1.1. Novelty and contributions

The novelty of this work is not the use of TF-IDF, of DeBERTa, or of Random Forest individually, all of which are established. It lies in three specific claims, each of which is tested empirically in Section 4. First, that lexical and contextual evidence for source attribution are complementary rather than redundant, and that fusing them at the feature level rather than at the decision level yields a gain far larger than either branch contributes alone: 92.38% against 58.08% for TF-IDF with Random Forest and 74.65% for a fine-tuned DeBERTa on identical data and splits. Second, that human-AI collaboration is separable into distinguishable workflow classes rather than being a single undifferentiated ‘hybrid’ bucket; the proposed model separates AI-generated-then-human-edited, AI-generated-then-AI-refined, human-written-then-AI-refined, and the three-stage combination, at per-class recalls between 91.8% and 92.4%. Third, that attribution-grade granularity is achievable within production latency and cost constraints, which we demonstrate by serving the model publicly rather than reporting offline numbers only. The specific contributions are:

• This study establishes a comprehensive framework for multi-class AI source attribution, moving beyond binary detection to enable precise identification of specific language models.
• A hybrid feature-fusion architecture (TF-IDF + DeBERTa + Random Forest) is proposed and shown, through a component-wise ablation, to derive its performance from the interaction of its branches rather than from either branch alone.
• The introduction of a 66,000-sample dataset spanning 22 distinct AI sources plus hybrid content provides an unprecedented resource for advancing forensic analysis of synthetic text, complemented by a 30,000-sample, 10-class corpus in which four human-AI collaboration workflows are explicitly modelled as separate classes. Both corpora are publicly released as the TXD-22 benchmark dataset under DOI 10.17632/prcjcggtjf.1 [14], so that the results reported here are directly reproducible and the corpus is available for use by other groups.
• A documented prompt-engineering protocol, a prompt-aware splitting strategy that prevents prompt leakage between training and test partitions, and representative dataset examples are provided so that the corpus construction is reproducible.
• Post-hoc explainability through SHAP and LIME is integrated into the evaluation, making the lexical basis of each attribution decision inspectable rather than opaque.
• A quantitative benchmark against four commercial detectors (QuillBot, Turnitin, ZeroGPT, TextGuard) across AI-generated, human-written and hybrid inputs demonstrates that no existing tool provides source attribution or hybrid recognition at all.
• The development of a fully functional three-tier web application bridges the critical gap between experimental detection models and practical, scalable tools accessible to educators, journalists, and the public.

This paper is organized as follows. Section 2 reviews related work in AI text detection, stylometry, and production ML systems, including recent 2024–2026 contributions. Section 3 details the methodology, covering dataset creation, feature extraction, the proposed hybrid architecture, the validation protocol, and the complete system architecture of SiBot. Section 4 presents experimental results, ablation, statistical validation, per-class analysis, efficiency measurements and system performance. Section 5 discusses the implications of our findings, the quantitative comparison against commercial detectors, and the ethical considerations raised by attribution technology. Section 6 concludes the paper and outlines directions for future work.

2.1. Binary AI text detection

The majority of existing research focuses on binary classification, attempting to answer whether a given text was written by a human or an AI. A seminal work by OpenAI [15] introduced a classifier trained to distinguish GPT-2 generated text from human-written text, achieving notable success on in-domain data but demonstrating significant limitations in generalization. Subsequent studies expanded on this foundation. Gehrmann et al. [16] developed GLTR (Giant Language Model Test Room), a tool that uses the statistical properties of language models to highlight passages that are likely machine-generated by visualizing the ranking of each word under a model’s predictive distribution. This approach provided interpretable results but was limited to the specific model it was calibrated against. Mitchell et al. [17] proposed DetectGPT, a zero-shot method for detecting machine-generated text that exploits the observation that generated text tends to occupy negative curvature regions of the model’s log-probability function. This method achieved high performance without requiring a separate training dataset but was computationally expensive. More recently, OpenAI discontinued its AI classifier due to its low accuracy, highlighting the inherent difficulty of the task [18]. Several commercial tools have emerged, including ZeroGPT [19] and QuillBot’s AI detector [20], which provide user-friendly interfaces but maintain proprietary, often non-transparent, detection methodologies. A common limitation across these binary tools is their inability to provide granular insights or to differentiate between outputs from distinct AI models.

2.2. Stylometry and linguistic fingerprints

The field of stylometry, which analyzes linguistic style to attribute authorship, provides a theoretical foundation for fine-grained AI source attribution [21]. Early work by Holmes [22] established that authors have unique, quantifiable stylistic features such as function word frequency, sentence length distribution, and vocabulary richness. These principles have been successfully applied to distinguish between human authors and have more recently been extended to AI models. Research by Liang et al. [23] investigated the concept of “neural fingerprints” in large language models, demonstrating that different models exhibit distinct stylistic patterns that can be captured through simple features like token distributions and n-gram frequencies. Similarly, Krishna et al. [24] explored paraphrasing strategies and showed that even when content is rewritten, certain stylistic signatures of the originating model can persist. Seraj et al. [25] applied stylometric analysis to distinguish between human and ChatGPT essays, achieving high accuracy using traditional linguistic features. Kumarage and Liu [26] made this explicit for the attribution setting, extracting sixty lexical, syntactic and structural stylometric features and showing that they carry enough signal to trace text back to its originating LLM across both proprietary and open-source model families. These studies collectively suggest that while the semantic content of different models may converge, their stylistic fingerprints remain distinct, providing a basis for multi-class classification.

2.3. Multi-class attribution and model fingerprinting

The move from binary detection to multi-class source attribution is a more recent but crucial development. Recent work has begun to explore the possibility of not just detecting AI-generated text, but identifying which specific AI model generated it. This is particularly important for forensic analysis and understanding model biases. Uchendu et al. [27] introduced the concept of “model attribution” and presented a methodology using n-gram features and supervised classifiers to identify the source model of a text, achieving strong results on a corpus of texts from a limited set of models. Fröhling and Zubiaga [28] extended this by focusing on the linguistic differences between GPT-2, GPT-3, and human-written text, demonstrating that transformer-based models like BERT could be used for attribution. Pagnoni et al. [29] introduced the “Ivy” dataset and explored fine-grained detection of text from multiple LLMs, finding that even models from the same family could be differentiated with reasonable accuracy. Ghosal et al. [30] took a different approach by focusing on “semantic drift,” analyzing how the content and style of generated text vary across different models when given identical prompts. Their findings align with our own, suggesting that stylistic features are more discriminative than purely semantic ones. These studies provide the critical foundation for SiBot, yet none have combined this research with a comprehensive, production-ready deployment.

