Source-linked AI summary
Overview of PAN 2026: Voight-Kampff Generative AI Detection, Text Watermarking, Multi-Author Writing Style Analysis, Generative Plagiarism Detection, and Reasoning Trajectory Detection
Janek Bevendorff, Maik Fröbe, André Greiner-Petter, Andreas Jakoby, Maximilian Mayerl, Preslav Nakov, Henry Plutz, Martin Potthast, Benno Stein, Minh Ngoc Ta, Yuxia Wang, Eva Zangerle
TL;DR
Reliable evaluation is needed as AI-assisted text production challenges detection and text forensics beyond surface characteristics. PAN 2026 presents five shared tasks spanning AI detection, watermarking, writing-style analysis, generative plagiarism, and reasoning trajectories, reflecting this evolving landscape.
Problem
LLM detection and textual-overlap methods are less relevant to modern plagiarism detection as the boundary of acceptable generated research content becomes less clear.
Source: §5 · p. 4Method
PAN 2026 organizes shared tasks on generative AI detection, text watermarking, multi-author style analysis, generative plagiarism, and reasoning trajectory detection.
Source: §1 · p. 2, §3 · p. 3, §4 · pp. 3–4Results
The task suite spans robust AI detection, watermarking, authorship-change analysis, generative plagiarism, and reasoning trajectory detection, reflecting the evolving landscape of AI-assisted text production.
Source: §7 · p. 6Takeaways & Limitations
Reasoning trajectory detection assesses how an LLM reasons as well as its conclusion, supporting work on trustworthy reasoning models.
Source: §6 · p. 6The 2025 plagiarism setup lacked realism because paragraph replacements were easy to detect, while substantial paraphrasing makes precise alignment less relevant.
Source: §5 · p. 4
Abstract
from arXiv · showhide
The goal of the PAN workshop is to advance computational stylometry and text forensics via objective and reproducible evaluation. In 2026, we run the following five tasks: (1) Voight-Kampff Generative AI Detection, particularly in mixed and obfuscated authorship scenarios, (2) Text Watermarking, a new task that aims to find new and benchmark the robustness of existing text watermarking schemes, (3) Multi-author Writing Style Analysis, a continued task that aims to find positions of authorship change, (4) Generative Plagiarism Detection, a continued task that targets source retrieval and text alignment between generated text and source documents, and (5) Reasoning Trajectory Detection, a new task that deals with source detection and safety detection of LLM-generated or human-written reasoning trajectories. As in previous years, PAN invites software submissions as easy-to-reproduce Docker containers for most of the tasks. Since PAN 2012, more than 1,100 submissions have been made this way via the TIRA experimentation platform.
1 Introduction
PAN is a workshop series and networking initiative for stylometry and digital text forensics. Its 2026 program builds on a history of shared tasks and datasets while continuing and introducing tasks across AI detection, writing analysis, plagiarism, watermarking, and reasoning trajectories.
1 Introduction
- Since its 2007 inception, PAN has organized 82 shared tasks and assembled many original evaluation datasets.Source: §1 · p. 2
- The 2026 program continues Voight-Kampff detection, multi-author writing style analysis, and generative plagiarism detection, and introduces text watermarking and reasoning trajectory detection.Source: §1 · p. 2
2 Task 1: Voight-Kampff Generative AI Detection
The task addresses the difficulty of detecting LLM-generated text across domains and under style obfuscation. Building on its 2025 robustness focus, the 2026 edition explores new ways to make and break successful detectors.
2 Task 1: Voight-Kampff Generative AI Detection
- LLM detection remains challenging across text domains and when style obfuscations mask telltale LLM fingerprints.Source: §2 · p. 2
- The 2025 task used a builder–breaker setup in which ELOQUENT participants created obfuscated datasets to challenge PAN participants’ systems.Source: §2 · p. 3
- For 2026, the task continues this research and explores new ways to make and break successful LLM detectors.Source: §2 · p. 3
3 Task 2: Text Watermarking
The new task tests text watermarking as a way to authenticate text despite unknown automated modifications. It evaluates both watermark inconspicuousness and robustness, using TIRA’s stateless sandbox to separate embedding from verification.
3 Task 2: Text Watermarking
- Participants embed a watermark into existing text, then verify it after unknown automated obfuscations intended to destroy the watermark.Source: §3 · p. 3
- Because the task uses existing text rather than relying specifically on LLMs, watermarking systems can authenticate text beyond machine-generated output.Source: §3 · p. 3
- TIRA executes submitted systems in a stateless sandbox between task steps to prevent knowledge leaks between embedding and verification.Source: §3 · p. 3
- Systems are evaluated for watermark inconspicuousness and robustness against unknown text modifications.Source: §3 · p. 3
4 Task 3: Multi-Author Writing Style Analysis
The task identifies positions where writing style, and thereby authorship, changes. The 2026 edition retains the established task definition while using fanfiction excerpts and controlled mixing to create datasets with different topical and stylistic similarities.
