The “Voight-Kampf” Generative AI Authorship Verification task aims to determine whether a text was generated by an AI or written by a human. The 2025 edition of the task explores two subtasks: Subtask 1 tests the detection of purely AI generated text with potentially unknown obfuscations, and as such continues our research from 2024. The task is again organized as a builder-breaker challenge together with the ELOQUENT lab. The PAN participants submitted 24 detectors. The best system archives a mean score of 0.99, the best baseline achieves a score of 0.92. ELOQUENT participants submitted 13 new test datasets with 22 obfuscated texts each. The most dificult dataset archives a mean C@1 − 1 score of 0.63. Subtask 2 investigates texts with six degrees of human-AI collaboration: (i) fully human-written, (ii) humanwritten, then machine-polished, (iii) machine-written, then machine-humanized (obfuscated), (iv) human-initiated, then machine-continued, (v) deeply mixed text,where some parts are written by a human and some are generated by a machine, and (vi) machine-written, then human-edited. The dataset contains over half a million examples in total and is composed from several relevant AI-detection datasets across multiple text genres. PAN participants submitted 21 detectors to subtask 2. The best system archives an F1 score of 0.65, the best baseline a score of 0.48. The data, baselines, and the code used for creating the datasets and evaluating the systems are available.1
Authorship verification is a fundamental task in author identification. PAN has continuously been organizing authorship verification tasks for years [ 1, 2, 3, 4] and with generative AI / LLM detection being fundamentally also an authorship verification task [5], we decided to “delve” into that realm.
For the 2025 edition of this task, we increase the dificulty of the binary detection setting and add a new subtask with a focus on detecting human-AI collaboration. In the first edition of this task [ 6], “Voight-Kampf” was organized as a classic authorship verification task: Given two texts, one authored by a human, one by a machine: pick out the human. Against our initial assumptions, the task was of little challenge to the participants’ systems even in the face of text obfuscation. Thus, for this year, we elevate the challenge in two directions.
First, we reduce the original “Voight-Kampf” setting to a binary task: Given a text, determine if it was written by a human or an AI model. In addition, we extend our selection of obfuscation methods to test model sensitivity and add larger, newer, and more dificult-to-detect models. As before, the PAN and ELOQUENT labs jointly organize this task: ELOQUENT participants provide very strongly obfuscated texts from a set of summaries, and PAN participants develop systems to robustly detect AI generated text. The ELOQUENT contributions are described in Section 2 and the PAN contributions are described in Section 3. As in the previous year, all systems are submitted as immutable docker containers via TIRA [7] for easy reproducibility.
Second, we organize a new subtask aiming to detect human-AI collaboration, since joint writing and editing of documents with AI tools has become commonplace and such cases are presumably harder to detect than in the binary case. The subtask is a logical continuation of previous shared tasks organized at SemEval and GenAIDetect [8, 9]. Subtask 2 asks to distinguish six cases of joint human-AI authorship: (i) fully human-written, (ii) human-written, then machine-polished, (iii) machine-written, then machine-humanized (obfuscated), (iv) human-initiated, then machine-continued, (v) deeply mixed text, in which some parts are written by a human and some are generated by a machine, and (vi) machine-written, then human-edited. This subtask is described in Section 4.
ELOQUENT participants generate datasets of machine text, attempting to break the classifier systems built by PAN participants.
In its first edition, the classifiers submitted by participants to the PAN lab handily classified the texts into human vs. machine. We found that of the submitted datasets in 2024, all were able to fool some of the classifier systems some of the time; but no generative model was consistently able to convince the better classifier systems that it was human. It was clear that machine-generated texts appeared to consistently hold to certain detectable stylistic indicator features. [6]
For the test set, 22 texts written by human authors, of between 350–700 words were selected. Most original texts were longer and a suitably long section of the text was selected. Summaries of each text were generated by the organizers using OpenAI’s ChatGPT service using the prompt “Summarize the main points of the following text and give an overall description of the genre and tone of the text.” Those summaries were then shared to the participants for their systems to generate short texts on the basis of the summaries. A sample summary test item is given in 1 and a list of item titles is given in 1.
A suggested prompt was given with the participants—“Write a text of about 500 words which covers the following items”—but the participants were free to formulate their own prompts as they saw fit. The generated texts were submitted by the participants through a submission form, and then further submitted by the organizers to the PAN lab for classification.
Genre: Mythological and Linguistic Ethnography / Cultural Anthropology Tone: Scholarly, reverent, and lyrical, blending academic analysis with a poetic appreciation of language, mythology, and cultural worldview. Finnish language and culture are deeply intertwined with nature, with precise and acoustically rich verbs used to describe natural elements like snow, wind, and animals.
Ancient Finns practiced animistic nature-worship, viewing all visible forces sun, moon, sea, earth as living, conscious beings.
Over time, belief evolved to include invisible spiritual beings, or habitat (genii/regents), who governed natural elements and had both form and spirit, though lesser ones were more formless and abstract. These haltiat were immortal and hierarchical, often ranked based on the significance of their domain (e.g., Tapio of the forest outranking Pilajatar, daughter of the aspen).
Finnish mythology emphasizes the independence and dignity of each deity, regardless of power; even a minor god rules absolutely within their sphere.
Deities were typically paired and familial, with the sky and celestial bodies being the earliest and most revered objects of worship, leading to the concept of Jumala, the thunder-home, as the supreme god.
GPT-2 Qwen3-8B Qwen3-8B Qwen3-8B Qwen3-8B Claude Sonnet3.5 Claude Sonnet3.5 0.136 Of the 49 registered participants, five teams submitted 10 experiments [ 10, 11, 12, 13, 14]. The attrition rate is great, similar to last year, and we will investigate the possibility of turning this task into a continuously open experiment with asynchronous submission.
Table 2 lists the participating systems and the classification results per system using the C@1 accuracy score aggregated from the PAN lab participants.
This year, we find that of the submitted ELOQUENT-generated datasets, all were able to fool some of the classifier systems some of the time, and several were able to do so better than chance. This is a considerable improvement from last year, and reflects a more directed experimentation in the submitted experiments, e.g., by inserting explicit human variation and departure from norm in the generated texts.
The human authored texts were on average misclassified as machine generated only 15% of the time and this is entirely due to two texts which both caused a majority of the classification systems to assign them a non-zero likelihood to be machine-generated: the excerpt from the Maastricht treaty (038) and the Intro to Large Language Models (045). All other human-authored texts were correctly assumed to be human-authored by every participating classifier system. The best generative models fooled the classifiers with many, but never with all of their generated texts. Some of the test items appear to have been easier to generate convincing human-like output for: Session Moderator Instructions (036), What is Free Software? (042), and the Wikipedia text on Safron (052). It is unclear if this is an efect of the qualities of the summaries or if the language model training data are better equipped for academic text than for other genres.
As a very general finding, it is clear that generative language models still have recognizable traits in their output and that classifiers are able to pick up on them quite efectively. Generating truly human-like text remains an open challenge for developers and operators of generative language models.
At PAN 2024 [15], we ofered, for the first time, the “Voight-Kampf” Generative AI Authorship Verification task [6], which attracted a large number of submissions. To start with, we formalized diferent task variants and ordered them from easiest to hardest (Figure 2). To establish a baseline, we decided to start with the easiest variant, in which participants were given a pair of texts, of which exactly one was of human and the other of machine origin.
Input / Task 1. { ? , ? } 2. { ? , ? } 3. { ? , ? } 4. { ? , ? } 5. { ? , ? } 6. { ? , ? } 7.
? →−
Possible Assignment Patterns 1. { A , M } 2. { A , M }, { A , A } 3. { A , M }, { M , M } 4. { A , M }, { A , A }, { M , M } 5. { A , M }, { A , A }, { A , B } 6. { A , M }, { A , A }, { A , B }, { M , M } 7.
A ,
M
For PAN 2025 [16], we move on to the harder variant, in which participants are given only one text. This variant reflects a more realistic scenario of authorship verification “in the wild,” and it also aligns with the settings commonly addressed in other LLM detection shared tasks. The PAN 2025 subtask 1 is again co-organized with the ELOQUENT [17] in a builder-breaker style.
The subtask 1 is in essence the classic binary detection task known also from other LLM detection shared tasks. However, we are testing the limits of the detectors by crafting a test set with text “obfuscations” that try to evade detection. Apart from drastic text length restrictions, the obfuscations we tested or received from ELOQUENT participants in the previous year had turned out to be mostly inefective. So this year, we tested what happens when the human writers obfuscate their style and whether machines can replicate this.
The training, validation, and test datasets were built from a selection of 19th-century English fiction from Project Gutenberg,1 the Extended Brennan-Greenstadt [18] corpus, and the Riddell-Juola [19] corpus. We also included a sample of the PAN’24 training and test sets (containing U.S. news articles from 2021) in this year’s training and test sets (the PAN’24 test data, which were never openly shared). Participants were free to use any other sources for augmenting their training data, including the PAN’24 training set.