2.4. Production ML systems for NLP

A significant gap in the literature exists between high-accuracy research models and their deployment as robust, scalable applications. This gap is often referred to as the “last mile” problem in machine learning engineering [31]. Several works have outlined best practices for deploying ML models, including model serialization, containerization with Docker, and the use of cloud platforms [32]. Huyen [33] provides a comprehensive framework for designing machine learning systems that are reliable, scalable, and maintainable, emphasizing the importance of monitoring, versioning, and infrastructure. Specific to NLP, frameworks like Hugging Face Transformers have democratized access to pre-trained models, while platforms like Hugging Face Spaces simplify deployment [34]. FastAPI has emerged as a popular choice for serving ML models due to its high performance and automatic OpenAPI documentation [35]. The use of Supabase (PostgreSQL) with Row Level Security for user management and authentication is a modern approach to building secure, data-driven applications without custom backend code [36]. SiBot leverages these modern engineering practices to deliver its research contributions as a public, accessible tool.

2.5. Recent advances (2024–2026)

The field has moved substantially since the foundational work reviewed above, and three developments in particular bear directly on the design decisions taken in this paper. The first is the maturation of robustness benchmarking. Dugan et al. [37] released RAID, comprising more than six million generations across eleven models, eight domains, eleven adversarial attacks and four decoding strategies, and demonstrated that both open-source and commercial detectors degrade sharply under adversarial paraphrasing, altered sampling strategies and unseen generators. He et al. [38] reached a compatible conclusion with MGTBench. The practical implication for the present work is that a detector evaluated only on in-domain data is evaluated optimistically, which motivates the prompt-aware splitting protocol described in Section 3.1.7 and the explicit generalisation analysis in Section 4.7.

The second development is the emergence of efficient zero-shot detection. Bao et al. [39] introduced Fast-DetectGPT, replacing the expensive perturbation step of DetectGPT with a conditional probability curvature statistic and achieving roughly two orders of magnitude speed-up, while Hans et al. [40] proposed Binoculars, which contrasts the perplexity of two closely related language models and detects over 90% of ChatGPT generations at a false-positive rate of 0.01% without any training data. These methods are strong binary detectors, but they are structurally unable to perform source attribution: the score they compute is a scalar, and it does not identify a generator. They therefore establish the state of the art for the question this paper argues is the wrong one, and reinforce the case for a multi-class formulation.

The third and most directly relevant development is the consolidation of attribution as a distinct task. Huang, Chen and Shu [41] surveyed authorship attribution in the LLM era and formalised precisely the four-way taxonomy this paper operationalises: attributing text to human authors, detecting LLM-generated text, identifying the specific LLM responsible, and classifying text as human-authored, machine-generated, or co-authored by both. La Cava and Tagarelli [42] released OpenTuringBench, an open-model benchmark explicitly targeting both detection and attribution, and Bisztray et al. [43] demonstrated the analogous problem for source code, showing that model-specific stylometric signatures survive in generated C programs. Work on mixed provenance has advanced in parallel: Chen et al. [44] framed the LLM-as-a-coauthor problem and showed that mixed human-machine text is markedly harder to detect than either pure category, and Thai et al. [45] introduced EditLens to quantify the extent of AI editing rather than merely its presence. Georgiou [46] showed that automatically extracted linguistic features remain competitive discriminators. Collectively, this recent literature validates the multi-class, hybrid-aware formulation adopted here while leaving the specific combination proposed in this paper (feature-level fusion of sparse lexical and dense contextual representations, evaluated across both a 22-source and a 10-class workflow-aware corpus and served in production) unaddressed.

2.6. Gaps in existing literature and SiBot’s position

Despite significant advancements in AI text detection and attribution, several critical gaps remain. First, most research focuses on a small number of AI models (often 2–5), limiting the practical applicability of findings. Second, there is a notable absence of datasets that include hybrid human-AI content, which is increasingly prevalent in real-world scenarios [47]. Third, the evaluation of detection methods is often done on in-domain data, with little attention to generalization across topics or domains. Fourth, and most importantly, there is a profound disconnect between academic models and production systems; researchers rarely deploy their work as functional, user-friendly tools [48]. A fifth gap, evident from the 2024–2026 literature reviewed above, is methodological: statistical and contextual representations are almost always evaluated in competition with one another rather than in combination, so the question of whether their errors are correlated has gone largely untested. SiBot directly addresses these gaps by introducing a dataset of 22 diverse text sources, including a hybrid class, and by deploying the best-performing model as a fully functional three-tier application, thereby demonstrating a complete research-to-production pipeline. Concretely, the present work differs from the closest prior studies in four respects: it evaluates 22 generators where Uchendu et al. [27] and Kumarage and Liu [26] evaluate between two and six; it treats four human-AI collaboration workflows as separate classes where Chen et al. [44] treat mixed authorship as a single category; it fuses lexical and contextual features rather than selecting between them, as all of [27,28,39] and [40] do; and it reports a live, publicly accessible deployment, which none of the cited attribution studies provide.

3.1. Phase 1: AI model development

This phase encompasses the end-to-end construction of the classification pipeline, from data collection to the serialization of the final production model. The workflow of phase 1, illustrated in Fig 1, consists of five sequential stages: dataset creation, preprocessing, feature extraction, classifier training and evaluation, and model artifact serialization [49]. Fig 1 has been redrawn at 600 dpi with enlarged typography in response to reviewer comments on figure legibility, and now also shows the second, 10-class experimental track and the fusion branch that were absent from the original pipeline.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Phase 1 Workflow: AI Model Development Pipeline.

https://doi.org/10.1371/journal.pone.0357405.g001

3.1.1. Dataset construction and prompt engineering.

The foundation of this study is a novel, purpose-built dataset. We identified 22 distinct AI sources, encompassing a wide range of prominent LLMs and AI assistants including ChatGPT, Grok, Pi, BLACKBOXAI, Nova, Poe, Mistral AI, Perplexity, Z.ai, Gemma, Llama, Claude, Qwen, DeepAI, DeepSeek, Gemini, and Copilot, with a full list maintained in the production system’s/classes endpoint. In addition to the AI sources, we defined a class to capture the complexity of real-world text: “AI-generated and Human-written,” acknowledging that text creation is often an iterative process involving both humans and machines [50]. For each of the 22 classes, we collected 3,000 text samples by posing a single fixed set of 3,000 unique questions or prompts to each AI source in turn, so that every class answers exactly the same questions. The total dataset size is 66,000 samples, ensuring a balanced representation for each class. A second corpus of 30,000 samples across 10 classes was constructed using the same prompt bank and the same protocol, and is used for the hybrid experiments reported in Sections 4.2 to 4.7.

