propose DocAIdet, a method designed to detect and localize AI-modified content in documents.

Critical strengths
  • The proposed dataset is potentially valuable for future research on AI-modified document detection and localization.
Critical weaknesses
  • The quality and representativeness of the proposed dataset raise several concerns. First, the dataset is constructed using relatively small and outdated open-source generative models, such as Llama-3.2-3B and Stable Diffusion v1.5. This raises questions about whether the generated documents adequately reflect the capabilities of more advanced generative models, making it questionable whether the reported findings still holds on more advanced models. Second, the paper lacks sufficient qualitative examples to demonstrate the quality of the generated documents. The examples provided in the Appendix are too small to be clearly examined. Furthermore, no human evaluation is conducted to assess the realism and quality of the generated documents. Consequently, claims such as "no shortcuts" are not sufficiently supported. Although the original and modified documents are rendered using the same engine, this does not necessarily eliminate potential formatting shortcuts. For example, AI-modified text may be excessively long, resulting in unnatural page layouts that could be exploited by detectors.

  • The experimental evaluation is not sufficiently comprehensive, and the generalization ability of the proposed method remains unclear. In particular, a piecewise baseline that separately applies text and image detectors should be one of the most important comparisons. However, the paper does not clearly specify which text and image detectors are used in this baseline, making it difficult to assess the fairness and strength of the comparison. Moreover, the evaluated text and image detectors appear to be relatively simple fine-tuned models, while stronger and more representative methods, such as DRCT, AIDE, and DDA for image detection, are not included. This limits the persuasiveness of the experimental conclusions. Additionally, Gemini-3-Flash (a model released near a year ago) already achieves strong performance, suggesting that the advantages of the proposed method over capable general-purpose models require further investigation. Another major concern is the generalization ability of DocAIdet. The proposed method is trained on a subset of the constructed benchmark, which raises the possibility of learning dataset-specific shortcuts. Even if the generators used for training and testing differ, the underlying data construction pipeline remains the same, potentially introducing shared artifacts or formatting patterns that the detector could exploit. The design of DocAIdet further motivates this concern. Given the relatively limited capabilities of its SmolVLM backbone, its strong performance in detecting AI-modified text rendered within document images is somewhat surprising and warrants further investigation. While this observation does not necessarily indicate overfitting, it raises questions about whether the model has learned generalizable detection features or relies on dataset-specific artifacts.

Questions for authors
  • Could the authors provide additional qualitative and quantitative evaluations to demonstrate the quality, realism, and reliability of the proposed benchmark? In particular, a human evaluation assessing document realism and potential formatting shortcuts would be valuable.

  • Could the authors incorporate stronger and more representative AI-generated text and image detectors, particularly in the piecewise baseline, to provide a more comprehensive and fair comparison?

  • Could the authors provide stronger evidence of the generalization ability of DocAIdet? For example, evaluating the method on independently collected real-world documents with disclosed AI usage, rather than documents generated through the same construction pipeline, would help assess whether the method generalizes beyond dataset-specific artifacts.

Reviewer_D3emOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces document-level detection and localization of AI-edited content. DocAIgen edits arXiv PDFs to build a 42,503-page benchmark with region-level labels, and DocAIdet, a 265M SmolDocling-based VLM, reaches 0.929 AUROC and 0.699 Dice on unseen generators, beating retrained piecewise baselines.

Critical strengths

S1: This paper defines a new task and releases a regrowable data engine. Pristine and edited pages share one renderer, and the vision-only stream cannot fit the training data, ruling out rendering shortcuts.

S2: This paper evaluates nine off-the-shelf detectors, retrained piecewise pipelines and three frontier MLLMs under one protocol with paired tests.

Critical weaknesses

W1: This paper trains and tests only on DocAIgen output, and even pristine pages are re-rendered by Docling. No original arXiv PDF or externally edited document is tested, so the evaluation largely validates its own pipeline. This paper's piecewise baselines do not use the best detector in each field.

W2: This paper does not show that document detection differs from context-aware text detection: the semantic-only stream already reaches about 0.5 Dice, and the detector combines standard parts on SmolDocling, with a generic backbone still reaching 0.905 AUROC.

Questions for authors

Q1: What text does the language decoder read: the PDF text layer, SmolDocling's own transcription, or teacher-forced tokens? Does the method work on scanned documents that have no text layer?

Q2: Can the authors test on original arXiv PDFs that were not re-rendered, and on documents edited outside DocAIgen, for example rewritten by frontier LLMs and then polished by humans?