The (Extended) Brennan-Greenstadt corpus has been used a lot in the fields of authorship verification and authorship obfuscation. For its construction, volunteers on Amazon Mechanical Turk were asked to submit existing writing samples (about 500 words) of their own. They were then asked to write another text describing their neighborhood, but in a way that hides their own writing style. No particular instructions were given for how this style obfuscation was to be achieved. They were also asked to write a third text in which they should try to imitate a particular style from a given sample of a novel. For PAN, we used the original and obfuscated neighborhood essays only. We chose this corpus in particular, as the obfuscated nature of the texts adds an interesting aspect to the classification task.
The Riddell-Juola corpus was created for a replication study of the original experiments by Brennan et al., but on a larger scale and with an additional control group. Volunteers were also asked to submit existing writing samples and a new piece about their neighborhood. However, the control group was not instructed to obfuscate their style.
You are an essay summarizer and a forensic writing style analyst. Given an essay, you summarize its key points in ten bullet points, extract the main topic in just a few words, and some other details. As additional details, classify the point of view from which the essay is written ("first-person", "third-person", "second-person") and the tense ("past-tense" or "present-tense"). If the essay is argumentative, classify the author’s stance on the main topic as "pro" or "con" or "neutral" if the article is not argumentative.
Point out any traits that make the author’s writing style unique, but phrase them as instructions for another writer who wants to imitate the style. For example: "Use very short sentences," "Use passive voice a lot," "Write in a nominal style with few adjectives," "Use very poetic and descriptive language," "Add spelling mistakes," "Start multiple sentences with the same word," "Use the word X more often than usual," "Write in long and nested sentences," "Use a lot of technical terms," "Use very simple language," "Use metaphors/similes/anaphoras/alliterations/...," etc.
Answer in structured JSON format (without Markdown formatting) like so: { "key_points": ["key point 1", "key point 2", ...], "main_topic": "essay topic", "pov": "narrative point of view, either first-person, third-person, or second-person", "tense": "essay tense, either past-tense or present-tense", "stance": "stance on topic, either pro, con, or neutral", "style_instructions": ["style instruction 1", "style instruction 2", ...] }
From Project Gutenberg, we sampled 927 English novels tagged as “19th-century” and “Fiction,” trying not to duplicate individual titles. The texts were cleaned of headers and footers and split into no more than 10 chunks of 500–700 words, resulting in a total of 9,185 original human texts.
From the essays, we generated JSON-structured bullet-point summaries with GPT-4o using the prompt displayed in Figure 3. The model was instructed to extract a list of key points from the text, but also stylistic markers such as the narrative point of view, the tense, stance, and any other traits that seem unique to that text (such as the use of many technical terms, very poetic language, etc.). The neighborhood texts were summarized with a similar prompt, but with the additional information that the text describes a neighborhood and that the author tried to hide their writing style. For the 19th-century fiction texts, a much simpler prompt was used that asked only for a bullet-point summary given a section of a novel.
From the summaries, we generated the final machine texts using 14 diferent LLMs: GPT-3.5 Turbo, GPT-4o, GPT-4o-mini, GPT-4.5-preview, OpenAI o1, OpenAI o1-mini, OpenAI o3-mini, Gemini 1.5 Pro, Gemini 2.0 Flash, DeepSeek-R1-Qwen-32b, Falcon3-10b, Llama3.1-8b, Llama3.3-70b, Ministral-8b-2410.
Using the summaries extracted earlier, we asked the models to write (
In addition to the style-informed prompts described above, we also prompted GPT-4o and OpenAI
o1 with five variations of the essay prompts, in which we instructed the models (
As a sixth obfuscation, we increased the temperature of DeepSeek R1 to a value of 1.1, the highest that still produced sensible text.
Finally, we also paraphrased already-generated texts from GPT-4o, GPT-4.5, and DeepSeek R1 using the adversarial DIPPER paraphraser [20]. Unfortunately, we were unable to get a satisfactory text quality from this. We used the implementation and pre-trained models provided by the authors, yet the paraphraser often generated long strings of repeated white space and punctuation marks and sometimes degenerated entirely into seemingly random tokens. We got this result with both the context-aware and the context-unaware model and with all sampling parameters we tested. To improve the quality, we tried to clean up at least some of the repeated characters using a regular expression.
From all generated texts, we randomly sampled a subset with a varying balance of human to LLM texts. We sampled from the Gutenberg set with a human:LLM ratio of 1:1, from Brennan-Greenstadt, Riddell-Juola, and PAN’24 training with a 1:4 ratio, and PAN’24 test with a 1:1 ratio.
The sampled texts were split into training, validation, and test sets. Texts created by OpenAI o1 and o1-mini texts were held back and included only in the test set. This was inspired by our earlier ifndings [ 5] that OpenAI o1 was much harder to detect when detectors were trained only on other LLMs. For this reason, we also created new o1 texts from the PAN’24 news article summaries. The texts provided by ELOQUENT participants were later added to the test set without balancing them. We kept track of the generating models and the individual source datasets, so we could later distinguish between diferent subsets of the full test set during the evaluation phase. For this reason, we distinguish between the terms dataset, which refers to all data from all sources, and data subset, which refers to texts from an individual source. The subsets were chosen so that each subset contains both classes (with the exception of the OpenAI o1 and DIPPER-obfuscated Gutenberg subsets), which is crucial for measures such as ROC-AUC and F1. A detailed listing of all subsets, splits, and class balances is given in Table 3.
We provide three baselines: Two zero-shot baselines and one supervised. Baseline (
We received submissions from 24 teams, 21 of which also submitted a work notebook describing their approach. The following section gives an overview of them with a short description for each. Table 4 summarizes the use of certain feature families and methods employed by the submitted systems. Basani and Chen [36] use GPT-2 to estimate mean, variance, skewness, and kurtosis of the negative per-token log likelihood (token “surprisal”) distributions of a text. In addition to these stationary statistical moments, they also calculate their first- and second-order rates of change. An XGBoost classifier is learned on these features for distinguishing between human and LLM texts. Huang et al. [40] fine-tune a RoBERTa model for binary classification, but does extensive data augmentation work. The existing training data is modified via synonym replacement and sentence reorganization, and noise is added via random repetitions or stop and fill word insertions. In addition, new generated documents are added using models not included in the training data, where diferent prompts are used to generate text across diferent genres to roughly double the number of generated training texts.
Jimeno-Gonzalez et al. [32] use stylometric and word-frequency features with a stacking ensemble of Random Forest, XGBoost, and LinearSVC, where a logistic regression classifier casts the ensemble vote. Specifically, the system uses the function-word ratio, average tf-idf scores, mean sentence length, mean word length, number of sentences, POS-tag distribution, and word frequency features over the vocabulary.
Kumar et al. [42] fine-tune a DistilBERT for binary classification on the provided training data. As features, the system uses the base model’s CLS embedding concatenated with five stylometric features: average word length, average sentence length, punctuation frequency, type-token ratio, and character entropy.
Larson [39] aim to be especially light-weight and use a Support Vector Machine with RBF kernel and class weighting for binary classification. As features, the system uses the 40 most frequent 1- and 2-grams, along with the 15 most frequent punctuation features, selected based on performance on the validation set.
Liang et al. [43] fine-tune a ModernBERT for binary classification and use a custom loss function weighted by an example’s dificulty.
Liu et al. [25] use an ensemble of a fine-tuned Qwen and a fine-tuned ModernBERT with contrastive loss to distinguish between human and LLM-generated text. The dataset was augmented by generating LLM paraphrases of the human texts. Macko [24] fine-tune a Qwen3-14B model via QLoRA for binary classification. What makes the approach work is the obfuscation via a homoglyph (replacement of characters with similar-looking ones Unicode characters) attack of parts of the training data (a variant of training data enhancement) and model selection on external training data to select the best model based on out-of-domain performance. The external training data is a collection of 2,000 examples across seven languages, sampled from 18 diferent AI-detection datasets of diferent genres and domains [44].
Marchitan et al. [30] describe two system: First, a voting-ensemble of LightGBM, XGBoost, Logistic Regression, and an SVM using embeddings from Qwen3-0.6B, and, second, fine-tuning a classification head on top of an LLM, where Qwen2.5-0.5B, Qwen3-0.6B, Mistral7B-v0.1, and Llama-3.1-8B were tested.
Ochab et al. [34] use an LGBM classifier on four types of stylometric features extracted using spaCy and inspired by the “stylo” R package. The dataset was augmented by adding about half a million texts from other AI detection datasets. 10-fold cross-validation was used to select the best hyperparameters for the model.
Ostrower et al. [35] propose three systems. The first system is a XGBoost ensemble using the Binoculars score, tf-idf scores, and BERT embeddings as features. The second, not submitted system uses a Maximum Likelihood Estimation based on various features: cohesiveness, type-token ratio, word count, stop-word frequency, non-sentiment word frequency, POS distribution. Cohesiveness is computed via the average BARTScore (between a text and multiple noisy, obfuscated copies) and multiple zero-shot LLM detectors (LRR, FastDetectGPT). The third approach uses adversarial training following Hu et al. [45].
Pudasaini et al. [37] use various ensembling strategies (voting, stacking, and boosting) with a number of fine-tuned pre-trained language models (such as DeBERTa, Longformer, RoBERTa, etc.). The results were tested on the PAN’25 dataset and the COLING’25 dataset.