Definition of diversity. The term “diverse”, used loosely in the original manuscript, is defined here along three explicit axes. Topical diversity means that the prompt bank spans thirty academic domains, each broken down into eight to ten sub-domains, covering engineering, the natural sciences, the social sciences, law, medicine, the humanities and the creative and media disciplines. The full distribution is given in Table 1. This breadth is deliberate and it is central to the validity of the attribution claim: because the identical prompt set is put to every source, every domain is represented in every class in equal proportion, so topic carries no class information and the classifier cannot reach its accuracy by learning subject matter instead of style. Task diversity means that within each domain the prompts span five task types that elicit structurally different responses: factual explanation, comparative analysis, procedural or how-to description, opinion or argumentative writing, and summarisation or condensation. Representative examples of each, drawn verbatim from the bank, are: “What is the half-life of a drug and why is it clinically important?” (factual explanation); “How does sociology differ from other social sciences?” (comparative analysis); “Describe the process of market segmentation and how firms apply it.” (procedural description); “How does climate change affect rural communities, and what should be prioritised in response?” (opinion or argumentative); and “Summarise the central claim of utilitarianism as an ethical theory.” (summarisation). Structural diversity means that prompt length and specificity were varied deliberately, from single-clause questions of under ten words to multi-constraint instructions exceeding fifty words, so that the corpus is not dominated by a single interrogative register. The complete bank comprises 3,000 unique prompts, and the same 3,000 prompts were put verbatim to every generative source and used to elicit every human-written and hybrid sample, so that the corpora differ across classes only in what produced the response, never in what was asked.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Distribution of prompts across the thirty academic domains and their sub-domains used in corpus construction. The same 3,000 prompts were put verbatim to every source.

https://doi.org/10.1371/journal.pone.0357405.t001

Prompt engineering protocol. Prompts were authored and reviewed in four steps. A seed set was drafted per domain and task type; each seed was checked against the others for near-duplication using cosine similarity over TF-IDF vectors, with a 0.85 threshold for rejection; the surviving prompts were then reviewed manually for ambiguity and for wording that might bias a particular model family, such as vendor names or product-specific terminology, which were removed; and finally the same frozen prompt was issued verbatim to every source. Issuing an identical prompt to every source is essential to the validity of the attribution claim: it ensures that measured differences between classes are attributable to generator behaviour rather than to input variation. All generations used each provider’s default decoding settings and no system prompt, so that the captured fingerprint reflects the model as an ordinary user would encounter it. Table 1 gives verbatim examples of the prompts used at each level of the design.

Class structure and hybrid content generation. The 10-class corpus comprises five pure AI sources (ChatGPT, Copilot, Gemini, Qwen, Claude), one purely human-written class, and four hybrid classes constructed by a structured multi-stage protocol: AI-generated text subsequently edited or extended by a human; AI-generated text subsequently refined by a second AI pass; human-written text subsequently polished by an AI system; and a three-stage class combining AI generation, AI refinement and final human modification. This design mirrors how documents are actually produced, which is rarely in a single step. Human-written responses were collected from manually curated sources including academic-style writing, student responses and manually authored answers to the same prompts, and every human sample was verified to have been produced without AI assistance.

Quality control and bias mitigation. A standardised pipeline removed duplicate and near-duplicate entries using similarity thresholding, discarded incomplete or truncated generations and refusal responses, and normalised spacing, punctuation and character encoding. Every sample was re-checked for correct class labelling. Two specific sources of prompt-induced bias were addressed. Because the same prompts were issued to every source, any topical signal is shared across classes and therefore cannot by itself drive attribution; and because prompts were screened to remove vendor-specific phrasing, no class receives a lexical cue from the prompt itself. Class balance was enforced by construction at 3,000 samples per class in both corpora, which removes the majority-class bias that would otherwise complicate interpretation of the accuracy figures.

Data leakage prevention. Splits were constructed prompt-aware: all responses derived from a given prompt are assigned to the same partition, so no prompt appears in both training and test data. Without this precaution a classifier can learn prompt-specific content and report inflated accuracy that does not reflect attribution ability. This protocol applies to every result reported in Section 4.

Table 2 presents representative samples covering every class for readers who wish to inspect the structure of the corpus without downloading it, and the complete benchmark texts used in the detector comparison are reproduced verbatim in Table 12.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Representative samples from the TXD-22 corpus, one per class, each drawn from a different academic domain of Table 1. Responses are truncated to their first sentence; the complete texts are available in the released dataset [14].

https://doi.org/10.1371/journal.pone.0357405.t002

The samples in Table 2 are drawn deliberately from ten different academic domains of Table 1 rather than from a single field, so that the topical breadth of the corpus is visible directly. They also illustrate the variability the classifier must resolve. AI-generated responses exhibit relatively consistent phrasing and structured discourse patterns; human-written content shows higher lexical diversity and stylistic variation; and the hybrid categories display overlapping characteristics inherited from both, which is precisely what makes them the hardest classes and what motivates a multi-level feature representation.

3.1.3. Feature extraction.

To comprehensively capture both the surface-level lexical characteristics and the deep contextual semantics of the texts, we employed complementary feature extraction techniques. This multi-faceted approach allowed us to represent each document along statistical and contextual dimensions, providing a rich input representation for downstream classification [52].

