On Improving the Efficiency of AI-Generated Text Detection

Bogdan Ichim

1,2

and Andrei-Cristian Năstase

Faculty of Mathematics and Computer Science, University of Bucharest, Str. Academiei 14, Bucharest, Romania

Simion Stoilow Institute of Mathematics of the Romanian Academy, Str. Calea Grivitei 21, Bucharest, Romania

Keywords: AI-Generated Text Detection, AI-Generated Content, Large-Scale AI Detection, AI-Generated Text

Classification, Dimensionality Reduction, PCA (Principal Component Analysis), TSVD (Truncated

Singular Value Decomposition), FastICA (Fast Independent Component Analysis), BERT Embeddings,

SimCSE Embeddings, TF-IDF (Term Frequency-Inverse Document Frequency).

Abstract: This paper proposes methods of making AI-Generated Text Detectors more computationally efficient

without paying a high price in prediction accuracy. Most AI-Detectors use transformer-based architectures

with high-dimensional text embedding vectors involved in the pipelines. Applying dimension reduction

algorithms to these vectors is a simple idea for making the whole process more efficient. Our experimental

results reveal that this may lead from 5 up to 500 times improvements in the training and inference times,

with only marginal performance degradation. These findings suggest that integrating such methods in large-

scale systems could be an excellent way to enhance the processing speed (and also reduce the electric

energy consumption). In particular, real-time applications might benefit from such enhancements.

1 INTRODUCTION

In parallel with the recent advances in artificial

intelligence, texts generated by computers have

become increasingly indistinguishable from human

writing (see Casal & Kessler, 2023; Jakesch et al.,

2023). Even AI-generated reviews are challenging to

distinguish (Ignat et al., 2024). Large Language

Models (LLM’s) can write extremely coherent (and

sometimes relevant) texts, and that, to some, might

signal serious future problems (Bender et al., 2021).

While AI-generated texts have many uses in

fields such as marketing (generating content,

thumbnails, video edits, and so on), or customer

service, as with anything, LLM’s are a double-edged

sword – potential misuses of such tools include the

possibility of impersonation, public opinion

manipulation, and disinformation in general.

Therefore, it's becoming more and more

important to be able to recognize AI-generated text

in order to preserve confidence in digital information

and communication. Robust and efficient detection

systems are essential for ensuring this.

Methods for detecting AI-generated texts have

been proposed by several authors as can be seen in

Section 2 of this paper. All these techniques are

computationally very expensive since the models

take a long time to train and the sizes of the datasets

can grow out of control. This may result in

significant costs for the hardware, for the electric

energy used and might even impact the environment.

There is interest for making such computations more

efficient (Patel et al., 2024).

By adding dimension reduction algorithms to the

AI-detection pipelines we may improve training and

inference times and at the same time significantly

lower the computational costs. While high-

dimensional text embeddings capture nuanced

semantic features, the dimensionality reduction

algorithms maintain the essential features with

minimal information loss, while at the same time

making models smaller, faster, and more efficient.

Therefore, the contributions in this paper are the

following:

 We have prepared a new dataset, which may

be useful also for other experiments;

 We are studying how dimensionality reduction

techniques like Principal Component Analysis

(PCA), Truncated Singular Value

Decomposition (TSVD) and FastICA can be

used to minimize the size of a host of different

text embedding vectors generated by SBERT,

SimCSE and the more traditional TF-IDF with

a minimal reduction in prediction accuracy.

Ichim, B. and N

astase, A.-C.

On Improving the Efﬁciency of AI-Generated Text Detection.

DOI: 10.5220/0013433600003928

In Proceedings of the 20th International Conference on Evaluation of Novel Approaches to Software Engineering (ENASE 2025), pages 731-738

ISBN: 978-989-758-742-9; ISSN: 2184-4895

731

The newly prepared dataset, together with the

implementation used for experiments, is available on

demand from the authors.

