Pith. sign in

REVIEW 4 major objections 8 minor 15 references

AI Generated Text Detection Using Instruction Fine-tuned Large Language and Transformer-Based Models

T0 review · 4 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Fine-tuning GPT-4o-mini with a single instruction prompt distinguishes human from machine text at 0.9547 macro F1 on the DeFactify unseen test set, while the best model for identifying the specific generator, BERT, reaches only 0.4698.

desk verdict A concrete but under-specified DeFactify shared-task result: simple instruction tuning of GPT-4o-mini reaches 95.5% macro F1 on binary detection, yet the paper never states how ~200 Azure-filtered test samples were handled, so the headline number is not fully verified. read the letter →

arxiv 2507.05157 v1 pith:BDB3XOQM submitted 2025-07-07 cs.CL cs.AI

Chinnappa Guggilla , Budhaditya Roy , Trupti Ramdas Chavan , Abdul Rahman , Edward Bowen This is my paper · ORCID
classification cs.CLcs.AI
keywords AI-generatedtextdetectioninstructionfine-tuningGPT-4o-miniBERTLLaMA-38BhumanvsmachineclassificationgeneratorattributionDeFactifybenchmark
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tests whether simple instruction fine-tuning can separate human-written from machine-generated text (Task-A) and identify which specific LLM generated a text (Task-B). On the DeFactify benchmark, fine-tuning OpenAI's GPT-4o-mini with a one-sentence prompt reaches a macro F1 of 0.9547 on the unseen test set for Task-A, while BERT, fine-tuned the same way, reaches 0.4698 for Task-B. The authors conclude that binary detection is largely achievable with simple prompts, but identifying the exact generator remains an open problem requiring larger models and richer instructions. If correct, this offers a low-effort recipe for building human-vs-machine text detectors and redirects attention to the harder attribution task.

What carries the argument

The mechanism is supervised fine-tuning from instruction-formatted examples: for each input text, a short classifier prompt such as 'classify whether the given text is written by human or machine' is prepended, and the target label is used as the supervised response. GPT-4o-mini and LLaMA-3 8B are adapted by continuing language-model training on these prompt-label pairs, while BERT is fine-tuned as a text classifier with the same labels. The central objects are the instruction-tuned dataset, the three fine-tuned models, and the macro F1 evaluation used by the benchmark. BERT's role is notable: despite being older and smaller than the generative LLMs, it gives the best attribution result (0.4698 Task-B F1) and a strong binary result (0.7670 Task-A F1), because it can be trained directly as a multi-class classifier.

What would settle it

Re-run the Task-A evaluation on the same test set with two rules for the ~200 content-filtered samples: count all of them as errors, and count all of them according to their true labels; if the macro F1 drops by more than a couple of points under the error rule, the claim of 0.9547 is not robust to the filtering artifact. The same recalculation should be repeated in a deployment setting with no content filtering applied at all.

Watch

Extended reading notes

Core claim

The paper's central claim is that instruction fine-tuning with simple prompts is enough to make a strong binary detector. After prepending a one-sentence classifier prompt to each text and fine-tuning on labeled examples, GPT-4o-mini reaches 0.9547 macro F1 in distinguishing human-written from machine-generated text on the DeFactify unseen test set. BERT, fine-tuned for the seven-way attribution task, reaches 0.4698 macro F1 in identifying which of six LLMs generated a text. LLaMA-3 8B, fine-tuned with 4-bit LoRA, achieves only 0.14 macro F1 on the same attribution task. The authors also report that all models' test performance falls well below their validation performance, with BERT dropping from near-perfect validation scores to 0.7670 (Task-A) and 0.4698 (Task-B) on the test set, and they conclude that pinpointing the generator will require larger models, longer contexts, and detailed prompting instructions.

Load-bearing premise

The reported Task-A score assumes that the roughly 200 test samples blocked by the provider's content filter were handled in a way that does not inflate performance; if those samples were simply dropped, the 0.9547 F1 could be an upper bound rather than an unbiased estimate.