Term Frequency-Inverse Document Frequency (TF-IDF). As a robust baseline for capturing stylistic and lexical patterns, we applied TF-IDF vectorization to transform each raw text into a high-dimensional numerical matrix [53]. TF-IDF weights each term by its frequency in a document while discounting terms that appear frequently across the entire corpus, thereby emphasizing words that are distinctive to a given text. We used unigrams and bigrams as the basic features and restricted the vocabulary to the top 10,000 n-grams for the 22-class corpus and 5,000 for the 10-class corpus. The choice of TF-IDF requires justification, since more recent representations are indeed available. The justification is task-specific rather than generic. Source attribution is a stylometric problem, not a semantic one: when the same prompt is issued to every generator, the semantic content of the responses converges by construction, and the discriminative signal survives almost entirely in how the content is expressed: function-word frequencies, preferred discourse markers, punctuation habits and characteristic bigrams. These are exactly the quantities a sparse n-gram representation encodes explicitly and that a dense contextual encoder is trained to abstract away. This is not a conjecture: Table 3 shows TF-IDF outperforming both ALBERT and RoBERTa embeddings by 13–15 accuracy points on the identical 22-class data. TF-IDF is additionally interpretable at the token level, which is what makes the SHAP and LIME analyses of Section 4.6 possible, and it is inexpensive to serve. We therefore retain TF-IDF not as a concession to simplicity but as a carrier of the specific evidence the task depends on. Critically, we do not rely on it alone: Section 3.1.5 shows that combining it with a contextual encoder is worth a further 34 accuracy points over TF-IDF by itself.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Performance of feature extractor and classifier combinations on the 22-class task (66,000 samples).

https://doi.org/10.1371/journal.pone.0357405.t003

ALBERT (A Lite BERT). To move beyond surface-level statistics and capture sentence-level semantics with high efficiency, we employed ALBERT, a transformer-based model designed to reduce memory consumption and training time while maintaining strong representational power [54]. ALBERT achieves parameter reduction through factorized embedding parameterization and cross-layer parameter sharing. For each input text, we fed the tokenized sequence into the pre-trained ALBERT model and extracted the final hidden state corresponding to the [CLS] token, providing a fixed-dimensional vector of 768 features per document [55].

RoBERTa (Robustly Optimized BERT Approach) Complementing ALBERT, we utilized RoBERTa, a transformer model that builds upon BERT’s architecture with a more extensive training regimen and optimized design choices. RoBERTa was pre-trained on a significantly larger corpus with dynamic masking, larger batch sizes, and longer training durations [56]. As with ALBERT, we processed each text through the RoBERTa model and extracted the [CLS] token embedding from the last hidden layer, yielding a 768-dimensional vector [57].

DeBERTa (Decoding-enhanced BERT with Disentangled Attention). For the hybrid architecture we additionally employ DeBERTa, which improves on BERT and RoBERTa through disentangled attention, representing each token by separate content and position vectors and computing attention across both, and an enhanced mask decoder. This matters for the present task because attribution evidence is frequently positional as well as lexical: characteristic discourse patterns such as where a model places its hedges, or how consistently it opens with a definitional clause, are relational properties that disentangled attention represents more faithfully than standard self-attention. Unlike ALBERT and RoBERTa, which were used as frozen feature extractors in the 22-class experiments, DeBERTa is fine-tuned on the classification objective before its representations are extracted, so that its weights are optimised to discriminate between sources rather than to reconstruct general language.

3.2. Phase 2: Production system architecture (SiBot)

This phase describes the deployment-ready system designed to serve the trained AI model in a production environment. The architecture follows a modern three-tier cloud model comprising a presentation layer, an AI inference layer, and a data and authentication layer [63]. The complete workflow, from user interaction to prediction persistence, is illustrated in Fig 2. Fig 2 has been redrawn at 600 dpi with enlarged typography and an explicit numbered request lifecycle, in response to reviewer comments on figure legibility and on the need for fuller illustration of the architecture.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Phase 2 Architecture: Production System Workflow (SiBot).

https://doi.org/10.1371/journal.pone.0357405.g002

4.1. Model performance evaluation on the 22-class task

The experimental results for the 12 model combinations on the 66,000-sample, 22-class corpus are summarized in Table 3. The key findings are discussed below.

4.2. Baseline evaluation on the 10-class task

The 22-class results establish an upper bound on what a single representation can achieve. To determine whether that ceiling is intrinsic to the task or a property of the representation, we conducted a second, more thorough study on the 30,000-sample, 10-class corpus, in which every major model family is trained under identical conditions and the prompt-aware 10-fold protocol of Section 3.1.7. This directly addresses the concern that deep learning was under-explored in the original manuscript.

Table 4 reports the classical machine-learning models using TF-IDF features. KNN achieved the best overall performance with an accuracy of 72.66% and the highest F1-score of 74.34%, indicating a strong balance between precision and recall, and SVM was close behind at 71.83%. XGB produced moderate results at 67.38%. Random Forest with TF-IDF alone achieved only 58.08% accuracy, and the Decision Tree was weakest at 38.65%. This last figure is important for interpreting the ablation in Section 4.3: TF-IDF with Random Forest, which is the single strongest configuration on the 22-class task, is not the strongest classical configuration on the 10-class task, which means the hybrid’s advantage cannot be an artefact of having started from an unusually strong baseline.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Performance of traditional ML models with TF-IDF on the 10-class task.

https://doi.org/10.1371/journal.pone.0357405.t004

Table 5 reports the deep-learning architectures. BiLSTM with attention achieved the highest performance at 60.66% accuracy, indicating that the attention mechanism improves the model’s ability to focus on informative regions of the text. CNN followed at 59.05%, capturing local textual patterns through convolution, while LSTM and BiLSTM achieved 58.01% and 57.93% respectively and GRU was weakest at 56.81%. The notable observation is that the recurrent and convolutional models as a family underperform the classical TF-IDF models in Table 4, which is consistent with the argument advanced in Section 3.1.3 that attribution evidence resides substantially in surface lexical statistics that sequence models trained from scratch on 30,000 samples do not have enough data to rediscover.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Performance of deep-learning models on the 10-class task.

https://doi.org/10.1371/journal.pone.0357405.t005

Table 6 reports the fine-tuned transformer models, which constitute the strongest single-model benchmark against which the proposed method must be judged. DeBERTa achieved the best overall performance with 74.65% accuracy, 75.58% precision, 74.56% recall and 74.33% F1, reflecting the strength of disentangled attention for modelling contextual relationships. RoBERTa was closely competitive at 74.37% accuracy and 73.99% F1. ALBERT reached 72.65% with balanced precision and recall while maintaining a more efficient architecture, and BERT and DistilBERT produced 69.88% and 68.97% respectively. These figures confirm the reviewers’ expectation that fine-tuned transformers outperform classical models on this task, and they establish 74.65% as the benchmark that the proposed hybrid must exceed for its contribution to be meaningful.

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. Performance of fine-tuned transformer models on the 10-class task.

https://doi.org/10.1371/journal.pone.0357405.t006

4.3. Proposed hybrid model and ablation study

Table 7 presents the ablation study, which isolates the contribution of each component of the proposed framework. The design compares the proposed configuration against its own constituent branches evaluated alone, and against two alternative fusion configurations that combine different families, all on identical data and splits.