2 RELATED WORK

Researchers have recognized the valuable role of

dimensionality reduction in the detection of AI-

generated text. For example, as shown in (Rojas-

Simón et al., 2024), feature reduction on ASCII-

based lexical representations improves classification

accuracy while reducing overheads. Our

experiments complement this type of work. While

embedding-based methods are capable of capturing

more semantic information, character-based

vectorization may serve as a viable and less

computationally expensive alternative. Hybrid

methods seem to be a very promising idea.

In (Singh et al., 2022) the authors proposed a

semantic feature selection process based on GloVe

word embeddings for eliminating redundant

features. Their method is called redundant feature

removal (rRF) and it provided a considerable

improvement in classification accuracy, beyond

reducing dimensionality. Besides that, they

introduced a novel performance metric (NPM) for

balancing between evaluation of feature-reduction

effectiveness and classification accuracy, which

could serve as a benchmark for improvements in

efficiency in AI text detection.

PCA and Autoencoders are proven to be good

dimensionality reduction techniques for increasing

classification accuracy and computational efficiency

of classic machine learning models (Khan et al.,

2022). This approach showed that PCA-based

preprocessing yielded significant improvements in

training and inference times, while also showing that

AutoEncoders (AE) could provide a viable

alternative for unsupervised data compression.

Indeed, their findings reveal that preprocessing is

essential to speeding up model convergence,

improving feature interpretability, and ensuring

robustness on various text datasets. The application

of AE’s alongside techniques like PCA, TSVD or

FastICA could achieve greater efficiency gains,

particularly in real-time AI text classification tasks.

In order to identify the optimal number of

features to use (Moulik et al., 2023) performed a

systematic analysis of feature dimensionality

reduction through PCA and decision-tree-based

classifiers (AdaBoost). Their results show that

reducing the feature space to 3-6 dimensions can

greatly speed up training time without sacrificing

classification performance - an insight directly

applicable to AI-generated text detection. Also, the

use of Kernel PCA (KPCA) with an RBF kernel

gives a valuable alternative to using regular PCA,

particularly in domains where modelling non-linear

relationships is crucial – in our case, preserving

semantic integrity is of utmost importance.

For the task of detecting AI-generated texts

supervised learning is the main technique used by

many authors. In a nutshell, various machine

learning classifiers are trained on labelled datasets

(comprising both human-written and AI-produced

texts) and used for predicting the class of unlabelled

data. For example, CamemBERT, CamemBERTa

and XLM-R are used in (Antoun et al., 2023);

ChatGPT was put up to this task in (Bhattacharjee &

Liu, 2023); a logistic regression model and two deep

classifiers based on RoBERTa are tested by (Guo et

al., 2023); Bag-of-Words with a logistic regression

classifier and a fine-tuned BERT model are used in

(Ippolito et al., 2020). Building up in complexity,

the RADAR framework is proposed by (Hu et al.,

2023), then in a recent concurrent work the

OUTFOX framework is introduced by (Koike et al.,

2024). Also, human-assisted detection methods were

proposed in (Dou et al., 2022).

Commonly, such classifiers use sentence

embeddings for the role of (text-extracted) features.

Those multi-dimensional vectors supposedly

describe the text's semantic content and can easily be

fed into the classifiers.

Fine-tuned transformers, as for example

RoBERTa (Liu et al., 2019), turned out to be very

successful at detection when trained on appropriate

datasets. Another possible approach is using

stylometric features for detecting AI-generated text

– the authors use differences and inconsistencies in

the writing style as markers of generated text

(Kumarage et al., 2023).

Finally, we also note that two recent surveys on

detecting of AI-generated texts are also available,

that is (Jawahar et al., 2020) and (Tang et al., 2024).

3 METHODOLOGY

3.1 Problem Statement

Assume we are presented with a two-column

dataset, the first column contains a “text” and the

second column contains a “label” which is 0 or 1 (in

our case 0 is the label for human-generated text and

1 is the label for AI-generated text). For the

associated binary classification problem the problem

is to find an efficient algorithm, that it we would like

to trade a little from the classification performance

ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering

732

in order to obtain a significant reduction in the

computational effort involved. In order to investigate

this problem we experiment with three text

embeddings algorithms, four machine learning

binary classification algorithms, as well as with

three dimensionality reduction algorithms.

