Traditional authorship attribution (AA) is often framed as a multiple-choice test: given a document and a list of ten suspects, which one wrote it? While this "closed-set" approach works in controlled experiments, it fails in the "wild-west" conditions of real-world forensics, historical analysis, and platform integrity. In these scenarios, the true author is frequently not in the database at all.
Researchers are shifting the paradigm from simple classification toward Open-World Authorship Attribution and Authorship Verification. This transition moves away from picking the "best fit" among known candidates and toward understanding the fundamental stylistic distance between any two pieces of text.
The Closed-Set Fallacy
In a closed-set environment, a model is forced to choose the most likely candidate from its training data. As Gordon Lucas and Janek Amann observe, if the true author is missing, these systems will incorrectly attribute the text to the most similar known candidate rather than identifying it as "unknown."
This limitation is a critical failure point in forensic linguistics. Thamar Solorio and Manuel Montes-y-Gómez argue that real-world cases require models that can handle an unknown individual not represented in the training data. Whether investigating criminal activity (Fredrik Johansson) or human trafficking (Vageesh Saxena), the system must be able to say "none of the above."
From Classification to Verification
To solve the open-world problem, researchers often reframe attribution as Authorship Verification. Instead of asking "Who wrote this?", the system asks "Did Author X write this?"
Cor J. Veenman describes this as a one-class problem. In this framing, the "target" class is a specific author, while the "non-target" class is effectively the rest of the human population. This is significantly more difficult because the non-target category is infinite and impossible to sample fully. Conrad Sanderson and Simon Günter highlight that this verification approach is the only way to determine if a text originated from an unknown external source.
Learning the Space of Style
If we cannot train a classifier on every person on Earth, how do we identify an unknown author? The answer lies in Representation Learning. Instead of learning to recognize specific authors, models learn to map text into a "similarity space" where documents by the same author cluster together, regardless of whether the model has seen that author before.
- Siamese Networks: Chakaveh Saedi and Mark Dras use Siamese architectures to learn a general "notion of authorship." By comparing pairs of texts, the model learns stylistic fingerprints that generalize to new, unseen authors.
- Learning in Similarity Space (LSS): Tieyun Qian and Bing Liu use this approach to link "sockpuppet" accounts. By focusing on the difference between writing styles rather than the styles themselves, the model can accurately identify when two IDs belong to the same person, even in an open-world setting.
- Universal Embeddings: Jitkapat Sawatphol and Sarana Nutanong emphasize creating stylistic embeddings that are regularized against topic bias. This ensures the model learns how a person writes (style) rather than what they write about (content), which is essential for identifying authors across different domains.
Scaling to the Unbounded World
The ultimate challenge is the Open-World Authorship Attribution task, where the candidate pool is effectively the entire internet. Xinhao Tan and Xinchao Wang recently proposed a two-stage framework for this: first, using hierarchical keyword searching to narrow the "infinite" world down to a manageable candidate pool, and then using Large Language Models (LLMs) to perform multi-perspective reasoning on the remaining suspects.
This shift is also revitalizing historical research. By moving away from closed-set models, researchers can more accurately analyze disputed classical texts, such as the Ars Rhetorica (Fine-Tuning Pre-Trained Language Models...) or works attributed to Plutarch (Ivan P. Yamshchikov). These methods allow scholars to provide evidence that a text was written by an "anonymous other" rather than being forced to pick from a list of famous historical figures.