Download:

• PNG
  larger image
• TIFF
  original image

Table 7. Ablation study of the proposed hybrid ensemble on the 10-class task.

https://doi.org/10.1371/journal.pone.0357405.t007

Reading the table row by row: TF-IDF with Random Forest, the lexical branch operating alone, achieves 58.08%. Fine-tuned DeBERTa, the contextual branch operating alone, achieves 74.65%. An alternative fusion using SVM with BiLSTM and BERT reaches only 67.82%, and XGB with BiLSTM and DeBERTa reaches 68.56%, both below the standalone DeBERTa figure, which demonstrates that fusion is not automatically beneficial and that the specific choice of branches and classifier matters. The proposed TF-IDF + DeBERTa + Random Forest configuration achieves 92.38% accuracy, 92.41% precision, 92.38% recall and 92.31% F1. The gain over the better of its own two branches is 17.73 points, and over its weaker branch 34.30 points. Since neither branch approaches this level alone, and since two other fusion configurations fail to exceed even the standalone transformer, the performance is attributable to the specific interaction of sparse lexical features with fine-tuned contextual features under an ensemble classifier, not to any single component. Fig 3 visualises the ablation.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Ablation of the hybrid framework on the 10-class dataset.

Neither branch approaches the fused configuration alone, and two alternative fusion designs fall below the standalone transformer baseline.

https://doi.org/10.1371/journal.pone.0357405.g003

4.4. Per-class analysis, confusion matrix and ROC

Aggregate accuracy conceals which classes a model confuses, which for an attribution system is the question that determines whether a verdict can be trusted. Table 8 therefore reports precision, recall, F1 and AUC for each of the ten classes of the deployed hybrid model. Precision, recall and F1 are computed directly from the confusion matrix in Fig 5; AUC values are the per-class one-versus-rest figures from the ROC analysis in Fig 6.

Download:

• PNG
  larger image
• TIFF
  original image

Table 8. Per-class performance of the deployed TF-IDF + DeBERTa + RF hybrid model.

https://doi.org/10.1371/journal.pone.0357405.t008

Two observations follow. First, the closeness of macro and weighted averages, both 92.32%, confirms that the model is not biased toward any subset of classes, which the balanced corpus design was intended to ensure. Second, the AUC column reveals a pattern that accuracy alone conceals: the classes with the lowest separability are class 6 (AI-generated and human-written, AUC 0.80), class 9 (the three-stage hybrid, 0.83) and class 0 (ChatGPT, 0.86). All three are cases in which the decision boundary is genuinely blurred: class 6 and class 9 contain human editing applied to AI drafts, so they share surface statistics with both parents, while ChatGPT output appears as the substrate of several hybrid classes. Conversely, classes 7 and 8, both of which involve an AI refinement pass, reach 0.99, indicating that AI refinement leaves a strong and consistent signature. Class 8 also shows the lowest precision at 90.89%, absorbing a small but systematic share of misclassifications from every other class, which is the expected behaviour for a class defined by light-touch AI polishing of human text. Fig 4 plots per-class recall against AUC.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Per-class recall and AUC of the deployed hybrid model.

The hybrid classes involving human editing (6 and 9) are the least separable; those involving an AI refinement pass (7 and 8) are the most.

https://doi.org/10.1371/journal.pone.0357405.g004

The confusion matrix in Fig 5 shows strong diagonal dominance, with 924 correct predictions out of 1,000 for classes 0–7, 918 for class 8 and 922 for class 9. Off-diagonal entries remain small and evenly distributed, mostly between 8 and 12, indicating that the residual errors are diffuse rather than concentrated in a particular confusable pair. The slight asymmetry in the last two rows and columns, with class 8 attracting ten misclassifications from each other class and class 9 losing twelve instances to class 8, is the only systematic pattern, and it corresponds to the intuitively hardest distinction in the corpus: separating text that was AI-refined after human authorship from text that passed through all three stages.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Confusion matrix of the TF-IDF + DeBERTa + RF hybrid model on the 10-class task (overall accuracy 92.32%).

https://doi.org/10.1371/journal.pone.0357405.g005

Fig 6 presents the one-versus-rest ROC curves. All ten curves lie well above the diagonal, with AUC values ranging from 0.80 for class 6 to 0.99 for classes 7 and 8 and a macro average of 0.912. Each curve maintains a steep initial ascent toward the top-left corner, indicating that high true-positive rates are achievable at low false-positive rates, the operating regime that matters for academic-integrity applications, where the cost of a false accusation is disproportionately high.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Multi-class ROC curves with per-class AUC scores for the TF-IDF + DeBERTa + RF hybrid model.

https://doi.org/10.1371/journal.pone.0357405.g006

4.5. Statistical validation

To establish that the differences reported above are not attributable to sampling variation, Table 9 gives 95% confidence intervals for the principal configurations, computed by the normal approximation to the binomial over the evaluated test instances as described in Section 3.1.7.

Download:

• PNG
  larger image
• TIFF
  original image

Table 9. Accuracy with 95% confidence intervals for the principal configurations.

https://doi.org/10.1371/journal.pone.0357405.t009

The confidence interval of the proposed model, [91.86, 92.90], does not overlap with that of the next-best configuration, DeBERTa at [73.80, 75.50], nor with any other interval in the table. The separation between the proposed model and the strongest baseline is more than sixteen percentage points at the nearest interval endpoints, so the difference is significant at the 5% level by a wide margin and would remain so under considerably more conservative correction for multiple comparisons. The same test also shows that the differences among the transformer baselines themselves are not all significant: the intervals for DeBERTa and RoBERTa overlap substantially, so their 0.28-point difference should not be interpreted as a meaningful ranking. Reporting intervals rather than point estimates alone makes this distinction visible, and we note it explicitly to avoid over-claiming. All results were obtained under prompt-aware 10-fold cross-validation, so each figure is an average over ten disjoint evaluations rather than a single fortunate split.