Teja et al. [31] fine-tune several pre-trained language models (such as DeBERTa, DistilBERT, XLNet, and others) with a mixture-of-experts gating mechanism. They tested a hard gating mechanism which selects only a single experts and a soft gating mechanism which uses a linear layer with softmax to select a weighted sum of expert outputs. A DeBERTa model with hard gating performed best on the validation set and was submitted to Tira.
Seeliger et al. [26] generate a matrix of cumulative binary correlation coeficients between terms and documents. Three diferent versions were submitted using word unigrams, bigrams, and trigrams as terms. The authors also supplied two baselines, one of which is a fine-tuned RoBERTa model. The approach allows for stationary analysis of the whole document, but also temporal analysis of the cumulative sum of correlation coeficients per word.
Sun et al. [38] use a combination of 25 stylometric and 25 entropy-based features and a voting ensemble of five diferent classifiers (Gaussian Naive Bayes, AdaBoost, LightGBM, CatBoost, Random Forest). The stylometric features selected via univariate feature selection from 101 features suggested by the Claude LLM. The entropy-based features are selected via univariate feature selection from 72 statistical features describing the distribution of per-token forward and backward cross-entropy losses, following Guo et al. [46]. These losses are computed by “regenerating” a training text via teacher forcing, given a generated summary as prompt, and measuring the loss of each predicted output logits to the target (forward) and to the last token of the prefix (backward), according to a Llama2-7b. Titze and Halvani [41] use of-the-shelf pre-trained LLMs to extract negative log likelihood “surprisal”, mean Shannon entropy, log rank, Jensen-Shannon divergence from the token representation of the text. A logistic regression classifier is learnt to combine the features into a final score. Valdez-Valenzuela generate a syntactic dependency graph representation of the text. Sentence-level graphs are merged into document-level graphs, which are embedded using a graph neural network for use in a dense neural network for classification. The dataset is augmented with three diferent kinds of obfuscations (shortening, Unicode replacement, paraphrasing) to make the system more robust. Völpel and Halvani [33] use linearized constituent trees n-grams as features for a feed-forward neural network classifier. The trees are parsed via the Constituent Treelib library [47] and segmented by traversing from every node to all possible leaves and counting all encountered paths of length 1-7. Voznyuk et al. [27] use multi-task learning with DeBERTa-v3 base. The system learns three tasks, each utilizing a diferent head: the binary Voight-Kampf task, a 3-class genre prediction task, and a 4-class model family prediction task. The model is then trained for all tasks, albeit only some losses are propagated to the base model to prevent overfitting. The results of the Voight-Kampf classification head are reported for the competition.
Yang and Yan [28] fine-tune a BERT model using genre-dependent contrastive loss. For this, the CLS output token is concatenated with a learned genre vector, which is then fed into an MLP with contrastive loss.
Zaidi et al. [29] fine-tune an uncased BERT-base for binary classification of human and LLM texts.
Team
The final system ranking based on the PAN and ELOQUENT test sets is listed in Table 5. Participating teams that submitted more than one system are ranked only with their best-performing system.
All systems were submitted and evaluated on Tira [7]. At test time, the participants had to calculate a score between 0 and 1 for each text, indicating the likelihood that the text was LLM-generated. A score of exactly 0.5 could be given to signal a non-decision.
For each participant, we computed a confusion matrix and the following scores, which we used in previous authorship verification shared tasks as well: • ROC-AUC: The area under the Receiver Operating Characteristic curve. • Brier: The complement of the Brier score (mean squared loss) • C@1: A modified accuracy score that assigns non-answers (score = 0.5) the average accuracy of the remaining cases [48]. • F1: The harmonic mean of precision and recall.
F1
Macko Valdez-Valenzuela
Liu Seeliger Voznyuk
Yang
Zaidi hello-world Baseline TF-IDF SVM
bohan-li Marchitan
Teja xlbniu Jimeno-Gonzalez
Völpel
Ochab Ostrower
Basani Pudasaini
Sun Larson Huang
Titze Baseline Binoculars Llama3.1
Kumar Baseline PPMd CBC
Liang 0.4 0.5
0.6 0.7 0.8 Mean Score Per Dataset
• F0.5u: A modified F 0.5 measure (precision-weighted F measure) that treats non-answers (score = 0.5) as false negatives [49].
• Mean: The arithmetic mean of all previous measures
The six measures are calculated for each system on each of the test data subsets described in Section 3.1.4. A final system score is calculated per measure as a macro average over all subsets. The resulting macro mean score is used to determine the system’s ranking in the final leaderboard. The mean scores for data subsets with only positive examples (i.e., the held-back OpenAI o1 and DIPPERobfuscated Gutenberg subsets) is calculated without the contribution of ROC-AUC, as the value would be undefined. The precision for the F 1 and F0.5u calculations is assumed to be 1.0 in those cases.
Table 5 shows the final macro scores per team and system. 2 If teams submitted multiple systems, only the best-performing system is considered in the ranking. Of the 24 submitted systems, 8 beat the strongest baseline (TF-IDF SVM) and 14 more beat the second-strongest baseline (Binoculars). Macko [24] lead the ranking with Valdez-Valenzuela (no notebook) and Liu et al. [25] being the runners-up.
Figure 4 shows the distribution of mean scores for each system over test data subsets from which the macro means are calculated. Most systems range between 0.7 and 1.0 for most subsets, apart 2Note: Bevendorf et al. [16] and an earlier draft of this paper described significantly worse scores for all systems. This was due to a score calculation error on the ELOQUENT data subset, which has been corrected in this version.
Riddell-Juola (Obfuscated) Brennan-Greenstadt (Obfuscated)
ELOQUENT PAN'24 Test PAN'24 (o1)
Riddell-Juola
Gutenberg Fiction
Brennan Greenstadt Gutenberg Fiction (Obfuscated)
Gutenberg Fiction (o1)
Riddell-Juola (o1)
Riddell-Juola
PAN'24 Test Riddell-Juola (Obfuscated)
Brennan Greenstadt tse ELOQUENT taBrennan-Greenstadt (Obfuscated) aD Gutenberg Fiction
PAN'24 (o1) Gutenberg Fiction (Obfuscated)
Gutenberg Fiction (o1)
Riddell-Juola (o1) Brennan-Greenstadt (Obfuscated)
Riddell-Juola (Obfuscated)
ELOQUENT
PAN'24 (o1) tse Gutenberg Fiction ta Gutenberg Fiction (Obfuscated) aD PAN'24 Test
Gutenberg Fiction (o1)
Riddell-Juola Brennan Greenstadt
Riddell-Juola (o1) 0.0 0.0 0.2 0.4 1.0 – C@1: All Systems from certain outliers. On the ELOQUENT subset, the majority of systems achieve scores between 0.7 and 0.8, dragging down their final scores. The best system by Macko [24] is an exception to this rule by displaying more than solid performance also on ELOQUENT. The mean false positive rate of all systems on ELOQUENT’s human texts is 0.15, which is probably best explained by the topic domain shift from the PAN training data. Cross-domain classification has been a longstanding problem in authorship identification [50], but also in LLM detection in particular [51].
If we analyze the distribution C@1 scores per subset (Figure 5), we can see the obfuscated BrennanGreenstadt and Riddell-Juola and ELOQUENT subsets seem to be the most dificult. However, if we look at false positive and false negative rates separately, we can find this efect only in the latter. That means LLMs were more successful than humans in obfuscating their distinct “LLM style,” whereas the obfuscated human texts still look human enough. This is not entirely unexpected, since the human authors were only asked to mask their personal style, not to “write like a machine” (and LLMs did not exist back then, anyway). The LLMs, on the other hand, apparently managed to deviate suficiently from their usual style that makes them otherwise stand out. This efect, however, must also be attributed in large part to the other obfuscations we applied. Due to the way the test set was sampled, only parts of these subsets are generated from the plain neighborhood prompt (57 % for Brennan-Greenstadt, 18 % for Riddell-Juola). Hence, further analysis is needed to distinguish between these two efects. Since these obfuscations are in part also purely prompt-based, we can assume, though, that LLMs did manage to change their style suficiently. Considering that the human texts were classified correctly for the most part, the topic / domain shift alone cannot serve as an explanation.
The Gutenberg data, on the other hand, seems to be the easiest, which may indeed be a result of the text homogeneity between training and test. The held-back OpenAI o1 texts also posed no particular challenge. This is in contrast to our previous findings [ 5]. This is even true for the TF-IDF SVM baseline, which is identical to the detector we used in those earlier experiments. This peculiarity definitely warrants more research. Perhaps the inclusion of GPT-4.5 texts in the training data played a role, but perhaps the included sample of only 300 texts was too small.
Figure 6 shows the efectiveness of individual obfuscations (not including the plain neighborhood prompt) described in Section 3.1.3. DIPPER paraphrasing appears to be highly efective in avoiding detection, but this comes at a heavy price with regard to text quality. Hence, this result should be taken with a grain of salt. The second-most efective obfuscation appears to be the prompt to write like a 7-year-old. This obfuscation uses only the prompt to change the model output, which appears to be suficient to successfully avoid detection in most cases. Prompting the model to change the word order to subject-object-verb shows a medium efectiveness. All three are even more efective if tested against only the top-10 systems, hinting at overfitting. Increasing the temperature, prompting for alliterations, or inserting random words into the prompt did not have any significant impact.
