Authorship Attribution in Multilingual Machine-Generated Texts¶

Conference: ACL 2026
arXiv: 2508.01656
Code: To be confirmed
Area: AIGC Detection
Keywords: Machine-generated text detection, Authorship attribution, Multilingual, Cross-lingual transfer, MULTITuDE

TL;DR¶

Existing research on machine-generated text authorship attribution (identifying which specific LLM or human produced a text) is almost entirely monolingual (primarily English). This paper is the first to formally define Multilingual Authorship Attribution (ML-MGT) and Cross-Lingual Authorship Attribution (CL-MGT). Through a systematic evaluation of 18 languages \(\times\) 8 generators (7 LLMs + human) using statistical methods, fine-tuned encoders, contrastive learning, and fine-tuned decoders, it finds that while fine-tuned/contrastive methods adapt well to multiple languages (best macro-F1 > 0.9), they degrade severely when transferring across different language families or writing systems, revealing the challenges of real-world multilingual scenarios.

Background & Motivation¶

Background: As LLM fluency approaches human levels, machine-generated text (MGT) becomes increasingly difficult to distinguish. Initial responses focused on binary MGT detection (determining if a text is machine-written). However, as the number of LLMs grows daily, knowing it is "machine-written" is insufficient; identifying the specific source model—fine-grained authorship attribution (AA)—is crucial for accountability, provenance, and preventing abuse.

Limitations of Prior Work: Research in AA remains largely monolingual. Most work focuses on English, with a few extensions to Russian or Spanish, lacking systematic study of multilingual attribution. Modern LLMs are inherently multilingual and used across various cultural contexts; whether AA methods verified only in English can function or generalize in real-world multilingual scenarios remains an unexplored area.

Key Challenge: Attribution is significantly more difficult than detection (multi-class classification across 8 balanced categories, where the random baseline is only 0.125 macro-F1). Multilingualism adds another layer of complexity—linguistic properties vary greatly across different language families and writing systems (Latin, Cyrillic, Arabic, Hanzi, Greek). A "generator fingerprint" learned in one language might not transfer to another.

Goal: Dissect this blind spot using three research questions:
RQ1: How do existing AA methods perform on Multilingual MGT (ML-MGT)?
RQ2: To what extent can AA methods transfer across languages and language families (CL-MGT)?
RQ3: How does the choice of generator model affect multilingual adaptability and cross-lingual generalization?

Core Idea: Instead of proposing a single new model, this work formally defines the ML-MGT/CL-MGT problems for the first time + builds a unified and comparable multilingual evaluation + systematically adapts and tests representative existing methods, providing the first systematic evidence of the difficulties and promising directions for this new problem.

Method¶

Overall Architecture¶

This is a problem definition + systematic empirical study paper rather than a single new method. The logical chain is: ① Formalize authorship attribution as a multi-class classification problem (ML-MGT) and define cross-lingual transfer as a special case (CL-MGT); ② Select a balanced evaluation set of 18 languages \(\times\) 8 generators based on the MULTITuDE dataset; ③ Adapt four representative classes of existing methods (statistical, fine-tuned encoders, contrastive learning, fine-tuned decoders) to this attribution task; ④ Design four evaluation tasks to answer RQ1 (multilingual adaptation), RQ2 (transfer across languages/families), and RQ3 (generator impact), using macro-averaged F1 as the metric.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["ML-MGT / CL-MGT<br/>Problem Formalization"] --> B["MULTITuDE Data<br/>Balanced set of 18 languages × 8 generators"]
    B --> C["Method Adaptation<br/>Statistical / Encoder / Contrastive / Decoder"]
    C --> D["Four-Task Evaluation<br/>macro-F1"]
    D -->|Joint training on all languages| E["RQ1 Multilingual Adaptation"]
    D -->|Train on one/few languages → Test on all| F["RQ2 Cross-lingual & Cross-family Transfer"]
    D -->|Split by generator| G["RQ3 Generator Influence"]

Key Designs¶

1. Problem Formalization of ML-MGT and CL-MGT: Splitting multilingual attribution into "Joint" and "Transfer" layers

The paper first defines authorship attribution clearly. Given a text set \(\mathcal{X}=\mathcal{X}_h\cup\mathcal{X}_m\) (human-written + machine text from a set of generators \(\mathcal{M}\)), each text belongs to a language in the set \(\mathcal{L}\). The goal of ML-MGT (Problem 1) is to learn a mapping \(f:\widehat{\mathcal{X}}\mapsto\mathcal{Y}=\{y_h\}\cup\mathcal{Y}_m\), identifying the author among 8 classes (human class \(y_h\) + \(|\mathcal{M}|\) machine generators). Language selection is controlled by strategy \(g(\cdot)\), assuming human and MGT texts are available in the same language pairs. CL-MGT (Problem 2) is a special case: when the training language set \(\mathcal{L}_{train}\subset\mathcal{L}_{test}\), the model must rely on cross-lingual knowledge transfer rather than memorization. This two-layer split is the framework for all subsequent experimental designs.