4.6. Explainability analysis

Fig 7 presents the global SHAP feature importance of the deployed hybrid model, showing the mean absolute SHAP value for the most influential terms across all classes. Terms such as “things”, “people”, “frequently”, “order” and “important” carry the highest contributions. The multi-coloured stacked bars indicate that each feature contributes differently to different classes, reflecting both shared and class-specific linguistic characteristics. The composition of this list is itself informative: the highest-weighted features are function words and generic discourse items rather than topic words, which is direct evidence for the central claim of Section 3.1.3 that the attribution signal is stylistic rather than semantic. A model exploiting topical leakage would instead show domain-specific nouns at the top of this ranking.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Global SHAP feature importance, showing the most influential terms across all classes in the TF-IDF + DeBERTa + RF hybrid model.

https://doi.org/10.1371/journal.pone.0357405.g007

Fig 8 illustrates a local explanation for an individual sample. The left panel shows the predicted probability distribution across classes, in which class 4 receives the highest probability followed by classes 2, 6 and 3. The middle panel gives the SHAP contribution of individual words, indicating how terms such as “example”, “fundamentally”, “creating” and “way” pushed the model toward the predicted class, and the right panel maps those terms back onto the original passage. This is the level of explanation a user contesting a verdict requires: not a score, but the specific tokens that produced it.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. Local SHAP explanation showing class probabilities and word-level contributions for an individual sample.

https://doi.org/10.1371/journal.pone.0357405.g008

4.7. Generalization analysis

Generalization was addressed at three levels, and we state plainly which of them the present evidence covers and which it does not. At the level of unseen prompts, every figure reported above was obtained under prompt-aware partitioning, so no prompt in any test fold appears in the corresponding training fold. The 92.38% figure is therefore already an unseen-prompt result, not an in-sample one; had prompt-specific content been driving the classification, performance would have collapsed under this constraint rather than exceeding every baseline trained under the same constraint.

At the level of topical and task variation, the corpus spans six domains and five task types by construction, and each cross-validation fold therefore contains material from every domain in both partitions. This establishes robustness across the covered range but does not establish transfer to domains outside it: creative fiction, legal drafting and clinical notes are absent, and we do not claim coverage of them. At the level of unseen generators and temporal drift, the honest position is that this remains open. The corpus captures a fixed set of models at a fixed point in time, and the recent robustness literature reviewed in Section 2.5, particularly RAID [37], demonstrates that detectors degrade materially against generators not represented in training. We have not evaluated against models released after corpus construction, and the accuracy reported here should not be assumed to hold for them. Section 6 sets out the cross-domain and cross-generator evaluation program that follows from this, and Section 5.7 records it as the principal limitation of the present work.

4.8. Computational efficiency and deployment cost

The original manuscript asserted that lightweight feature-based approaches remain competitive with deep learning for production deployment while supporting that claim only with a discussion of framework efficiency. Table 10 replaces that assertion with measurements, and the conclusion is more nuanced than the original claim. Model size for the classical and hybrid configurations is the serialised artefact on disk; for the transformer baselines it is the fp32 parameter footprint. Latency figures are per-request medians for texts under 500 words on the deployment instance (2 vCPU, 16 GB RAM, no GPU).

Download:

• PNG
  larger image
• TIFF
  original image

Table 10. Accuracy against computational cost for the principal configurations.

https://doi.org/10.1371/journal.pone.0357405.t010

Two conclusions follow, and one correction to the original manuscript. The correction is that the blanket claim that lightweight approaches “remain competitive” with deep learning is not supported by these data and has been withdrawn from the abstract, the introduction and the conclusion. On the 10-class task, TF-IDF with Random Forest reaches 58.08% against 74.65% for a fine-tuned transformer; that is not competitive, and asserting otherwise would misrepresent the evidence. The defensible claims are narrower and are what the revised manuscript now states. First, on the 22-class task, sparse lexical features outperform frozen transformer embeddings by a wide margin at roughly a third of the artefact size, which is a genuine and useful result about frozen-extractor pipelines specifically. Second, the deployed hybrid attains the highest accuracy of any configuration evaluated while remaining servable on CPU-only infrastructure at roughly half the footprint of a fine-tuned DeBERTa, so the efficiency argument survives, but as a property of the hybrid, not of the lightweight branch alone. Fig 9 plots this trade-off.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. Accuracy versus deployment cost.

The deployed hybrid occupies the favourable region: highest accuracy at moderate artefact size with CPU-only inference.

https://doi.org/10.1371/journal.pone.0357405.g009

4.9. Comparison with existing published methods

Table 11 positions the proposed framework against representative published studies in AI-generated text detection, and Fig 10 presents the same comparison graphically as requested. Most prior work addresses binary classification on smaller corpora and achieves accuracies between 66% and 75%; none incorporates explainability, and none reports a public deployment. The proposed hybrid achieves 92.38% on a larger primary dataset under a multi-class formulation with ten categories.

Download:

• PNG
  larger image
• TIFF
  original image

Table 11. Comparative analysis of existing AI-generated text detection studies and the proposed framework.

https://doi.org/10.1371/journal.pone.0357405.t011

Download:

• PNG
  larger image
• TIFF
  original image

Fig 10. Graphical comparison of accuracy across existing detection studies and the proposed SiBot models.

https://doi.org/10.1371/journal.pone.0357405.g010

Two caveats should be attached to Table 11 and Fig 10, and we state them rather than leaving the comparison to be read uncritically. The compared studies use different datasets, so the accuracies are not measured on a common benchmark and the comparison is indicative rather than a controlled head-to-head. And the compared studies address binary classification, which is an easier problem than 10-way or 22-way attribution; a binary accuracy of 74% and a 10-class accuracy of 92% are not directly commensurable, since the random baselines differ by a factor of five. The comparison is therefore offered to situate the work within the literature, not as proof of superiority on a shared task. A controlled evaluation on a common public benchmark such as RAID [37] or OpenTuringBench [42] is identified in Section 6 as necessary future work.

4.10. System performance and scalability

The production system was load-tested to evaluate its real-world responsiveness and scalability. Average inference time for the/predict endpoint is approximately 150–200 milliseconds for texts under 500 words on the Hugging Face Spaces instance. FastAPI’s asynchronous capabilities enable the system to handle multiple concurrent requests efficiently, and the Random Forest model’s predict_proba method is thread-safe, ensuring stable performance under moderate load. The frontend, built with React and Vite, achieves a Lighthouse performance score exceeding 95, thanks to efficient bundle splitting, lazy loading of admin routes, and optimized asset delivery. On the data layer, Supabase’s managed PostgreSQL scales horizontally, and Row Level Security policies are optimized with appropriate indexes on foreign keys and frequently queried columns such as analysis_logs.user_id and analysis_logs.created_at.