3.2 Dataset

We have created a new dataset by merging several

publicly available labelled sets (from GitHub,

HuggingFace, and Kaggle). We ensured consistent

labelling schemes using Pandas. In our approach to

merging we also took care to remove the duplicate

entries. Next, we manually scanned samples that

were flagged by first-order heuristics: repeated

words, odd punctuation patterns, or really awkward

semantic structures. Simple examples of "unnatural"

samples include placeholder-like tokens (like

"XXXXXXXX"), repeated nonsensical phrases, or

alignment issues seen in pairs of text and labels.

Although subjective, this split was consistent in its

rule: if the text was obviously incomplete, silly, or

contained placeholder tokens, it was removed. Then,

spelling errors were rectified across the corpus with

the use of specialized libraries (SymSpell and

Neuspell). We found this step justified because

typos can introduce a large set of near-duplicate

terms in the TF-IDF vectors – diluting its power.

Also, LLMs rarely generate typos, so leaving such

mistakes in our dataset could introduce an

(exploitable) classifier bias. Typographical variance

should not be a good feature in this case.

Figure 1. The balance between human and AI-generated

text classes.

As can be seen in Figure 1, the resulting new

prepared dataset is quite balanced between the

human and AI-generated text classes. It contains

598441 samples that were further split as 80% train

dataset and 20% test dataset.

3.3 Technical Implementation

We have used three approaches in order to generate

text embeddings:

 TF-IDF representations (TfidfVectorizer);

 SBERT embeddings;

 SimCSE embeddings.

As output we get matrices of sizes (598441,

1000) for TF-IDF and (598441, 768) for SBERT and

SimCSE. (In fact this CUDA-accelerated process

yielded gigabytes of data.) The reason we included

TF-IDF representations was that we wanted to see

how relatively easily calculable embeddings can

stack up against transformer-based embeddings. We

note that TF-IDF representations took significantly

less time to generate – they were finished in a few

minutes, compared to a few hours for SBERT and

SimCSE. And to our surprise – in several

experiments, they deliver quite similar performance.

Term Frequency-Inverse Document Frequency

representations (TF-IDF) is a statistical method that

is frequently used in natural language processing

tasks for representing sentences. It was introduced as

a response to the traditional Bag-of-Words

approaches in order to account for the importance of

a word in a document relative to a corpus. As a

refresher, the TF-IDF value is the product of two

values – the term frequency and the inverse

document frequency. Term frequency reflects the

number of times a word appears in a document from

a collection (which can be interpreted as the

relevance of the word within the document itself),

while inverse document frequency expresses how

common or rare a word is across all documents in

the collection (the assumption here is that less

frequent words are of higher importance). By

combining these two metrics through multiplication,

the TF-IDF value effectively highlights words that

are significant in a specific document similar to a

Bag-of-Words approach, but the key difference lies

in downplaying common words that appear across

many documents.

Sentence-BERT (SBERT) embeddings – were

derived from BERT embeddings and are a way of

extracting text features at the sentence-level,

essentially creating embeddings for whole sentences.

While BERT alone excels at producing

contextualized word embeddings, it cannot create

fixed-sized vectors for entire sentences, which is a

crucial requirement for tasks like semantic similarity

On Improving the Efﬁciency of AI-Generated Text Detection

733

and clustering. SBERT approaches this issue by

fine-tuning BERT on sentence pairs (via a siamese

network). During inference, SBERT encodes each

sentence independently into a fixed-dimensional

vector (768 in this case), preserving the semantic

relationships between sentences. As a result, we can

use these vectors to do things like semantic search

and question answering with a significantly reduced

computational cost when compared to traditional

models.