2. Balanced Multilingual Evaluation Set: Filtering 18 languages \(\times\) 8 generators from MULTITuDE to control comparability

To ensure cross-lingual comparisons are not "contaminated by data bias," the paper uses the MULTITuDE (v3) dataset. It includes articles generated by 7 LLMs (Mistral-7B-Instruct, OPT-IML-Max-30B, v5-Eagle-7B, Vicuna-13B, Llama-2-70B-Chat, Aya-101, GPT-3.5-Turbo) using the same news headline prompts, plus human news from MassiveSum. Each language uses the same set of generators, settings, and domains, designed specifically for unbiased cross-lingual comparison. Out of 21 available languages, 18 were selected based on: (i) complete language-generator balance, and (ii) reaching at least 95% of the target sample size (approx. 1000 per generator for training, 300 for testing). The final set covers 8 language families and 5 writing systems (12 Latin / 3 Cyrillic / 1 Arabic / 1 Hanzi / 1 Greek), with uniform class distribution.

3. Multilingual Adaptation Suite: Unified adaptation of four technical routes to the attribution task

The paper adapts four categories of representative methods to AA. Statistical Methods: Zero-shot features are extracted using Fast-DetectGPT (mGPT-13B as reference/sampling model) and Binoculars (Falcon-7B as observer), followed by a Logistic Regression classifier; the stronger StatEnsemble feeds nine statistical features (Binoculars, Fast-DetectGPT, perplexity, Rank, log-rank, log-likelihood, Entropy, LLM-Deviation, DetectLLM-LRR) into an MLP. Fine-tuned Encoders: RoBERTa-large (monolingual English) and XLM-RoBERTa-large (multilingual), fine-tuned following existing work (lr 2e-6, max length 512). Contrastive Learning: The OTBDetector (top performer on OpenTuringBench) is adapted by replacing Longformer with XLM-RoBERTa-large to ensure multilingualism, using contrastive loss to separate latent representations of generators. Fine-tuned Decoders: Both mdok (based on Qwen3-4B-Base + QLoRA) and Qwen3-4B-Base are adapted with a multi-class classification head and fine-tuned. This covers the full spectrum from zero-shot statistical to discriminative fine-tuning and contrastive separation.

An Illustrative Example: How RQ2 Cross-family Transfer Exposes Vulnerability¶

Taking "Writing System Transfer" as an example: the paper trains on English and Spanish (representing Latin script) or Russian (Cyrillic script) and tests across all 18 languages. Results (see RQ2 data) show that while transfer within the same family/script is acceptable, performance drops significantly when a model trained on Latin is tested on Cyrillic, Arabic, or Hanzi. This concretizes the vulnerability behind the ">0.9 macro-F1" joint training figures: models largely learn "generator fingerprints within specific languages." Changing the script or family causes these fingerprints to mismatch. This is a core warning: the multilingual capability of current AA methods is overestimated by "joint training" results.

Key Experimental Results¶

Main Results (RQ1: Multilingual Adaptation)¶

Joint training on all 18 languages, 8 classes, reporting macro-averaged F1 (random baseline 0.125). 5 out of 8 detectors reached macro-F1 \(\ge\) 0.75; fine-tuning and contrastive methods adapted best.