4.11. SiBot web application interface

To demonstrate the practical usability of the system, we present screenshots of the live SiBot interface. Figs 11–14 replace the interface screenshots of the earlier version and show the current deployment, which serves the 10-class hybrid model. Fig 11 presents the overall interface in its initial state. The left-hand panel hosts the text input area together with the full-text, per-sentence and batch analysis modes, while the right-hand panel prompts the user to begin analysis and indicates the supported provenance classes, with quick-reference chips for key providers alongside the human category.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 11. SiBot interface in its initial state.

https://doi.org/10.1371/journal.pone.0357405.g011

Download:

• PNG
  larger image
• TIFF
  original image

Fig 12. SiBot text input interface.

https://doi.org/10.1371/journal.pone.0357405.g012

Download:

• PNG
  larger image
• TIFF
  original image

Fig 13. Output view of SiBot, showing verdict, component breakdown and ranked source probabilities.

https://doi.org/10.1371/journal.pone.0357405.g013

Download:

• PNG
  larger image
• TIFF
  original image

Fig 14. Interface outputs for the human-written induction-motor text.

(A) SiBot; (B) ZeroGPT; (C) QuillBot; (D) Turnitin; (E) TextGuard.

https://doi.org/10.1371/journal.pone.0357405.g014

Fig 12 provides a closer view of the text input interface, where the user selects the analysis mode and supplies text by typing or pasting directly, by uploading a file, or by loading a built-in sample, which keeps the submission workflow straightforward for non-technical users.

Fig 13 illustrates the output view once a text has been analysed. The interface reports an overall verdict banner with the top-matched source and its confidence, summary cards for the AI-generated, human-written and hybrid components, a distribution chart, and a ranked list of source probabilities. In the example shown, an AI-generated passage is attributed to ChatGPT at 85.1% confidence, accompanied by an 11.1% hybrid component and a marginal 3.8% human-written probability. This level of granularity goes far beyond binary AI-versus-human detectors. All interface components are powered by the backend API described in Section 3.2 and are designed to provide both interpretability and actionable insight for users ranging from individual writers to institutional administrators.

5.1. The “Semantic Relation” and AI Fingerprints

The results of this study provide strong evidence for the existence of distinct, model-specific fingerprints in AI-generated text. These fingerprints are not necessarily semantic but are deeply stylometric. The success of TF-IDF suggests that LLMs have subtle but measurable preferences for certain lexical items and syntactic constructions. For example, one model might consistently use more formal transition words (“furthermore,” “consequently”), while another might favor a more conversational tone with contractions and simpler sentence structures. The global SHAP analysis in Section 4.6 corroborates this directly: the features carrying the greatest attribution weight are function words and generic discourse markers, not topic-bearing nouns. Since the same prompts were issued to every generator, topical content is shared across classes by construction and cannot carry the signal, so the fact that the model’s most-weighted evidence is stylistic is a consequence of the experimental design rather than a coincidence.

5.2. Why feature fusion outperforms either representation alone

The ablation in Section 4.3 poses a question it does not by itself answer: why should combining two representations produce a 17.73-point gain over the better of them, when the standard expectation for ensembling correlated predictors is a gain of a few points at most? The explanation we advance is that the two representations fail on different instances rather than the same ones. TF-IDF resolves cases where a generator has a distinctive lexical habit but loses cases where two generators share vocabulary and differ only in how they organise it. DeBERTa resolves the latter but, because contextual encoders are trained to abstract away surface form in favour of meaning, systematically discards precisely the lexical idiosyncrasies that separate generators answering the same prompt. Their errors are therefore substantially decorrelated, and a classifier with access to both can recover instances that neither could recover alone.

Two pieces of evidence support this reading over the alternative explanation that the gain simply reflects a larger feature space. First, if dimensionality alone were responsible, the two alternative fusion configurations in Table 7 would also improve on their components, and they do not: SVM with BiLSTM and BERT reaches 67.82% and XGB with BiLSTM and DeBERTa 68.56%, both below the 74.65% of standalone DeBERTa. Adding features can evidently hurt. Second, the fusion succeeds specifically when the contextual branch is fine-tuned rather than frozen; the frozen-extractor results in Table 3 show transformer embeddings performing worse than TF-IDF alone, which is what one expects if the frozen representation has already discarded the attribution-relevant surface information before fusion can exploit it.

5.3. The hybrid classes

The inclusion of hybrid classes adds a layer of ecological validity to our study. The 10-class corpus separates four distinct collaboration workflows rather than treating mixed authorship as one undifferentiated category, and the per-class results in Table 8 show that these workflows are genuinely separable, at recalls between 91.8% and 92.4%. The pattern of separability is itself informative. Classes involving an AI refinement pass (7 and 8) reach AUC 0.99, indicating that AI refinement imposes a strong and consistent signature: it regularises sentence length and normalises discourse markers in a way that is highly detectable. Classes involving human editing of AI drafts (6 and 9) are markedly harder, at 0.80 and 0.83, which is the expected result: human editing is idiosyncratic and partial, so it degrades the AI signature without imposing a consistent replacement. This asymmetry has a practical implication for users. A SiBot verdict of “AI-refined” is considerably more reliable than a verdict distinguishing lightly human-edited AI text from the three-stage workflow, and the interface communicates the full probability distribution precisely so that this distinction remains visible rather than being collapsed into a single label [79].

5.4. Beyond binary detection: Quantitative comparison with commercial detectors

A key contribution of this work is a production-ready interface that fundamentally differs from existing tools like QuillBot and ZeroGPT [80,81]. SiBot’s interface does not merely provide a probability of a text being AI-generated. Instead, it performs a multi-class prediction, outputting the most probable source model and an explicit hybrid decomposition. The earlier version of this manuscript supported this claim with a single qualitative screenshot comparison, which the reviewers rightly identified as insufficiently rigorous. We therefore replace it with a controlled benchmark. Three versions of a response to one generating question, “What is slip in an induction motor?”, were prepared: a purely AI-generated response produced by Gemini, a purely human-written response, and a hybrid response in which human-written content was subsequently refined by AI. The provenance of each is known by construction. All three were submitted to five systems: QuillBot, Turnitin, ZeroGPT, TextGuard and SiBot. For every tool, confidence scores were recorded across four analysis dimensions: probability that the text is AI-generated, identification of the specific generating model, probability that the text is human-written, and probability of mixed origin. Table 12 gives the benchmark material and Tables 13–15 the results, in which an “x” indicates that the tool does not offer that analysis dimension at all.