Our last type of text embeddings, Simple

Contrastive Sentence Embeddings (SimCSE), is a

recent and more sophisticated approach to

generating sentence embeddings. It enhances the

semantic content of the sentences through

contrastive learning. The main point is training the

model to distinguish between similar and dissimilar

pairs of sentences. During training, the model is

provided with sentences and their slightly perturbed

versions (with the help of data augmentation

techniques) as positive pairs, and unrelated

sentences as negative pairs. Of course, the model

builds upon pre-trained transformers like BERT or

RoBERTa and is fine-tuned to bring the embeddings

of positive pairs closer together while pushing apart

the embeddings of negative pairs in the vector space.

This supposedly helps capture fine-grained semantic

nuances. There are many possible variants of loss

functions for this task – for example, TripletLoss,

MultipleNegativesSymmetricalLoss, and Softmax to

name a few – all of them perform well for various

use-cases.

For the binary classification task, the following

well-known machine learning algorithms were used:

 Logistic Regression;

 K-Nearest-Neighbors (KNN);

 Naïve Bayes;

 Neural Network.

These classification algorithms were chosen

almost arbitrarily for our experiments, therefore we

do not further elaborate on how each algorithm

works.

The dimensionality-reduction algorithms chosen

were

 PCA;

 TSVD;

 FastICA.

All these algorithms are available in Python from

scikit-learn (Pedregosa et al., 2011).

Principal Component Analysis (PCA) is a well-

known linear dimensionality reduction technique. In

fact, the principal components are just the

eigenvectors of the input data covariance matrix

ordered by the size of the associated eigenvalues. By

selecting the top k principal components we can

effectively reduce the data dimension – while

simultaneously maintaining most of the data

variance. The implicit assumption is that the

variance captures the most significant information in

the data.

Truncated Singular Value Decomposition

(TSVD) is a variant of Singular Value

Decomposition (SVD) and another popular

technique. It is most commonly used when we want

to reduce the dimensionality of sparse matrices (as

for example CSR matrices).

Finally, Fast Independent Component Analysis

(FastICA) is a faster way of doing Independent

Component Analysis (ICA). In short, it is a

computational way of separating multivariate signals

into additive, independent components. In other

words, its scope is to transform multivariate data

points into statistically independent components.

The first step in FastICA is to center the data

(subtract the mean) and whiten it, that is alter the

data points to become uncorrelated and have unitary

variance. Whitening is typically achieved through

Principal Component Analysis or Singular Value

Decomposition – in a sense up to this point this

technique it is a combination of the previous two

methods. After this pre-processing step, the

algorithm iteratively adjusts a set of weights in order

to maximize the statistical independence of the

rotated components. This is a very strong condition

requiring infinite data to check; therefore a proxy

called “maximal non-Gaussianity” is used. The

algorithm stops when certain convergence criteria

are met (for example, the difference from the

changed weights to the initial is smaller than an

epsilon).

4 EXPERIMENTS

The experiments were performed using Jupyter

Notebook (The Jupyter Development Team, 2015),

scikit-learn (Pedregosa et al., 2011) and PyTorch

(Paszke et al., 2019).

4.1 Performance Metrics

Since our dataset is balanced and we are interested

in a binary classification task, we considered (test)

accuracy to be good as a base metric for evaluating

the performance.

ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering

734

𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =

𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑐𝑜𝑟𝑟𝑒𝑐𝑡 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝑠

𝑡𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝑠

The accuracy takes values between 0 and 1 and

the bigger the accuracy, the better the model at the

classification task.

4.2 Tuning Model Parameters

For hyperparameter tuning (see Table 1), we have

chosen to use Optuna (Akiba et al., 2019). This is a

Bayesian hyperparameter optimizer that is very

popular in machine learning competitions (for

example the Kaggle competitions). Optuna basically

is an open-source framework for hyperparameter

optimization designed to automate and streamline

the process of finding the best hyperparameters for

various machine learning models. In a nutshell, the

user needs to define the hyperparameter space and

then Optuna repeats the process of suggesting

hyperparameters by evaluating model performance

and learning from the results. It dynamically adjusts

its search strategy based on past trials to find better

configurations more efficiently. Its algorithms

intelligently navigate the hyperparameter space,

reducing the time and resources required for manual

tuning – it also helps that it is framework-agnostic,

meaning it works with any possible algorithm – the

user just needs to specify the value to be maximized

(which is in our case accuracy).