Editorial extensions

If this is right

  • A detector for human-vs-machine text can be built by fine-tuning a commercial LLM with a one-sentence instruction prompt, reaching 95.47% macro F1 on the benchmark's unseen test set.
  • Identifying the exact generator remains largely unsolved: the best attribution score is 46.98% macro F1 (BERT), and LLaMA-3 8B reaches only 14%, far below a usable level.
  • Validation performance is much higher than test performance (BERT reaches 100% validation F1 for Task-A versus 76.7% on test), so models fine-tuned this way can overfit to the validation distribution.
  • Simple prompts suffice for binary detection, but the paper concludes that attribution will require larger models, longer contexts, and prompts that describe each generator's linguistic characteristics.
  • On the test set, combining GPT-4o-mini's Task-A result with BERT's Task-B result yields the paper's headline pair of 0.9547 and 0.4698; neither model alone solves both tasks.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper does not state whether the roughly 200 test samples blocked by the provider's content filter were excluded from the F1 denominator; a reader deploying this detector in an environment without such filtering should treat 0.9547 as a ceiling rather than the expected operational accuracy.
  • The large validation-to-test gap suggests the benchmark's train and test distributions differ substantially, so a user checking only validation scores would overestimate real-world performance by a wide margin on Task-B.
  • Because the winning recipe is just a prepended instruction and standard fine-tuning, the same approach could be tested cheaply for other binary provenance questions, such as detecting machine-translated or AI-summarized text, an experiment the paper does not run.
  • The paper's own data analysis shows human text has wider lexical diversity than machine text; adding lexical-diversity and sentence-length features as auxiliary signals could be a testable way to improve Task-B, where simple prompts underperform.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 8 minor

Summary. This manuscript reports experiments on the DeFactify 2025 Text Counter Turing Test benchmark for two tasks: Task-A, binary classification of human-written versus machine-generated text, and Task-B, identification of the specific LLM (among six LLMs plus human) that produced a given text. The authors fine-tune GPT-4o-mini (Task-A only), LLaMA-3 8B (Task-A and Task-B), and BERT (Task-A and Task-B) using simple instruction prompts. The abstract claims accuracies of 0.9547 on Task-A for GPT-4o-mini and 0.4698 on Task-B for BERT. Validation results show very high F1 scores (up to 1.00 for BERT on Task-A), while test results show large drops for BERT and LLaMA-3 8B on both tasks. The paper concludes that simple prompts are effective for Task-A but that Task-B requires more sophisticated approaches.

Significance. The empirical result that a fine-tuned GPT-4o-mini reaches 0.9547 macro F1 on Task-A of an external shared benchmark is a potentially useful data point for the AI-text-detection community, and the benchmark choice avoids circularity in the evaluation. The paper does not introduce a new method, but it provides a head-to-head comparison of three model families under fine-tuning on the same data, with hyperparameters tabulated. However, the contribution is limited by the lack of code release, the absence of confidence intervals, and the unspecified handling of filtered test samples. The Task-B results (0.4698 for BERT, 0.14 for LLaMA-3 8B) are modest, and the paper's own analysis is thin.