The overall best performing system by Macko [24] had a false-negative rate of 42 % on DIPPERparaphrased texts, but was entirely unafected by all other obfuscations.
The rise of AI-assisted writing is transforming the traditional notion of authorship and steadily blurring the boundary between human and machine contributions. As AI involvement now spans the full spectrum—from none to complete [52], new questions of ethical and intellectual accountability arise. Subtask 2 seeks to address these challenges by asking the participants to classify a collaboratively authored human-AI document into one of six categories, defined from an ethical and intellectual accountability perspective: i. fully human-written; ii. human-written, then machine-polished; iii. machine-written, then machine-humanized (obfuscated); iv. human-initiated, then machine-continued; v. deeply mixed text, where some parts are written by a human and some are generated by a machine; vi. machine-written, then human-edited.
The training and validation sets for Subtask 2 were constructed from existing datasets focused on ifne-grained machine-generated text detection. We provided 288,918 examples for training and 72,661 for validation. For testing, we collected four new datasets, with an emphasis on student essays and peer reviews. We additionally incorporated five recently released datasets to comprehensively evaluate the generalization of detection systems across unseen generators and domains. The test set consists of 140,756 instances. The dataset statistics, including the distribution across six categories, are shown in Table 6. All subsets underwent the cleaning procedure: we removed duplicates and filtered out texts shorter than 30 characters. Below, we describe each component dataset in detail, with an overview provided in Table 7.
MixSet [53] includes machine-polished human-written texts and human-edited machine-generated texts. We sampled 3,491 out of 3,600 texts, ensuring a minimum text length of 30 characters. LLM-DetectAIve [54] is a large-scale benchmark designed to improve machine-generated text detection across diferent domains and text variations. The dataset is based on M4GT [ 55], which includes a mix of human-written and machine-generated texts from sources such as arXiv, WikiHow, Wikipedia, Reddit, student essays (OUTFOX), and peer reviews (PeerRead). LLM-DetectAIve extends the M4GT dataset by adding new machine-generated texts from more advanced models (e.g., GPT-4o), machine-humanized MGTs, and human-written texts that were polished using LLMs. The resulting dataset comprises 91,358 fully machine-generated texts, 103,852 machine-humanized texts, and 107,900 human-polished texts. LLM-DetectAIve includes outputs from a diverse set of LLMs, including Llama38B/70B, Mixtral-8x7B, Gemma-7B, Gemma-2-9B, GPT-4o, Gemini-1.5-Pro, and Mistral-7B. For Subtask 2, we excluded the fully machine-generated texts and used 286,169 examples from the remaining two categories.
RoFT [56] is a collection of over 21k human annotations paired with error classifications to investigate how various variables such as model size, decoding strategy, and fine-tuning afect human detection performance. The dataset is in English and contains domains like recipes, presidential speeches, short stories, and New York Times. Each instance is initiated by a human and then continued by either a human or an LLM including GPT-2 [68], GPT-3 [69] and CTRL [70]. We preprocess the data by performing basic cleaning, removing duplicates and sanity entries. This results in a total of 9,148 instances that are either fully human-written or human-initiated and machine-continued. RoFT-ChatGPT [57] is an augmented version of the Real or Fake Text (RoFT) dataset. The augmented version includes GPT-3.5-turbo generations. The dataset is entirely in English and covers domains such as short stories, recipes, New York Times news articles, and presidential speeches. It consists of 6,940 samples of text that are human-initiated and machine-continued, ofering a rich resource for evaluating the interaction between human authorship and AI-generated text. Subtask 2 Component Datasets used for training, development and test sets. Size refers to the number of examples we sampled from a given dataset, rather than the full set of original dataset. In addition to leveraging existing datasets, we collected three datasets based on ICNALE, NLPeer and Peersum, and we initiated a dataset MBZUAI-CLEF based on our real-life usage.
Training and Development Sets Dataset Co-author [58] detected sentence-level boundaries of human-AI collaborative mixed texts. They identified challenges in detecting AI-generated sentences in mixed texts, such as human writers’ selecting and even editing AI-generated sentences based on personal preferences; the frequent change of authorship between neighboring sentences within the mixed text; and the short length of text segments within mixed texts, which provides on limited stylistic cues for reliable authorship determination.
Coauthor comprises 1,447 writing sessions produced through the collaboration of human writers, recruited from Amazon Mechanical Turk, and a GPT-3 assistant [69]. These sessions encompass both creative and argumentative writing. Writers were provided with a prompt as a starting point and could either compose independently or request sentence suggestions from the assistant. The suggested sentences could be accepted or rejected, and this iterative process continued until the essay was completed. This setup results in a highly mixed text, where some parts are authored by humans and others are generated by the machine.
TriBERT [59] is a dataset of human-AI collaborative writing. First, the authors collected the humanwritten essays from U.S.junior high school students (grades 7-10). They then designed 8 text modification tasks, each involving the removal of specific segments from the original human-written texts, followed by LLM-generated fill-in. This approach resulted in 8 distinct forms of human-AI hybrid writing. In our six-label space, texts that start with a human-written introduction and then were completed by LLMs are mapped as human initiated, machine continued, and all other cases are classified as deeply mixed. This transformation doubles, resulting in 34,272 texts, and then we sampled 21,178. LAMP [60] consists of 1,057 paragraphs generated by LLMs and subsequently edited by professional writers. The original paragraphs were sourced from reputable publication venues, including The New Yorker, The New York Times, and Dear Sugar. These paragraphs cover various domains, such as fiction, food writing, and internet advice. In total, approximately 1,200 paragraphs were selected, with the Literary Fiction genre accounting for 80% and the remaining 20% categorized as Creative Non-Fiction.
The instructions for generating these paragraphs were created using instruction back-translation [71]. In this process, GPT-4o was prompted to summarize each paragraph into writing instructions. After manual verification, ill-formed or overly specific instructions were filtered out, resulting in a total of 1,057 high-quality instructions. These instructions were then used to generate additional writings using GPT-4o, Claude-3.5-Sonnet, and Llama 3.1-70b. Finally, the LLM-generated responses were further refined and edited by a team of 18 professional writers, following a comprehensive edit taxonomy informed by expert writing practices. 4.1.2. Test Set Beemo [61] Benchmark of Expert-edited Machine-generated Outputs (Beemo) is a dataset designed to support fine-grained detection of machine-generated text (MGT), particularly in multi-author scenarios where LLM outputs are refined by either human experts or other LLMs. It comprises a total of 19.6k English texts spanning five tasks: open-ended generation, rewriting, summarization, and open/closed QA. We used 17,331 examples, removing the purely machine-generated texts. Specifically, this includes 2,184 human-written texts, 2,183 machine-written, then human-edited texts, with the rest as machinewritten, then machine-humanized.
AI-Polished-Text Evaluation (APT-Eval) Based on the 300 purely human-written texts sampled
from MixSet [53], Saha and Feizi [62] collected 11.7K machine-polished data using four diferent models:
GPT-4o, Llama3.1-70B, Llama3-8B, and Llama2-7B with two polishing strategies. For degree-based
polishing, the LLM is prompted to refine the text in four varying degrees of modification (
HART [63] is a dataset for fine-grained machine-generated text detection, including categories of purely human-written, human-written and then machine-polished, machine-generated and then machine-humanized, purely machine-generated text. There are 16K English examples, and 4K examples for the other four languages including Chinese, Arabic, French, and Spanish. Six LLM generators were involved in data collection, i.e., GPT-3.5-turbo, GPT-4o, Claude-3.5-Sonnet, Gemini-1.5-pro, Llama-3.370B-instruct, and Qwen-2.5-72B-instruct. They additionally gathered 250 machine-generated humanedited texts. This dataset can be used as part of our test set.
LLMDetect [64] is designed to distinguish between four categories of text origin: Human-Author (fully human-written), LLM-Creator (entirely machine-generated), LLM-Polisher (human-written text subsequently refined by an LLM), and LLM-Extender (human-written text extended or expanded by an LLM). LLMDetect includes the Hybrid News Detection Corpus (HNDC) for training detectors, as well as DetectEval, a comprehensive evaluation suite that considers five distinct cross-context variations and two multi-intensity variations within the same LLM role. This allows for a thorough assessment of detectors’ generalization and robustness across diverse contexts. academic paper/abstract writing, paper summary, peer review, rebuttal, Email writing, administrative letters, recommendation requests, ethical response letters, commitment letters PhD SOP review, teaching statements, job interview preparation, oral interview explanations, volunteer applications Story writing, daily chatting, continued writing GitHub issues, README files, report writing, NLP slides summary
Prompt refining
ALTA 2024 Shared Task Mollá et al. [65] employ four distinct construction patterns to organize
human and machine-generated sentences: (
We collected three datasets based on ICNALE3, Peersum, and NLPeer, and we further gathered a dataset from scratch based on authors’ daily usage of LLMs.
corpus. We sampled 5,843 topic-controlled essays produced by more than 5,500 college students (incl. grad students) in ten countries/regions in Asia (China, Hong Kong, Taiwan, Indonesia, Japan, Korea, Pakistan, Philippines, Singapore, Malaysia, and Thailand), as well as English native speakers. Based on these human-written essays, we further generated 2,624 human-written, then machine-polished, 1,000 machine-written, then machine-humanized, and 656 human-initiated, then machine-continued text, using the latest SOTA LLM including GPT-4.1, Qwen3-30B-a3b, Llama-3.3-70B, Deepseek-chat-v3, Gemini-2.5-pro-preview, resulting in a total of 10,123 examples.