4.3 Performance Results

The results obtained in our experiments for the

performance of the machine learning models are

contained in Table 2, in which we use the

abbreviations TF for TF-IDF, SB for SBERT and SC

for SimCSE. Is easy to see that the neural network

classifier delivers the best accuracy with all

embedding types and dimensionality-reduction

algorithms. This came as expected. It’s worth

mentioning that TF-IDF representations performed

even better than the transformers in some cases on

some reduced datasets. Naïve Bayes was clearly the

worst performing by far – it had the best accuracy of

only 79% and the worst one of 67%. In all

experiments done KNN showed an edge over

Logistic Regression.

Regarding the performance of the text

embeddings, while SimCSE embeddings are clearly

delivering the best overall performance, we note that

TF-IDF and SimCSE are delivering quite similar

performance after the dimensionality-reducing

algorithms are applied to the datasets – the

performance difference between them is not

significant. This is a quite surprising result and great

news since it is much easier computationally to

compute TF-IDF embeddings. SBERT embeddings

were systematically the worst-performing

embeddings.

Table 1: Hyperparameter Tuning Summary.

Model Hyperparameters tuned with Optuna

Logistic

Regression

C (Inverse of regularization strength), penalty ('l1', 'l2', 'elasticnet'),

solver ('newton-cg', 'lbfgs', 'liblinear', 'sag', 'saga')

K-Nearest

Neighbors

number of neighbors, weights ('uniform', 'distance'),

algorithm ('ball_tree', 'kd_tree', 'brute'), leaf_size

Naïve

Bayes

var_smoothing (represents the portion of the largest variance

of all features that is added to variances for calculation stability)

Neural

Network

number of hidden layers, number of neurons per layer, activation functions ('relu', 'sigmoid',

'tanh'), optimizer ('adam', 'sgd'), learning rate, batch size, dropout rate

Table 2: Experimental Results – Test Accuracies (% of correct predictions on test).

Dimension

Reduction

Logistic Regression K-Nearest Neighbors Naïve Bayes Neural Network

TF SB SC TF SB SC TF SB SC TF SB SC

None 89% 86% 91% 92% 91% 93% 79% 67% 76% 95% 93% 97%

PCA 84% 74% 82% 92% 90% 93% 75% 68% 72% 94% 91% 94%

TSVD 83% 74% 82% 91% 90% 93% 74% 67% 72% 93% 91% 94%

FastICA 84% 74% 82% 91% 90% 93% 74% 70% 74% 93% 92% 95%

On Improving the Efﬁciency of AI-Generated Text Detection

735

Table 3: Computational costs – Time to finish training (minutes).

Logistic Regression K-Nearest Neighbors Naïve Bayes Deep Neural Network

TF SB SC TF SB SC TF SB SC TF SB SC

None 6.75 100 10 0.16 0.12 0.1 4.75 1 1.25 197 210 110

PCA 0.21 0.25 0.12 0.01 <.01 0.01 0.06 0.06 0.06 62 27 20

TSVD 1 0.41 0.25 0.01 <.01 <.01 0.06 0.07 0.06 20 73 19

FastICA 0.06 0.06 0.19 0.01 <.01 0.01 0.06 0.06 0.06 48 50 19

Table 4: Computational costs – Inference on test set (seconds).

Logistic Regression K-Nearest Neighbors Naïve Bayes Deep Neural Network

TF SB SC TF SB SC TF SB SC TF SB SC

None 0.92 0.25 0.11 <.01 0.01 0.03 5.54 5 4.74 5.47 6.78 5.81

PCA 0.01 0.01 0.01 <.01 <.01 <.01 0.30 0.26 0.28 4.90 0.85 0.36

TSVD 0.01 <.01 0.01 <.01 <.01 <.01 0.28 0.32 0.28 1.29 0.33 2.42

FastICA 0.01 <.01 <.01 <.01 <.01 <.01 0.29 0.30 0.29 1.85 1.26 2.52

Finally, we note that applying the dimension

reduction techniques in order to significantly

decrease the computational effort seems only to lead

to a relatively minor decrease in the performance of

the classification task.