major comments (4)
  1. [Section 4, Table 6] The paper reports that 'Open AI's GPT_4o-mini fine-tuned model did not provide answers for approximately 200 test samples' due to Azure content filtering, and that 'This has limited the GPT_4o-mini model performance by 2%,' but it never states whether these samples were excluded from the denominator of the reported Task-A F1 score of 0.9547 or counted as errors. If they were excluded, the headline F1 is conditional on the responsive subset and is potentially biased upward, especially because Section 2.1 notes the dataset includes medical and election topics that are plausible triggers for content filtering. The authors must specify the exact scoring protocol, report the number of valid samples used in the denominator, and if possible provide the F1 computed with the filtered samples scored as errors or re-run the evaluation with the filter disabled.
  2. [Abstract, Section 4, Table 6] The abstract states that 'the fine-tuned version of GPT_4o-mini and the BERT model has achieved accuracies of 0.9547 for Task-A and 0.4698 for Task-B,' but Section 4 explicitly says performance is measured using macro-averaged precision, recall and F1, and Table 6 columns are headed 'Task-A F1' and 'Task-B F1.' The reported values are therefore macro F1 scores, not accuracies. This mislabeling in the abstract should be corrected (e.g., change 'accuracies' to 'macro F1 scores') to avoid misleading readers about the nature of the headline result.
  3. [Tables 4-6] The validation-to-test F1 drops are dramatic: BERT falls from 1.00 to 0.7670 on Task-A and from 0.98 to 0.4698 on Task-B, while LLaMA-3 8B falls from 0.93 to 0.14 on Task-B, which is at the chance level for a 7-class problem. The paper merely states that the models 'have not generalized well on the test dataset' without investigating possible causes such as overfitting, label leakage, or distribution shift. In particular, a validation F1 of exactly 1.00 for BERT on Task-A is a red flag that the validation set may not be representative of the test distribution; the authors should analyze this discrepancy and discuss its implications for the reliability of the reported test numbers.
  4. [Section 3.2, Table 3] The fine-tuning configuration for GPT-4o-mini is reported as 'Batch Size: 2 to 4; Epochs: 1 to 2,' which is a range rather than a single configuration. The paper does not state the exact values used for the final model, nor the number of training examples used, nor the model-selection criterion on the validation set. Without these details, the reported 0.9547 is not independently reproducible. Please provide the exact settings and selection procedure.
minor comments (8)
  1. [Section 2.1] The sentence 'The labels in train and validation set are equally distributed as shown in the Table 1' is contradicted by Table 1, which shows 7,321 human versus 43,926 machine samples in the training set and 1,569 versus 9,414 in the validation set. Please correct the statement to reflect the actual imbalanced distribution.
  2. [Section 3.2] The text says 'The training and validation losses for Tasks-A and B are plotted in Figure 2a and Figure 3, respectively,' but Figure 3 is an illustration of a filtered input sample, and Figure 2b contains the Task-B loss curves. The reference should be to Figure 2a and Figure 2b.
  3. [References] Reference [3] contains the placeholder 'Accessed: (Use the date of access).' Please provide the full access date.
  4. [Throughout] Model naming is inconsistent: 'GPT_4o-mini' appears in the abstract and most sections, while Table 6 uses 'GPT_4-o.' Please standardize to one name (e.g., 'GPT-4o-mini').
  5. [Section 3] The sentence 'For Task-A, we fine-tuned OpenAI's GPT_4o-mini and Google's BERT' omits LLaMA-3 8B, which is listed for Task-A in Table 2 and reported in Table 4. Please correct the list of models.
  6. [Section 3.2, Table 6] The statement 'The final results shown in the Table 6 are obtained by combining GPT_4O-mini Task-A results with BERT model Task-B results' is ambiguous because the table also reports BERT Task-A and LLaMA-3 8B Task-B results. Clarify which rows constitute the final combined system.
  7. [Tables 4-6] The paper reports no confidence intervals or significance tests for any F1 scores; in particular, the claim that GPT-4o-mini outperforms BERT on Task-A (0.9547 vs. 0.7670) would benefit from a statistical test or at least a statement about variance.
  8. [Author footnote] The author footnote contains the stray string 'envel⌢pe-⌢penc' before the email address; this appears to be a formatting artifact that should be removed.

Circularity Check

0 steps flagged · score 1.0 of 10

No material circularity: the headline F1 numbers are external-benchmark measurements, and the only self-citation is non-load-bearing related work.

full rationale

The paper's central quantitative claims are measurements of fine-tuned GPT_4o-mini, BERT, and LLaMA-3 models on the DeFactify workshop's externally provided validation and test sets (Refs [11,12]). The labels and data splits come from that benchmark, not from the paper's own assumptions or fitted quantities. The models are trained on the provided training subset and evaluated on held-out test data, so the reported F1 scores are not defined in terms of the model outputs themselves. The only author-affiliated citation is Ref [7] (Abburi et al., which includes co-author E. Bowen), but it is cited only as related prior work: 'Harika A. et al. [7], used ensemble of five pre-trained transformer based models to generate the robust features and various ML algorithms...' This citation is not load-bearing; it does not justify the benchmark, the labels, the fine-tuning procedure, or the claimed F1 values. The Section 4 discussion of roughly 200 test samples filtered by Azure OpenAI content policy is a genuine limitation and reproducibility concern about whether the 0.9547 F1 denominator includes those samples, but it is not circularity: the reported number is a measurement on an external test set, not a construct that reduces to its own inputs. No self-definitional equations, fitted-input-called-prediction steps, uniqueness imports, or ansatz-smuggling citations appear. Score 1 reflects only the presence of a minor, non-load-bearing self-citation.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new theoretical constructs. The central results rest on the DeFactify benchmark, hand-chosen fine-tuning configurations, and several unverified assumptions about label quality, distribution shift, and the treatment of failed API responses.