Peersum [66] is a dataset designed for generating meta-reviews of scientific papers based on reviews.
The meta-reviews can be interpreted as abstractive summaries of reviews, multi-turn discussions
among reviewers and the paper author, and the paper abstract. We sampled 1,000 human-written
metareviews and the corresponding reviews. We used GPT-4.1-2025-04-14, GPT-4.1-mini-2025-04-16, and
Deepseek-R1 to further produce 2,000 human-written, then machine-polished, 2,000 human-initiated,
then machine-continued, and 270 machine-written, then machine-humanized cases, in total of 5,270.
NLPeer [67] is intended for the study of peer review and approaches to NLP-based assistance to peer
review. Based on human-written reviews, we used GPT-4.1 and GPT-4.1-mini for human-written, then
machine-polished (
MBZUAI-CLEF is a dataset we collected from scratch. The human-written texts cover six categories including academic writing, professional correspondence, application materials, creative writing, technical and support writing, and prompting engineering, with 378 examples in total, as shown in Table 8. These human-written drafts were polished using GPT-4o, after which the authors further edited the polished outputs to adapt them to real-world applications.
The Subtask 2 was conducted in two phases: Development Phase. Only training and development data were provided to participants, with gold labels available for the development set. Although they were not allowed to use external training and validation data, data augmentation strategies such as back-translation, synonym replacement, random word deletion, and replacement were allowed.
Participants competed against each other to achieve the best performance on the development set. A live leaderboard on CodaLab was made available to track all submissions.4 Teams could make an unlimited number of submissions, and the best score for each team, regardless of the submission time, was displayed in real time on CodaLab.
Test Phase. Participants were given approximately one week to prepare their predictions. Then, participant teams submitted their results to CodaLab.5 They could submit multiple runs, but they would not receive feedback on their performance. Only the latest submission from each team was considered “oficial,” and was used for the final team ranking. In total, 22 teams submitted results, of which 16 submitted system description papers. After the competition concluded, we released the gold labels for the test set. Furthermore, we kept the submission system open for the test dataset for post-shared task evaluations and to monitor the state of the art.
Predictions of all systems were submitted and evaluated in CodaLab. At the test time, participants
assigned the predicted label among [0, 1, 2, 3, 4, 5] for each text, indicating its category. Participants
in the leaderboard were ranked by macro-recall. Macro-recall is selected as the primary evaluation
metric for two reasons: (
To establish a baseline, we fine-tuned a pre-trained transformer-based model RoBERTa on the training set. Fine-tuning was performed using the Hugging Face Trainer API with the following configuration: learning rate of 2 × 10− 5, batch size of 16 for both training and evaluation, weight decay of 0.1, and a total of 3 training epochs. Checkpoints were evaluated at the end of each epoch, and the best-performing model on the development set was retained for subsequent testing. The baseline achieved a macro-recall of 68.67 % on the development set, with corresponding macro-F1 and accuracy scores of 61.26 % and 56.71 %, respectively. 4.5. Submissions 22 teams submitted their predictions to CodaLab, of which 16 submitted notebook papers [24, 26, 30, 29, 27, 31, 72, 76, 77, 74, 79, 81, 78, 75, 73, 80]. The performance of 14 teams is above the baseline, and 8 teams are below fine-tuned RoBERTa-base, as shown in Table 9. One team submitted their prediction ifle after the deadline, which we marked in gray. Additionally, four teams submitted files with IDs misaligned with the test set. The test set contains IDs ranging from 0 to 141,410, with 655 entries filtered out, rather than forming a continuous range from 0 to 140,755. For fairness, we ranked all submissions based on their original versions. However, we also corrected the misaligned IDs and re-evaluated the 4https://codalab.lisn.upsaclay.fr/competitions/22620 5https://codalab.lisn.upsaclay.fr/competitions/22934
Team QLoRa PEFT fine-tuned Qwen3-4B-Base.
Under-sample high-frequency classes and adopt data augmentation for underrepresented classes, along with R-Drop regularization for DeBERTa-v3-base fine-tuning. – Shared Transformer Encoder between several classification heads trained to distinguish the domains.
Combine the deep language understanding of DeBERTa-v3-large and the high-dimensional mapping ability of StarBlock2d.
DeBERTa enhanced by contextual and geometric attention Use DeBERTa-v3-Large Fine-tune Gemma-2 2B for sequence classification with multiple classification heads. – Fine-tune DeBERTa-V3-Large and combining multi-scale features. – Combine the contextual strength of BERT with the sequence modeling capabilities of Transformer layers.
Fine-tune DeBERTa-V3-Large and combine it with BiLSTM and attention mechanism.
Soft and Hard Mixture of Experts (MoE) architectures with DeBERTa-V3-Large – Cumulative sum of token-Level correlation signals – Fine-tune RoBERTa Fine-tune RoBERTa with class balancing, data augmentation, and calculation of specific weights for each unbalanced class. – Combine features at diferent layers extracted using Transformers with layer-wise projection and attentive pooling.
Fine-tune transformer models with data augmentation strategies on underrepresented classes (late submission). – – – – – F1 afected submissions to better reflect the actual performance of their detection systems. The following analysis is based on these corrected results.
Many teams fine-tuned DeBERTa-v3-large and achieved better results than RoBERTa. Larger language models such as Qwen-3 4B and Gemma-2 2B were superior to DeBERTa and RoBERTa, see more in Table 10. The performance drop observed on the test set compared to the development set highlights the need for further improvement in fine-grained human-AI collaborative text detection.
M L P l l a m S
M L P e g r a
L ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ g n i n u t e n i F ✓ . g u A a t a
D ✓ ✓ ✓ ✓ ✓ . b m o C s e r u t a e F ✓ e l b m e s n E ✓ ✓ s d a e H .l t u M Qwen3-4B (QLoRA PEFT) DeBERTa-v3-base (R-Drop) DeBERTa-base (Multi-task) DeBERTa-V3-Large (Hybrid) DeBERTa-V3-Large (Geom. Attn) DeBERTa-V3-Large Gemma-2 2B (Multi-head) DeBERTa-V3-Large (Multi-scale) BERT_T DeBERTa-V3-Large + BiLSTM DeBERTa-V3-Large (MoE) Token-level correlation signals RoBERTa-base (Class weight) RoBERTa-base (Layer proj.) RoBERTa-large BERT-base (Sent. segment)
The highest-ranking system, mdok [24], fine-tuned the Qwen3-4B-Base model using QLoRA for eficient parameter adaptation. In comparison, Bohan Li [72] used DeBERTa-v3-base enhanced with data augmentation, balancing, and R-Drop regularization, showing strong performance despite a smaller backbone.
Several teams built upon DeBERTa-V3-Large for human-AI collaborative writing detection. StarBERT [73], Atu [74], TaoLi [75], and Zhangzhiliang [79] all fine-tuned this model, each introducing unique extensions: StarBERT proposed a hybrid classification framework, Atu incorporated contextual and geometric attention, TaoLi focused on fine-grained categorization, and Zhangzhiliang added a BiLSTM layer with attention mechanisms. Advacheck [27] also used DeBERTa-base within a multi-task learning framework to distinguish between human, machine, and hybrid authorship styles.
DetectTeam [77] enhanced DeBERTa-V3-Large with a Feature Pyramid Network to capture multiscale features, while CNLP-NITS-PP [31] employed the Mixture-of-Experts (MoE) architectures with SoftMoE and HardMoE for dynamic model routing.
Other teams opted for RoBERTa-based architectures. VerbaNex [80] fine-tuned RoBERTa-base and addressed data imbalance using augmentation, oversampling, undersampling, and loss weighting. Nexus [29] fine-tuned RoBERTa-Large, applying targeted data augmentation techniques including back-translation, synonym/antonym substitution, and random deletion.
ReText.Ai [76] adopted a multi-head classification strategy over the Gemma-2 2B model, while WeiDongWu [78] introduced BertT, a hybrid model combining BERT-base with an extra Transformer encoder and dropout layers for six-class collaborative classification.
Unibuc-NLP [30] merged transformer layers using projection and attentive pooling. Steely [26] converted text into interpretable word-level correlation signals. Fuchuan [81] applied a BERT-based model for sentence-level classification on segmented inputs.