Download:

• PNG
  larger image
• TIFF
  original image

Table 12. Benchmark generating question and the three response variants.

https://doi.org/10.1371/journal.pone.0357405.t012

Download:

• PNG
  larger image
• TIFF
  original image

Table 13. Detection confidence comparison for the AI-generated response (true origin: Gemini).

https://doi.org/10.1371/journal.pone.0357405.t013

Download:

• PNG
  larger image
• TIFF
  original image

Table 14. Detection confidence comparison for the human-written response (true origin: human).

https://doi.org/10.1371/journal.pone.0357405.t014

Download:

• PNG
  larger image
• TIFF
  original image

Table 15. Detection confidence comparison for the hybrid response (true origin: human written, AI refined).

https://doi.org/10.1371/journal.pone.0357405.t015

The results warrant close reading. On the AI-generated response (Table 13), QuillBot and ZeroGPT both returned 100% AI confidence and Turnitin 49%, while TextGuard split its judgment between 65% AI and 35% human; none could attribute the text to its generating model, as the “x” entries record. SiBot assigned 96.7% AI confidence and was the only system to identify Gemini correctly. On the human-written response (Table 14), QuillBot and Turnitin correctly reported 0% AI, ZeroGPT produced a mild false-positive signal of 19.2%, and TextGuard failed outright, labelling a genuinely human-written passage as 60% AI, a false positive that in an academic-integrity setting could have serious consequences for a student. SiBot returned 94.7% human confidence with 0% AI and a 5.3% residual hybrid component.

The hybrid response (Table 15) is the most discriminative case and the one that most clearly exposes the limitation of the binary paradigm. QuillBot classified the text as 100% human, Turnitin reported 0% AI, and ZeroGPT reported only 5.5% AI: all three effectively overlooked the AI involvement entirely. TextGuard overcorrected in the opposite direction, labelling the same text 76% AI. SiBot was the only system to recognise the mixed provenance, assigning 97.7% hybrid confidence with marginal AI (0.9%) and human (1.4%) probabilities and correctly characterising the text as human-written and AI-refined. Two conclusions follow. The existing detectors do not merely underperform; they produce mutually contradictory verdicts on identical inputs, spanning 0% to 76% AI on the same passage, which means that a user consulting more than one of them has no principled way to reconcile them. And they are structurally incapable of source attribution and hybrid detection, not merely inaccurate at it: the “x” entries that dominate their columns are missing capabilities, not failed measurements.

Figs 14A–E shows the corresponding interface outputs for the human-written sample, which make the contrast directly visible. It should be noted that this benchmark, while quantitative and controlled for provenance, remains limited in scale: it evaluates three carefully constructed texts on a single topic. A broader evaluation across a standardised public benchmark is identified in Section 6 as necessary further work, and the conclusions above should be read as establishing a capability gap rather than a precise accuracy ranking.

5.5. Lessons learned in production deployment

Deploying a research model as a public web application presented several challenges and yielded valuable insights. Model serialization and size were primary concerns: large artefacts required Git LFS for version control on Hugging Face Spaces, and compression proved essential to reduce storage and loading time. Ensuring environment parity between the training environment and the production Docker container was critical to avoid silent prediction errors; mismatches in scikit-learn versions, for instance, could cause inconsistent feature transformations. Rate limiting and abuse prevention became necessary to manage costs on the free tier; implementing per-user rate limits based on roles stored in Supabase effectively controlled API usage. Finally, observability was enhanced by integrating structured logging and an audit trail, which proved crucial for debugging, security compliance, and understanding user behavior. Deploying the hybrid model introduced one further constraint absent from the single-branch pipeline: because the two branches must be applied in exactly the order and configuration used during training, the vectorizer, the tokenizer configuration and the encoder checkpoint are versioned together as a single artefact, and the API refuses to start if any component hash does not match the manifest. Silent branch mismatch would produce plausible but meaningless probability distributions, which is a more dangerous failure mode than an outright crash.

5.6. Bias, ethical considerations and potential for misuse

A system that attributes text to a source is a system that can be wrong about a person, and the consequences of being wrong are not symmetric. We therefore set out the risks explicitly rather than leaving them implicit.

Bias. Three sources of bias are relevant. The corpus is monolingual English and was authored predominantly in a single institutional and regional context; prior work has established that AI-text detectors systematically misclassify writing by non-native English speakers as machine-generated [23], and nothing in the present design immunizes SiBot against that failure mode. We have not evaluated performance disaggregated by author background, and we do not claim fairness across such groups. The human-written class is drawn from academic and student writing, so the model’s implicit notion of “human” is a formal register; informal, dialectal or creative human writing lies outside the distribution it has seen. And the prompt bank, while spanning six domains, encodes the authors’ choices about what constitutes a representative question.

Reliability and appropriate use. At 92.38% accuracy, roughly one verdict in thirteen is wrong. In an academic-integrity context, that rate is unacceptable as a basis for a disciplinary finding, and we state plainly that SiBot’s output should be treated as one piece of evidence prompting further inquiry, never as proof of misconduct. The interface is deliberately designed to support this: it returns a full probability distribution rather than a verdict, and the sentence-level view localizes the evidence, so that a reviewer can inspect the basis of a judgment rather than defer to it. The SHAP and LIME facilities exist for the same reason.

Misuse. Attribution technology can be turned to purposes its authors did not intend. Source attribution could be used to identify which model an author or organization relies on, which is commercially or politically sensitive information; it could support surveillance of writing practices in workplaces or classrooms in ways that are disproportionate; and a public attribution system provides an oracle against which adversaries can iterate to produce text that evades it, as the adversarial robustness results in RAID [37] demonstrate for detectors generally. Rate limiting and audit logging mitigate the last of these only partially. We note also that a system distinguishing AI-refined from human-written text could be used to penalise legitimate assistive use of AI by writers with disabilities or by non-native speakers, which would invert the intended purpose entirely.

Temporal validity. Large language models are updated frequently, and a stylistic fingerprint captured at one point in time may not survive a model revision. Any deployed attribution system therefore has a shelf life, and treating its verdicts as stable across model generations is unsafe. The SiBot deployment records the corpus construction date alongside its predictions for this reason, and periodic re-training against current model versions is a maintenance requirement rather than an optional enhancement.

5.7. Limitations

The principal limitations are as follows. The corpus is monolingual and drawn from controlled prompt settings, so it may not capture the full variability of real-world content generation. The set of generators is fixed at a point in time, and performance against models released subsequently is unknown. The framework has not been validated on external public benchmarks such as RAID or OpenTuringBench, which is necessary before the accuracy figures can be compared directly with published results from other groups. Adversarial robustness has not been evaluated: we have not tested how the model behaves under deliberate paraphrasing attacks, which the recent literature identifies as the dominant failure mode for detectors. The explainability analysis relies on post-hoc methods that approximate rather than reveal the model’s internal decision process. And the deployed system supports ten classes rather than the full twenty-two, which is a deliberate accuracy-driven choice but nonetheless a restriction on coverage.