Method	Type	Average F1 (All Languages)	Note
Qwen3-4B-Base	Fine-tuned Decoder	0.93	Top performer
mdok	Fine-tuned Decoder (QLoRA)	0.93	>0.9 in most languages
OTBDetector	Contrastive Learning	0.90	7× fewer params, only ~3% drop
XLM-R-large	Multilingual Encoder	0.84
RoBERTa-large	English Encoder	0.75	Surprisingly difficult in English
StatEnsemble	Statistical Ensemble	0.45	Statistical methods are weak
Fast-DetectGPT	Single Statistical	0.23
Binoculars	Single Statistical	0.16	Slightly above random

Cross-lingual/Cross-family Transfer (RQ2) and Generator Influence (RQ3)¶

Evaluation Dimension	Setting	Key Observation
Multilingual Adaptation (RQ1)	Joint training on all languages	Fine-tuning/Contrastive F1 >0.9, Statistical \(\le\) 0.45
Per-language Transfer (RQ2)	Train on en/es/ru (single/combined) → Test on all	Performance fluctuates heavily based on language similarity
Per-family/Script Transfer (RQ2)	Latin (en+es) vs. Cyrillic (ru) training → Test on all	Most significant degradation across different writing systems
Generator Influence (RQ3)	Class-level F1 by generator	Generator identity interacts with language context

Key Findings¶

Fine-tuning/Contrastive > Statistical; Decoder/Contrastive > 0.9 in most languages: Qwen3-4Base, mdok, and OTBDetector lead significantly. Pure zero-shot statistical methods (Fast-DetectGPT 0.23, Binoculars 0.16) are nearly unusable for attribution—statistical signals for "detection" are insufficient for "attribution."
OTBDetector offers high efficiency: Despite being 7× smaller than Qwen3-4B-Base/mdok, its F1 only drops by 3%, attributed to sharper decision boundaries from contrastive loss.
English text is surprisingly difficult to attribute: Even with the monolingual English pre-trained RoBERTa, F1 on English remains lower (around 0.72/0.65), suggesting generator fingerprints in high-resource languages may not be easier to distinguish.
Cross-family/Cross-script is the true bottleneck: Methods transfer reasonably well within similar language families, but performance degrades significantly across Latin \(\leftrightarrow\) Cyrillic \(\leftrightarrow\) Arabic \(\leftrightarrow\) Hanzi, influenced by both target language properties and generator identity.

Highlights & Insights¶

Problem Formalization as a Contribution: Formalizing "Multilingual Authorship Attribution" and "Cross-lingual Transfer" as ML-MGT/CL-MGT provides a reusable evaluation framework for future work.
"High Joint Training Scores" can be Misleading: The >0.9 F1 in RQ1 is impressive, but RQ2 reveals severe flaws in cross-family transfer. This contrast reminds the community not to be misled by a single joint multilingual number; transferability is the true test of real-world scenarios.
Transferable Experimental Design: Using "same generator, same domain, same setting, varying only language" with balanced data isolates the effects of language/family, a control-variable approach applicable to any "cross-lingual robustness" evaluation.

Limitations & Future Work¶

Only used one dataset (MULTITuDE) and focused on the news domain—whether findings hold across social media, code, or dialogue remains unverified.
The generator set consists of 7 relatively early LLMs (GPT-3.5, Llama-2, Vicuna, etc.); newer models (GPT-4 class, Claude, newest open-source models) are not covered, and fingerprint separability might change with model iterations.
The paper stops at "revealing problems + evaluating existing methods" and does not propose a new method for cross-family generalization; the solution for CL-MGT remains an open problem.
Evaluation focuses on macro-F1, with limited discussion on calibration or adversarial robustness (e.g., paraphrasing/obfuscation) critical for deployment.

vs. Binary MGT Detection (DetectGPT/Binoculars, etc.): Detection only determines "Human vs. Machine"; this work performs fine-grained attribution. Experiments prove statistical signals effective for detection fail for attribution.
vs. Monolingual Attribution (OTBDetector / OpenTuringBench): These methods are SOTA in monolingual English. This work adapts them to 18 languages, exposing cross-lingual transfer gaps through multilingual stress testing.
vs. Existing Multilingual MGT Datasets (M4GT-Bench / RAID): MULTITuDE was chosen because it maintains consistent generators/settings/domains across all languages, enabling unbiased cross-lingual comparison rather than just increasing language count.

Rating¶

Novelty: ⭐⭐⭐⭐ First systematic study of multilingual/cross-lingual authorship attribution.
Experimental Thoroughness: ⭐⭐⭐⭐ Wide coverage (18 langs \(\times\) 8 generators \(\times\) 4 method classes), but limited to a single dataset/domain.
Writing Quality: ⭐⭐⭐⭐ Clear problem definitions and rigorous control-variable design.
Value: ⭐⭐⭐⭐ Established evaluation standards for provenance/accountability and identified the cross-family transfer challenge.