0 1
2 3 Predicted label
4
Confusion Matrix of lza le2 678 6158 16330 75 b a lloG3 684 4544 5941 10926 625 d
336 1657 5 209 1090 1007 10 0 1
2 3 Predicted label 93 4 61 5 82 148 5 0 25954 6135 1223 74 387 736 0 25840 6124 1371 73 386 715 1 2966 33975 5087 31 4 100 279 514 191 11411 5
4 100 278 518 191 11408 5 le2 599 5965 17217 125 275 1053 b a llodG3 417 1827 3014 16761 691 92 5000 30000 25000 20000 15000 10000 5000 25000 20000 15000 10000 5000 30000 25000 20000 15000 10000 5000
We show the macro-averaged recall and label-wise recall scores of 22 submissions for Subtask 2 in Table 11. A clear trend emerges in terms of label dificulty. Label 4 (deeply-mixed text) consistently receives the highest recall across systems, indicating it is the easiest class to detect. This is followed by label 1 (human-written, then machine-polished) and label 0 (purely human-written), which also achieve high recall scores for most systems. In contrast, label 5 (machine-written, then human-edited) proves to be the most challenging, with most systems performing poorly on this category, and only a few submissions (e.g., team 7) achieving a recall above 9.0 %.
This trend partially aligns with the label distribution in the training data. The easiest labels 1 (humanwritten, then machine-polished) and 0 (purely human-written) are among those with the highest number of training examples. In contrast, labels 5 (machine-written, then human-edited) and 3 (human-initiated, then machine-continued) are the most challenging to detect, which also corresponds to their small training sizes. While label 2 (machine-written, then machine-humanized) has a comparable amount of training data to label 1 (91k vs. 95k), its performance is significantly worse across most systems.
The confusion matrices in Figure 7 further confirm this pattern. For all top-performing systems (Teams 1–3 and lza), label 2 is frequently misclassified as 1. This suggests that machine-humanized texts (label 2) share stylistic cues with machine-polished texts (label 1) likely due to the dominating influence of machine-generated text in both. Conversely, label 5 exhibits low recall and high confusion with label 2, indicating that human edits often obscure the origin of machine-generated text, making it particularly dificult to detect.
Taken together, these findings underscore the nuanced dificulty of distinguishing between varying degrees of human-AI collaboration. While abundant data helps, the boundary between polished and humanized content remains inherently fuzzy, calling for more fine-grained modeling approaches beyond token-level patterns or surface style cues.
The system performance across the nine component datasets in the testbed is shown in Table 12. The top-ranked systems achieve the highest accuracies on datasets such as NLPeer (peer review), ALTA (essay), and ICNALE (essay), which consist primarily of academic and well-structured text genres. These datasets likely provide clearer linguistic signals and stylistic features for distinguishing diferent human-AI collaborative texts. For instance, the best-performing system achieves 90.97 % accuracy on ALTA and 96.90 % on NLPeer. These results suggest that model generalization is strongest in formal writing settings, particularly in educational and peer-review contexts.
In contrast, performance drops significantly on datasets such as MBZUAI-CLEF, Peersum, and Beemo, which contain more diverse genres and informal domains. MBZUAI-CLEF, for example, covers a wide range of real-world writing scenarios, including rebuttals, admin letters, emails, oral presentations, and GitHub README files, but consists of only 1,115 examples. All systems perform poorly on this dataset, with accuracies below 48 %, highlighting the dificulty of detecting fine-grained signals in low-resource, noisy, or mixed-genre settings. Similarly, Peersum and Beemo include peer review and open-domain generation tasks, both of which challenge the systems due to variability in style, structure, and intent.
We also observe that datasets with larger size and consistent formatting, such as LLMDetect (48k examples), HART (24k), and ALTA (16.5k), tend to support more stable performance across submissions. This indicates that both data volume and stylistic consistency facilitate more efective fine-tuning and generalization. The wide variance in performance across component datasets underscores the importance of robustness and adaptability in collaborative text detection. Future work should prioritize developing detection systems capable of generalizing across low-resource, diverse, and informal domains, where style-based cues may be weaker or inconsistent.
The PAN Subtask 1 received submissions from 24 teams, of which 21 also submitted a work notebook. Most of the submitted systems were quite strong, with the best system being almost perfect on the PAN test data. However, unlike in 2024, the texts submitted by ELOQUENT participants posed a dificult challenge for PAN systems. In part, this can be explained by the topic shift between the ELOQUENT dataset and the training data provided by PAN, but (almost) all systems also struggled with certain obfuscations applied to the texts, including ones that relied solely on prompting techniques without changing the actual topic.
The PAN Subtask 2 introduced a nuanced, six-category classification of human-AI collaboration, reflecting the complex reality of modern text production and arising ethical and intellectual accountability challenges. While participants’ systems could distinguish certain categories like “deeply-mixed text” with high recall, they struggled significantly to identify “machine-written, then human-edited” texts—a crucial and challenging real-world scenario. Performance was heavily influenced by the distribution of training data and the domain of the text, with systems performing better on structured academic writing than on more diverse, informal datasets. The dificulty of distinguishing between machine-polished and machine-humanized text further underscores the fuzzy boundaries in collaborative writing. Together, these findings indicate that while progress has been made, the reliable detection of AI authorship, especially in its more subtle and varied forms, remains an open and pressing challenge.
The work of Janek Bevendorf, Matti Wiegmann, Maik Fröbe, Martin Potthast, and Benno Stein on PAN subtask 1 has been funded as part of the OpenWebSearch project by the European Commission (OpenWebSearch.eu, GA 101070014).
Text, datasets, experiments, and analyses in this paper were created and conducted by the authors themselves. Generative AI was used by some authors for assistance in the writing process, but no substantial parts were generated by it. G. Puccetti, T. Arnold, SemEval-2024 Task 8: Multidomain, Multimodel and Multilingual MachineGenerated Text Detection, in: Proceedings of the 18th International Workshop on Semantic Evaluation (SemEval-2024), 2024, pp. 2057–2079. [9] Y. Wang, A. Shelmanov, J. Mansurov, A. Tsvigun, V. Mikhailov, R. Xing, Z. Xie, J. Geng, G. Puccetti, E. Artemova, J. Su, M. N. Ta, M. Abassy, K. A. Elozeiri, S. E. D. A. El Etter, M. Goloburda, T. Mahmoud, R. V. Tomar, N. Laiyk, O. M. Afzal, R. Koike, M. Kaneko, A. F. Aji, N. Habash, I. Gurevych, P. Nakov, GenAI Content Detection Task 1: English and Multilingual Machine-Generated Text Detection: AI vs. Human, in: F. Alam, P. Nakov, N. Habash, I. Gurevych, S. Chowdhury, A. Shelmanov, Y. Wang, E. Artemova, M. Kutlu, G. Mikros (Eds.), Proceedings of the 1stWorkshop on GenAI Content Detection (GenAIDetect), International Conference on Computational Linguistics, Abu Dhabi, UAE, 2025, pp. 244–261. URL: https://aclanthology.org/2025.genaidetect-1.27/. [10] S. Saha, R. Das, D. Das, JUNLP_SS at ELOQUENT Lab 2025: Humanizing AI - Enhancing the Realism of Machine Generated Text, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), 26th Working Notes of the Conference and Labs of the Evaluation Forum, CLEF 2025, CEUR-WS, 2025. [11] A. Creo, M. Hormazábal-Lagos, H. Cerezo-Costas, P. Alonso-Doval, HumanAIzers in VoightKampf at ELOQUENT 2025, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), 26th Working Notes of the Conference and Labs of the Evaluation Forum, CLEF 2025, CEUR-WS, 2025. [12] P. Vachharajani, Literal Re-translation as a Method for AI Text Disguise and Detection Evasion, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), 26th Working Notes of the Conference and Labs of the Evaluation Forum, CLEF 2025, CEUR-WS, 2025. [13] R. R. Gunti, The Data-Centric Approach for the Voight Kampf Task, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), 26th Working Notes of the Conference and Labs of the Evaluation Forum, CLEF 2025, CEUR-WS, 2025. [14] M. Hoveyda, Bypassing Human/Machine Classifiers by Prompting LLMs for Naturally Imperfect Text, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), 26th Working Notes of the Conference and Labs of the Evaluation Forum, CLEF 2025, CEUR-WS, 2025. [15] A. A. Ayele, N. Babakov, J. Bevendorf, X. B. Casals, B. Chulvi, D. Dementieva, A. Elnagar, D. Freitag, M. Fröbe, D. Korenčić, M. Mayerl, D. Moskovskiy, A. Mukherjee, A. Panchenko, M. Potthast, F. Rangel, N. Rizwan, P. Rosso, F. Schneider, A. Smirnova, E. Stamatatos, E. Stakovskii, B. Stein, M. Taulé, D. Ustalov, X. Wang, M. Wiegmann, S. M. Yimam, E. Zangerle, Overview of PAN 2024: Multi-Author Writing Style Analysis, Multilingual Text Detoxification, Oppositional Thinking Analysis, and Generative AI Authorship Verification, in: Experimental IR Meets Multilinguality, Multimodality, and Interaction. 15th International Conference of the CLEF Association (CLEF 2024), volume 14959 of Lecture Notes in Computer Science, Springer, Berlin Heidelberg New York, 2024, pp. 231–259. doi:10.1007/978-3-031-71908-0\_11. [16] J. Bevendorf, D. Dementieva, M. Fröbe, B. Gipp, A. Greiner-Petter, J. Karlgren, M. Mayerl, P. Nakov, A. Panchenko, M. Potthast, A. Shelmanov, E. Stamatatos, B. Stein, Y. Wang, M. Wiegmann, E. Zangerle, Overview of PAN 2025: Voight-Kampf Generative AI Detection, Multilingual Text Detoxification, Multi-Author Writing Style Analysis, and Generative Plagiarism Detection, in: J. C. de Albornoz, J. Gonzalo, L. Plaza, A. G. S. de Herrera, J. Mothe, F. Piroi, P. Rosso, D. Spina, G. Faggioli, N. Ferro (Eds.), Experimental IR Meets Multilinguality, Multimodality, and Interaction. Proceedings of the Sixteenth International Conference of the CLEF Association (CLEF 2025), Lecture Notes in Computer Science, Springer, Berlin Heidelberg New York, 2025. [17] J. Karlgren, E. Artemova, O. Bojar, V. Mikhailov, M. Sahlgren, E. Velldal, L. Øvrelid, ELOQUENT CLEF shared tasks for evaluation of generative language model quality, 2025 edition, in: Advances in Information Retrieval: 47th European Conference on Information Retrieval, ECIR 2025, Lucca, Italy, April 6–10, 2025, Proceedings, Part V, Springer-Verlag, Berlin, Heidelberg, 2025, pp. 366–372.