free parameters (3)
  • BERT fine-tuning configuration = max_seq_len=512, max_features=20000, lr=2e-5, batch_size=6, epochs=3
    Hand-chosen hyperparameters from the ktrain library defaults; not tuned per class or analyzed for sensitivity. The central Task-A and Task-B F1 values depend on these choices.
  • LLaMA-3 8B fine-tuning configuration = lr=2e-4, max_steps=400, seq_len=8000, 4-bit LoRA
    Learning rate and steps are defaults from Unsloth's FastLanguageModel; no sweep is reported. The Task-B F1 of 0.14 may partly reflect this configuration.
  • GPT-4o-mini fine-tuning configuration = batch_size=2 to 4, epochs=1 to 2
    Reported as ranges, so the exact configuration used for the 0.9547 Task-A F1 is not uniquely specified.
assumptions (4)
  • domain assumption DeFactify dataset ground-truth labels are correct and complete
    All reported F1 scores depend on the workshop's labels for train, validation, and test; the paper does not audit label quality.
  • domain assumption Test set is drawn from the same distribution as training/validation
    The paper uses validation scores to motivate the approach but reports a large drop on test (BERT Task-A 100% to 76.7%), indicating distribution shift; no analysis of the test distribution is provided.
  • ad hoc to paper Excluding the dataset's original prompts prevents instruction-tuning leakage
    Section 3.1 states that prompts in the dataset are excluded, but no leakage test is reported.
  • domain assumption Azure content-policy filtering does not bias the Task-A test evaluation
    Section 4 reports ~200 filtered samples; the handling of these samples is unspecified, yet the headline F1 depends on it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AI Generated Text Detection Using Instruction Fine-tuned Large Language and Transformer-Based Models." pith.science (2026). https://pith.science/paper/BDB3XOQM

@misc{pith2026250705157,
  author       = {Pith},
  title        = {Pith review of: AI Generated Text Detection Using Instruction Fine-tuned Large Language and Transformer-Based Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BDB3XOQM}},
  note         = {Machine review of arXiv:2507.05157}
}
read the original abstract

Large Language Models (LLMs) possess an extraordinary capability to produce text that is not only coherent and contextually relevant but also strikingly similar to human writing. They adapt to various styles and genres, producing content that is both grammatically correct and semantically meaningful. Recently, LLMs have been misused to create highly realistic phishing emails, spread fake news, generate code to automate cyber crime, and write fraudulent scientific articles. Additionally, in many real-world applications, the generated content including style and topic and the generator model are not known beforehand. The increasing prevalence and sophistication of artificial intelligence (AI)-generated texts have made their detection progressively more challenging. Various attempts have been made to distinguish machine-generated text from human-authored content using linguistic, statistical, machine learning, and ensemble-based approaches. This work focuses on two primary objectives Task-A, which involves distinguishing human-written text from machine-generated text, and Task-B, which attempts to identify the specific LLM model responsible for the generation. Both of these tasks are based on fine tuning of Generative Pre-trained Transformer (GPT_4o-mini), Large Language Model Meta AI (LLaMA) 3 8B, and Bidirectional Encoder Representations from Transformers (BERT). The fine-tuned version of GPT_4o-mini and the BERT model has achieved accuracies of 0.9547 for Task-A and 0.4698 for Task-B.

Figures

Figures reproduced from arXiv: 2507.05157 by the authors.

Figure 1
Figure 1. An illustration of the sample format includes an example prompt, a human-generated story, and text generated by the gemma-2-9b model from the text detection training dataset. [12] [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Train and validation losses of fine-tuned BERT model for Task-A and Task-B 4. Results and Analysis The performance of proposed methodology is measured using macro-averaged precision, recall and F1 scores. The results of fine-tuned models such as GPT_4o-mini, LLaMA-3 8B and BERT for Task-A on the validation set are shown in the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. An illustration of text portion of an input sample that triggered no answer from the model [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Distribution of lexical diversity in text detection training dataset [12] [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Distribution of average sequence length in text detection training dataset [12] [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 10 canonical work pages

  1. [1]