4 Task 3: Multi-Author Writing Style Analysis
- Multi-author style analysis identifies positions where writing style changes, supporting applications such as intrinsic plagiarism detection and authorship verification.Source: §4 · p. 3
- The task has evolved from author identification and grouping to paragraph-level style-change detection while controlling for simultaneous authorship and topic changes,,,,.Source: §4 · p. 3
- The 2026 task uses fanfiction excerpts and an established mixing method based on semantic and stylistic representations to create three datasets with different topical and stylistic similarities.Source: §4 · p. 4
5 Task 4: Generative Plagiarism Detection
The task redesigns generative plagiarism detection around source documents, addressing realism and alignment challenges in the previous setup. It includes source retrieval and text alignment, with cases of merging and expanding plagiarism.
5 Task 4: Generative Plagiarism Detection
- The task targets a plagiarism setting where blurred boundaries around acceptable generated content make detecting disingenuous research publications more difficult.The paper notes that methods based on LLM detection and textual overlap are less relevant for modern plagiarism detection.Source: §5 · p. 4
- The 2025 setup inserted plagiarized paragraphs into genuine publications, a scenario the authors describe as unrealistic and relatively easy to detect with embedding similarity.As textual overlap decreases, identifying precise plagiarized-paragraph boundaries becomes less meaningful.Source: §5 · p. 4
- The new dataset will generate plagiarism from one or more source documents, including merging multiple sources and expanding one source across multiple plagiarized paragraphs.This design is intended to reinstate source retrieval as a crucial part of plagiarism detection.Source: §5 · p. 4
- The dataset will add PubMed medical and chemical publications and JSTOR humanities and social-science publications, while largely retaining the prior subtask structure.The planned domain expansion is intended to enhance dataset relevance.Source: §5 · p. 5
- Source Retrieval asks systems to identify relevant documents in a candidate corpus or determine that an authentic suspicious document has no source there.Text Alignment matches suspicious-document paragraphs to source paragraphs, allowing one-to-many merging and many-to-one expanding correspondences.Source: §5 · p. 5
6 Task 5: Reasoning Trajectory Detection
The new task evaluates both the origin and safety of reasoning trajectories and final answers. It combines source detection with safety judgments at the reasoning-step and outcome levels.
6 Task 5: Reasoning Trajectory Detection
- The task addresses risks from spurious, non-logical, or unsafe intermediate reasoning that can accompany incorrect or harmful answers, as well as safe conclusions reached through deceptive or misaligned paths.It is intended to deepen understanding of AI-generated reasoning and support future improvements in reasoning and safety.Source: §6 · p. 5
- Source Detection classifies the source of a reasoning trajectory and final answer as AI-generated or human-written for queries mainly involving math, coding, and real-life financial reasoning.The task will curate human-written trajectories and collect outputs from open-source and commercial models, alongside existing math and coding datasets.Source: §6 · pp. 5–6
- Safety Detection judges both the reasoning trajectory and final answer as safe or unsafe across risky, jailbreak, and benign queries containing risky tokens.Training and validation will use annotated safety corpora, while the test set will add latest-model traces with self-correction, reflection loops, and tool use.Source: §6 · pp. 5–6
- The evaluation will run outside the TIRA sandbox, with participants submitting output files and no evaluation-time code execution, model inference, or hardware constraints.Participants may develop systems with their chosen local or cloud resources, models, and APIs.Source: §6 · p. 6
- The task evaluates how a model reasons as well as what it concludes, linking reasoning interpretability with safety assessment.It encourages detection of unsafe or synthetic trajectories, reasoning-transparency quantification, and explanations of detection decisions.Source: §6 · p. 6
7 Conclusion and Future Work
The conclusion presents five PAN 2026 shared tasks spanning AI detection, watermarking, authorship analysis, plagiarism detection, and reasoning trajectories. It frames them as responses to AI-assisted text production and sets out priorities for future evaluation work.
7 Conclusion and Future Work
- PAN 2026 covers five tasks: robust AI detection, text watermarking, multi-author style analysis, generative plagiarism detection, and reasoning trajectory detection.The authors describe the set as reflecting the evolving landscape of AI-assisted text production.Source: §7 · p. 6
- The authors call for rigorous evaluation frameworks that go beyond surface-level text characteristics as AI-assisted text production evolves.This is the stated rationale connecting the task portfolio to the broader evaluation need.Source: §7 · p. 6
- Future work will improve robustness, extend datasets across domains and languages, and refine evaluation to support trustworthy and reliable text analysis systems.Source: §7 · p. 6