URL: https://doi.org/10.1007/978-3-031-88720-8_56. doi:10.1007/978-3-031-88720-8_56. [18] M. Brennan, S. Afroz, R. Greenstadt, Adversarial stylometry: Circumventing authorship recognition to preserve privacy and anonymity, ACM Transactions on Information and System Security 15 (2012). URL: http://dx.doi.org/10.1145/2382448.2382450. doi:10.1145/2382448.2382450. [19] H. Wang, P. Juola, A. Riddell, Reproduction and replication of an adversarial stylometry experiment, arXiv [cs.CL] (2022). URL: http://arxiv.org/abs/2208.07395. arXiv:2208.07395. [20] K. Krishna, Y. Song, M. Karpinska, J. Wieting, M. Iyyer, Paraphrasing evades detectors of AIgenerated text, but retrieval is an efective defense, in: A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, S. Levine (Eds.), Advances in Neural Information Processing Systems 36 (NeurIPS 2023), volume 36, 2023, pp. 27469–27500. URL: https://proceedings.neurips.cc/paper_files/paper/2023/ hash/575c450013d0e99e4b0ecf82bd1afaa4-Abstract-Conference.html. [21] A. Hans, A. Schwarzschild, V. Cherepanova, H. Kazemi, A. Saha, M. Goldblum, J. Geiping, T. Goldstein, Spotting LLMs with Binoculars: Zero-shot detection of machine-generated text, in: Forty-first International Conference on Machine Learning, {ICML} 2024, Vienna, Austria, July 21-27, 2024, volume abs/2401.12070, OpenReview.net, 2024, pp. 17519–17537. URL: https://dl.acm.org/doi/10.5555/3692070.3692768. doi:10.48550/arXiv.2401.12070. [22] D. Sculley, C. E. Brodley, Compression and machine learning: A new perspective on feature space vectors, in: Data Compression Conference (DCC’06), IEEE, 2006, pp. 332–341. URL: https://ieeexplore.ieee.org/abstract/document/1607268?casa_token=EwORidpcgTwAAAAA: zONgvLu6aVgw-jrz0A-5JHXs-SdAljLqNXabhAQh6w685CRwAXXe7FxcD67SDkf6Ztfaj6AEwWA. doi:10.1109/dcc.2006.13. [23] O. Halvani, C. Winter, L. Graner, On the usefulness of compression models for authorship verification, in: Proceedings of the 12th International Conference on Availability, Reliability and Security, volume Part F1305, ACM, New York, NY, USA, 2017, pp. 54:1–54:10. URL: http: //dx.doi.org/10.1145/3098954.3104050. doi:10.1145/3098954.3104050. [24] D. Macko, mdok of KInIT: Robustly Fine-tuned LLM for Binary and Multiclass AI-Generated Text Detection, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [25] J. Liu, L. Kong, Z. Peng, F. Chen, Generative AI Authorship Verification Based on ContrastiveEnhanced Dual-Model Decision System, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [26] M. Seeliger, P. Styll, M. Staudinger, A. Hanbury, Human or Not? Light-Weight and Interpretable Detection of AI-Generated Text, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [27] A. Voznyuk, G. Gritsai, A. Grabovoy, Team Advacheck at PAN: Multitasking Does All the Magic, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [28] J. Yang, K. Yan, Genre-Aware Contrastive Learning for AI Text Detection: A RoBERTa-Based Approach, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [29] S. A. Zaidi, H. T. Ahmed, S. A. Akbar, Z. Shakeel, F. Alvi, A. Samad, Team Nexus Interrogators at PAN: Voight-Kampf Generative AI Detection, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [30] T.-G. Marchitan, C. Creanga, L. P. Dinu, Unibuc - NLP at “Voight-Kampf” Generative AI Detection PAN 2025, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [31] L. Teja, A. Yadagiri, P. Pakray, Team CNLP-NITS-PP at PAN: Advancing Generative AI Detection: Mixture of Experts with Transformer Models, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [32] M. Jimeno-Gonzalez, E. Martínez-Cámara, P. G. Noelia Fernandez, L. A. U. na López, Team SINAIINTA at PAN 2025: Uncovering machine generated text with linguistic features, in: Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [33] F. Völpel, O. Halvani, Adept: AI-Generated Text Detection Based on Phrasal Category N-Grams, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [34] J. K. Ochab, M. Matias, T. Boba, T. Walkowiak, StylOch at PAN: Gradient-boosted trees with frequency-based stylometric features, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [35] B. Ostrower, P. Doongare, M. T. Unnikrishnan, Binoculars, BART, and Adversaries: Multi-Faceted AI Text Detection for PAN 2025, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [36] A. R. Basani, P.-Y. Chen, DivEye at PAN 2025: Diversity Boosts AI-Generated Text Detection, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [37] S. Pudasaini, L. Miralles-Pechuán, D. Lillis, M. L. Salvador, Enhancing AI Text Detection with Frozen Pretrained Encoders and Ensemble Learning, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [38] Y. Sun, S. Afanaseva, K. Stowe, K. Patil, Bi-directional Cross-entropy loss and Stylometric Feature combined Classifier, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [39] J. Larson, Generative AI detection using simple Feature Selection and SVM, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [40] J. Huang, H. Cao, X. Lin, Z. Han, Application and Analysis of Roberta-base Model Fine Tuning Based on Data Enhancement in AI Text Detection, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [41] S. Titze, O. Halvani, LOG-AID: Logit-Based Statistical Features for AI Text Detection, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [42] R. Kumar, A. Trivedi, O. Varshney, Voight-Kampf AI Detection Sensitivity : IIITS@CLEF’25, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [43] Z. Liang, K. Sun, H. Cao, J. Luo, Z. Han, Research on Text Author Classification Based on ModernBERT and Gradient Loss Function, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [44] D. Macko, R. Moro, I. Srba, Increasing the robustness of the fine-tuned multilingual machinegenerated text detectors, arXiv preprint arXiv:2503.15128 (2025). [45] X. Hu, P.-Y. Chen, T.-Y. Ho, RADAR: Robust AI-text detection via adversarial learning, Neural Information Processing Systems abs/2307.03838 (2023) 15077–15095. URL: https://proceedings.neurips. cc/paper_files/paper/2023/hash/30e15e5941ae0cdab7ef58cc8d59a4ca-Abstract-Conference.html. doi:10.48550/arXiv.2307.03838. arXiv:2307.03838. [46] H. Guo, S. Cheng, X. Jin, Z. Zhang, K. Zhang, G. Tao, G. Shen, X. Zhang, Biscope: Ai-generated text detection by checking memorization of preceding tokens, Advances in Neural Information Processing Systems 37 (2024) 104065–104090. [47] O. Halvani, Constituent Treelib - A Lightweight Python Library for Constructing, Processing, and Visualizing Constituent Trees., 2024. URL: https://github.com/Halvani/constituent-treelib. doi:10.5281/zenodo.10951644. [48] A. Peñas, Á. Rodrigo, A Simple Measure to Assess Non-response, in: Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, 2011, pp. 1415–1424. URL: https://aclanthology.org/P11-1142.pdf. [49] J. Bevendorf, B. Stein, M. Hagen, M. Potthast, Generalizing unmasking for short texts, in: Proceedings of the 2019 Conference of the North, Association for Computational Linguistics, Stroudsburg, PA, USA, 2019, pp. 654–659. URL: https://aclanthology.org/N19-1068.pdf. doi:10. 18653/v1/n19-1068. [50] M. Kestemont, K. Luyckx, W. Daelemans, T. Crombez, Cross-Genre Authorship Verification Using Unmasking, English Studies 93 (2012) 340–356. URL: http://dx.doi.org/10.1080/0013838X.2012. 668793. doi:10.1080/0013838X.2012.668793. [51] Y. Wang, J. Mansurov, P. Ivanov, J. Su, A. Shelmanov, A. Tsvigun, C. Whitehouse, O. M. Afzal, T. Mahmoud, T. Sasaki, T. Arnold, A. F. Aji, N. Habash, I. Gurevych, P. Nakov, M4: Multi-generator, Multi-domain, and Multi-lingual Black-Box Machine-Generated Text Detection, in: Y. Graham, M. Purver (Eds.), Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), Association for Computational Linguistics, 2024, pp. 1369–1407. URL: https://aclanthology.org/2024.eacl-long.83.pdf. [52] J. Hutson, Human-ai collaboration in writing: A multidimensional framework for creative and intellectual authorship, International Journal of Changes in Education (2025). [53] Q. Zhang, C. Gao, D. Chen, Y. Huang, Y. Huang, Z. Sun, S. Zhang, W. Li, Z. Fu, Y. Wan, L. Sun, LLM-as-a-coauthor: Can mixed human-written and machine-generated text be detected?, in: K. Duh, H. Gomez, S. Bethard (Eds.), Findings of the Association for Computational Linguistics: NAACL 2024, Association for Computational Linguistics, Mexico City, Mexico, 2024, pp. 409–436. URL: https://aclanthology.org/2024.findings-naacl.29/. doi: 10.18653/v1/ 2024.findings-naacl.29. [54] M. Abassy, K. Elozeiri, A. Aziz, M. N. Ta, R. V. Tomar, B. Adhikari, S. E. D. Ahmed, Y. Wang, O. Mohammed Afzal, Z. Xie, J. Mansurov, E. Artemova, V. Mikhailov, R. Xing, J. Geng, H. Iqbal, Z. M. Mujahid, T. Mahmoud, A. Tsvigun, A. F. Aji, A. Shelmanov, N. Habash, I. Gurevych, P. Nakov, LLMDetectAIve: a tool for fine-grained machine-generated text detection, in: D. I. Hernandez Farias, T. Hope, M. Li (Eds.), Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, Association for Computational Linguistics, Miami, Florida, USA, 2024, pp. 336–343. URL: https://aclanthology.org/2024.emnlp-demo.35. [55] Y. Wang, J. Mansurov, P. Ivanov, J. Su, A. Shelmanov, A. Tsvigun, O. Mohammed Afzal, T. Mahmoud, G. Puccetti, T. Arnold, A. Aji, N. Habash, I. Gurevych, P. Nakov, M4GT-bench: Evaluation benchmark for black-box machine-generated text detection, in: L.-W. Ku, A. Martins, V. Srikumar (Eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Association for Computational Linguistics, Bangkok, Thailand, 2024, pp. 3964–3992. URL: https://aclanthology.org/2024.acl-long.218. doi:10.18653/v1/2024.acl-long. 218. [56] L. Dugan, D. Ippolito, A. Kirubarajan, S. Shi, C. Callison-Burch, Real or fake text?: Investigating human ability to detect boundaries between human-written and machine-generated text, 2022.