    Solaiman, M

    I. Solaiman, M. Brundage, J. Clark, A. Askell, A. Herbert-Voss, J. Wu, A. Radford, G. Krueger, J. W. Kim, S. Kreps, M. McCain, A. Newhouse, J. Blazakis, K. McGuffie, J. Wang, Release strate- gies and the social impacts of language models, 2019. URL: https://arxiv.org/abs/1908.09203. arXiv:1908.09203

    arXiv 2019

  2. [2]

    Uchendu, Z

    A. Uchendu, Z. Ma, T. Le, R. Zhang, D. Lee, Turingbench: A benchmark environment for turing test in the age of neural text generation, 2021. URL: https://arxiv.org/abs/2109.13296. arXiv:2109.13296

    arXiv 2021

  3. [3]

    Needle, Ai detection: How to pinpoint ai generated text and imagery [+ detection tools], https://blog.hubspot.com/marketing/ai-detection, 2024

    F. Needle, Ai detection: How to pinpoint ai generated text and imagery [+ detection tools], https://blog.hubspot.com/marketing/ai-detection, 2024. Accessed: (Use the date of access)

    work page 2024

  4. [4]

    Abdali, R

    S. Abdali, R. Anarfi, C. Barberan, J. He, Decoding the ai pen: Techniques and challenges in detecting ai-generated text, in: Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 6428–6436

    work page 2024

  5. [5]

    J. Wu, S. Yang, R. Zhan, Y. Yuan, D. F. Wong, L. S. Chao, A survey on llm-generated text de- tection: Necessity, methods, and future directions, 2024. URL: https://arxiv.org/abs/2310.14724. arXiv:2310.14724

    arXiv 2024

  6. [6]

    C. Mao, C. Vondrick, H. Wang, J. Yang, Raidar: generative ai detection via rewriting, 2024. URL: https://arxiv.org/abs/2401.12970. arXiv:2401.12970

    arXiv 2024

  7. [7]

    A Simple yet Efficient Ensemble Approach for AI-generated Text Detection

    H. Abburi, K. Roy, M. Suesserman, N. Pudota, B. Veeramani, E. Bowen, S. Bhattacharya, A simple yet efficient ensemble approach for ai-generated text detection, 2023. URL: https://arxiv.org/abs/ 2311.03084. arXiv:2311.03084

    work page Pith review arXiv 2023

  8. [8]

    OpenAI, Gpt-4 technical report, 2023

    work page 2023

Show all 15 references
  1. [9]

    Accessed: 2024-07-18

    OpenAI, Gpt-4o mini: Advancing cost-efficient intelligence, 2024. Accessed: 2024-07-18

    2024

  2. [10]

    J. D. M.-W. C. Kenton, L. K. Toutanova, Bert: Pre-training of deep bidirectional transformers for language understanding, in: Proceedings of naacL-HLT, volume 1, Minneapolis, Minnesota, 2019, p. 2

    2019

  3. [11]

    R. Roy, G. Singh, A. Aziz, S. Bajpai, N. Imanpour, S. Biswas, K. Wanaskar, P. Patwa, S. Ghosh, S. Dixit, N. R. Pal, V. Rawte, R. Garimella, A. Das, A. Sheth, V. Sharma, A. N. Reganti, V. Jain, A. Chadha, Overview of text counter turing test: Ai generated text detection, in: pr...

    2025

  4. [12]

    R. Roy, G. Singh, A. Aziz, S. Bajpai, N. Imanpour, S. Biswas, K. Wanaskar, P. Patwa, S. Ghosh, S. Dixit, N. R. Pal, V. Rawte, R. Garimella, A. Das, A. Sheth, V. Sharma, A. N. Reganti, V. Jain, A. Chadha, Defactify-text: A comprehensive dataset for human vs. ai generated text d...

    2025

  5. [13]

    M. H. Daniel Han, U. team, Unsloth, 2023. URL: http://github.com/unslothai/unsloth

    2023

  6. [14]

    E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen, Lora: Low-rank adaptation of large language models, arXiv preprint arXiv:2106.09685 (2021)

    2021 arXiv

  7. [15]

    A. S. Maiya, ktrain: A low-code library for augmented machine learning, arXiv preprint arXiv:2004.10703 (2020). arXiv:2004.10703. Figure 4: Distribution of lexical diversity in text detection training dataset [12] Figure 5: Distribution of average sequence length in text detec...

    2020 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.