Q3: Can the authors add a human study rating whether the generated paragraphs, tables and figures are plausible, and report how often figure edits used the random-image fallback?

Q4: Can the authors give a text-only detector each block together with its neighboring blocks? This would show how much layout and component-role cues add beyond plain context.

Reviewer_FjUgOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies a task of AI-generated documents detection and localisation. Its contributions are the following: (i) DocAIgen: a pipeline for building a collection of matched real and AI-generated documents with region-level annotations, and (ii) DocAIdet: an AI model that combines vision and language processing parts for detecting and localising AI-edited content in documents.

The paper argues that existing image-based and text-based methods for detecting AI-generated content underperform when applied to the task of AI-generated documents detection, using this to justify building a new benchmark and a framework for AI-generated documents detection. Results suggest that the proposed DocAIdet achieves higher ROC AUC and DICE scores on the proposed benchmark compared to several considered baseline approaches.

Critical strengths
  • The task of detecting AI-generated content in documents is meaningful and timely, as is the idea of developing a dedicated benchmark and detectors specifically adapted to this task
  • The proposed approach, DocAIdet, demonstrates the potential of combining image-based and text-based methods for detecting AI-generated content in documents
Critical weaknesses
  • The proposed benchmark lacks sufficient statistical coverage and characterization. When a new benchmark is presented, it should be accompanied by rigorous statistical analysis, such as distribution of the proportion of AI-generated content per document/page, distributions of numbers and lengths of modified text blocks, distributions of document types, scientific topics and lengths, distributions of the amount of content of each type in the documents, etc. Such analysis is important for identifying potential biases in the benchmark and ensuring that the performance metrics of methods evaluated on it are interpreted in light of these biases. The paper provides some component-level statistics and ablation studies, but these are far from sufficient.
  • The set of considered baselines is incomplete. I am more familiar with the field of AI-generated text detection, so I cannot reliably assess the selection of image-based baselines. However, many important text-based detection methods are missing, including Binoculars [1], Fast-DetectGPT [2], topology- and dimensionality-based detectors [3, 4, 5], likelihood- and entropy-based detectors, RepreGuard [6], and ImBD[7]. When proposing a new benchmark and an accompanying novel method, it is crucial to provide a comprehensive evaluation against established approaches.
  • The binary labelling scheme is questionable. As far as I understand, in DocAIgen, a page is labeled as AI-generated if at least one of its components has been replaced with an AI-generated surrogate. This design equates cases that can be processed very differently by different AI-generated content detection methods: identifying a document as AI-generated where only one component has been replaced is a different task from identifying a document where half of the content is replaced. The paper reports recall stratified by edited area but still lacks a systematic evaluation of the full classification task under controlled modification ratios.
  • The generative models used to replace content in DocAIgen are relatively small and outdated. The performance of AI-generated content detection methods depends heavily on the models used to generate the training and test data. Thus, performance on the current benchmark cannot be readily generalized to settings involving more modern generative models.
  • The proposed DocAIdet architecture is overly complex, with important component-level ablations missing, e.g. gated vs ungated cross-attention, single-scale vs multi-scale features, cross-attention vs simple fusion. Without these ablations, it is difficult to assess which components are necessary and which contribute to the framework's robustness.
  • Some experiments are confounded. For example, in the cross-generation evaluation, both the text generator and the source documents are changed, which entangles the effects of generalization to unseen models and distribution shift. Page-level bootstrap resampling is also questionable because multiple pages originate from the same documents

[1] Hans et al., Spotting LLMs With Binoculars: Zero-Shot Detection of Machine-Generated Text

[2] Bao et al., Fast-DetectGPT: Efficient Zero-Shot Detection of Machine-Generated Text via Conditional Probability Curvature

[3] Tulchinskii et al., Intrinsic Dimension Estimation for Robust Detection of AI-Generated Texts

[4] Kushnareva et al., Artificial Text Detection via Examining the Topology of Attention Maps

[5] Cherniavskii et al., Acceptability Judgements via Examining the Topology of Attention Maps

[6] Chen et al., RepreGuard: Detecting LLM-Generated Text by Revealing Hidden Representation Patterns

[7] Chen et al., Imitate Before Detect: Aligning Machine Stylistic Preference for Machine-Revised Text Detection

Questions for authors
  • Can you provide sufficient statistical coverage and characterization for the benchmark?
  • Can you add established baselines to the comparison?
  • Can you provide ablations on all the main components of DocAIdet?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.