URL: https://arxiv.org/abs/2212.12672. arXiv:2212.12672. [57] L. Kushnareva, T. Gaintseva, G. Magai, S. Barannikov, D. Abulkhanov, K. Kuznetsov, E. Tulchinskii, I. Piontkovskaya, S. Nikolenko, Ai-generated text boundary detection with roft, 2024. URL: https: //arxiv.org/abs/2311.08349. arXiv:2311.08349. [58] Z. Zeng, S. Liu, L. Sha, Z. Li, K. Yang, S. Liu, D. Gašević, G. Chen, Detecting ai-generated sentences in human-ai collaborative hybrid texts: challenges, strategies, and insights, in: Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI ’24, 2024. URL: https://doi.org/10.24963/ijcai.2024/835. doi:10.24963/ijcai.2024/835. [59] Z. Zeng, L. Sha, Y. Li, K. Yang, D. Gašević, G. Chen, Towards automatic boundary detection for human-ai collaborative hybrid essay in education, 2023. URL: https://arxiv.org/abs/2307.12267. arXiv:2307.12267. [60] T. Chakrabarty, P. Laban, C.-S. Wu, Can ai writing be salvaged? mitigating idiosyncrasies and improving human-ai alignment in the writing process through edits, 2025. URL: https://arxiv.org/ abs/2409.14509. arXiv:2409.14509. [61] E. Artemova, J. S. Lucas, S. Venkatraman, J. Lee, S. Tilga, A. Uchendu, V. Mikhailov, Beemo: Benchmark of expert-edited machine-generated outputs, in: L. Chiruzzo, A. Ritter, L. Wang (Eds.), Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), Association for Computational Linguistics, Albuquerque, New Mexico, 2025, pp. 6992–7018. URL: https:// aclanthology.org/2025.naacl-long.357/. [62] S. Saha, S. Feizi, Almost ai, almost human: The challenge of detecting ai-polished writing, 2025.
URL: https://arxiv.org/abs/2502.15666. arXiv:2502.15666. [63] G. Bao, L. Rong, Y. Zhao, Q. Zhou, Y. Zhang, Decoupling content and expression: Two-dimensional detection of ai-generated text, arXiv preprint arXiv:2503.00258 (2025). [64] Z. Cheng, L. Zhou, F. Jiang, B. Wang, H. Li, Beyond binary: Towards fine-grained llm-generated text detection via role recognition and involvement measurement, arXiv preprint arXiv:2410.14259 (2024). [65] D. Mollá, Q. Xu, Z. Zeng, Z. Li, Overview of the 2024 ALTA shared task: Detect automatic AI-generated sentences for human-AI hybrid articles, in: T. Baldwin, S. J. Rodríguez Méndez, N. Kuo (Eds.), Proceedings of the 22nd Annual Workshop of the Australasian Language Technology Association, Association for Computational Linguistics, Canberra, Australia, 2024, pp. 197–202.
URL: https://aclanthology.org/2024.alta-1.17/. [66] M. Li, E. Hovy, J. Lau, Summarizing multiple documents with conversational structure for meta-review generation, in: H. Bouamor, J. Pino, K. Bali (Eds.), Findings of the Association for Computational Linguistics: EMNLP 2023, Association for Computational Linguistics, Singapore, 2023, pp. 7089–7112. URL: https://aclanthology.org/2023.findings-emnlp.472/. doi: 10.18653/v1/ 2023.findings-emnlp.472. [67] N. Dycke, I. Kuznetsov, I. Gurevych, NLPeer: A unified resource for the computational study of peer review, in: A. Rogers, J. Boyd-Graber, N. Okazaki (Eds.), Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Association for Computational Linguistics, Toronto, Canada, 2023, pp. 5049–5073. URL: https://aclanthology. org/2023.acl-long.277/. doi:10.18653/v1/2023.acl-long.277. [68] A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever, Language models are unsupervised multitask learners, 2019. URL: https://api.semanticscholar.org/CorpusID:160025533. [69] T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. M. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, D. Amodei, Language models are few-shot learners, 2020. URL: https://arxiv.org/abs/2005.14165. arXiv:2005.14165. [70] N. S. Keskar, B. McCann, L. R. Varshney, C. Xiong, R. Socher, CTRL: A conditional transformer language model for controllable generation, CoRR abs/1909.05858 (2019). URL: http://arxiv.org/ abs/1909.05858. arXiv:1909.05858. [71] X. Li, P. Yu, C. Zhou, T. Schick, O. Levy, L. Zettlemoyer, J. Weston, M. Lewis, Self-alignment with instruction backtranslation, arXiv preprint arXiv:2308.06259 (2023). [72] B. Li, H. Qi, K. Yan, Team Bohan Li at PAN: DeBERTa-v3 with R-Drop regularization for Human-AI Collaborative Text Classification, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [73] M. Zheng, Y. Zhong, F. Liu, T. Xian, M. Xie, W. Wu, Z. Zhang, Q. Sun, StarBERT: A Hybrid Neural Network Model for Human-AI Collaborative Text Classification, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [74] T. Xian, Y. Zhong, F. Liu, M. Xie, Q. Sun, M. Zheng, W. Wu, Z. Zhang, DBG: Human-AI Collaborative Text Classification with DeBERTa-enhanced Contextual and Geometric Attention, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [75] T. Li, Fine-Grained Human-AI Collaborative Text Classification Using DeBERTa, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [76] D. Ignatenko, K. Zaitsev, O. Shkriaba, ReText.Ai Team at PAN 2025: Applying a Multiple Classification Heads to a Transformer Model for Human-AI Collaborative Text Classification, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [77] Q. Sun, L. Ma, W. Yang, T. Xian, M. Xie, W. Wu, Z. Zhang, M. Zheng, DeBERTa-FPN: Fusion Feature Pyramid Network for Human-AI Collaborative Text Classification, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [78] W. Wu, W. Yang, Z. Zhang, M. Xie, M. Zheng, T. Xian, Q. Sun, Bert_T for Human-AI Collaborative Text Classification, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [79] Z. Zhang, W. Yang, W. Wu, M. Xie, M. Zheng, Q. Sun, T. Xian, DBA: A Hybrid Neural Network Model for Generative Human-AI Collaborative Text Classification, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [80] D. G. Sánchez, J. Jimenez, M. Ramírez, J. Martinez, RoBERT-IA: Human-AI Collaborative Text Classification, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025. [81] Y. Fuchuan, H. Cao, H. Zhongyuan, Sentence-Level AI-Generated Text Detection with Fine-Tuned BERT, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org, 2025.