Overview of PAN 2026: Voight-Kampff Generative AI Detection, Text Watermarking, Multi-Author Writing Style Analysis, Generative Plagiarism Detection, and Reasoning Trajectory Detection

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.

💡 Research Summary

The paper presents an overview of the PAN 2026 shared‑task workshop, detailing five distinct tasks that together address the most pressing challenges in computational stylometry and text forensics today. The first task, Voight‑Kampff Generative AI Detection, builds on the successful 2024 and 2025 editions and focuses on detecting LLM‑generated text even when it is mixed with human‑written passages or deliberately obfuscated through synonym substitution, sentence reordering, or other adversarial transformations. Participants must submit Docker containers that are evaluated on the TIRA platform for robustness against unknown modifications, thereby pushing detection methods beyond simple fingerprinting toward truly resilient AI‑authorship attribution.

The second task, Text Watermarking, introduces a two‑stage pipeline: participants embed an invisible watermark into an existing piece of text, then a separate system attempts to recover it after the text has been subjected to a series of unknown automated obfuscations. This task expands the scope of watermark research from LLM‑output‑only scenarios to any textual artifact, emphasizing inconspicuousness and durability as primary evaluation dimensions.

The third task, Multi‑Author Writing Style Analysis, revisits a long‑standing PAN challenge with a fresh dataset derived from fan‑fiction. Fan‑fiction provides long, topically coherent excerpts while allowing controlled mixing of multiple authors, enabling the creation of three difficulty levels based on stylistic and topical similarity. Participants must develop profiling methods that pinpoint exact positions where the writing style—and thus authorship—changes, a capability crucial for downstream plagiarism detection and authorship verification.

The fourth task, Generative Plagiarism Detection, addresses the emerging reality that LLMs can produce sophisticated, near‑verbatim or heavily paraphrased plagiarism. Unlike earlier PAN editions that inserted isolated plagiarized paragraphs, the new dataset simulates realistic plagiarism by merging content from multiple sources, expanding single sources into several plagiarized paragraphs, and covering diverse domains such as biomedical literature (PubMed) and humanities/social sciences (JSTOR). The task is split into source‑retrieval and text‑alignment subtasks, demanding both effective document‑level search and fine‑grained alignment capable of handling merging and expanding plagiarism patterns.

The fifth and most novel task, Reasoning Trajectory Detection, focuses on the internal reasoning steps of LLMs. Subtask 1 (Source Detection) requires systems to decide whether a given reasoning trace and its final answer originate from a human or an AI model, using queries from mathematics, coding, and real‑world finance. Subtask 2 (Safety Detection) asks participants to label both the reasoning trajectory and the final answer as safe or unsafe across three query categories: risky content requests, jailbreak attempts, and benign queries containing risky tokens. This task moves beyond answer‑centric evaluation to assess why an LLM may produce unsafe or misleading conclusions, encouraging the development of interpretable and safety‑aware reasoning detectors.

All tasks continue PAN’s tradition of reproducible evaluation via Docker containers on the TIRA platform, with the exception of Reasoning Trajectory Detection, which uses a separate submission format to allow unrestricted hardware and model choices. The paper emphasizes that, since its inception in 2007, PAN has organized 82 shared tasks and received over 1,100 reproducible submissions, establishing a robust infrastructure for benchmarking. By combining established challenges (robust AI detection, multi‑author style change) with new frontiers (watermark robustness, generative plagiarism, reasoning safety), PAN 2026 aims to drive research toward systems that can reliably detect, attribute, and assess the safety of AI‑generated text across domains and languages, ultimately contributing to a more trustworthy information ecosystem.