4.4 Computational Effort

The text encodings and the training of the machine

learning models were accelerated with CUDA – we

have used an RTX 4060 Mobile graphics chip made

by NVIDIA for this. Generating the full-text

encodings took around 4 hours in total. Then

performing the dimensionality reduction took around

20 minutes for each of the three algorithms –

FastICA was (ironically) the slowest to finish – this

makes sense since at the processing step FastICA

includes the previous two. All processes, even the

CPU-bound ones (like, for example, PCA and KNN)

benefited from parallelization. The maximum

amount of RAM memory needed was 10 GB. We

completed a total of 720 training cycles (that is 48

Optuna runs with 15 trials each).

The training and inference times are presented in

Tables 3 and 4, in which we use the abbreviations

TF for TF-IDF, SB for SBERT and SC for SimCSE.

The reader should note that training the models

after a dimensionality reduction algorithm was

applied is 5 to 500 times faster than on the full

dataset. In Table 4 we can see a similar trend for

inference time as well. (The times presented in Table

4 are for the test dataset inference).

4.5 Threats to Validity

Internal Validity: The preparation procedure for the

dataset relies on subjective decisions about what

counts as "unnatural" texts. While we attempted to

systematically codify guidelines, there may remain

some bias regarding which texts were filtered out.

Although hyperparameter tuning with Optuna is

robust, it can still overfit to the training split if not

carefully cross-validated.

External Validity: The data were sourced from

publicly available sources (GitHub, HuggingFace,

Kaggle), which harbour potential biases with respect

to domain coverage. Therefore, caution must be

exercised in generalizing to highly professional or

domain-specific texts, for example, in the case of

legal or medical texts. The dimension-reduction

methods could perform differently when analysing

extremely large corpora or new languages.

Construct Validity: The metric chosen for

performance evaluation was classification accuracy.

While it covers much ground in terms of statistical

performance, other metrics such as F1, precision,

recall could prove useful for a more complete

analysis. Future studies could measure robustness to

adversarial attacks, interpretability, or real-time

adaptiveness, but this is beyond the scope of analysis

in this work.

Robustness: There was no systematic analysis on

how dimension-reduced embeddings stand up to

adversarial examples specifically engineered to fool

AI detectors. The preliminary findings imply that,

due to a domain shift or the introduction of synthetic

perturbations, the detection accuracy drops.

ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering

736

Therefore future works should explore the

intricacies of the vulnerabilities in detail.

5 CONCLUSIONS

In this study we tackled the idea of applying

dimensionality reduction techniques (PCA, TSVD,

and FastICA) to multidimensional text embeddings

vectors generated via different approaches. Our

experiments show that across the board, irrespective

of the type of text embeddings or dimension

reduction technique, applying such methods to text

embeddings vectors may lead to an important

reduction in the computational effort without a

corresponding significant reduction in performance.

In particular, applying dimensionality reduction

techniques can drastically improve the running times

and overall efficiency of AI text detectors. Through

all experiments, such techniques not only

accelerated training and inference times but also

preserved performance. By using dimension

reduction techniques on can successfully manage the

trade-off between reducing model complexity and

maintaining the accuracy of the predictions made.

This makes them perfectly suitable for deployment,

especially in resource-constrained environments.

Regarding the text embeddings we also note that

SimCSE consistently delivered the best accuracy,

with the lightweight TF-IDF proving to be a fierce

competitor beating the SBERT on multiple trials.

We think that incorporating these methods

provides a viable pathway towards optimizing large-

scale AI detectors, contributing to both their

scalability and efficiency. Other important practical

implications might be speeding up real-time systems

like social-media and content monitoring which

proves to be an increasingly important problem.

Also, applying dimensionality reduction in resource-

constrained environments can reduce the memory

and energy requirements enough to make detection

significantly cheaper, but also feasible on smaller

devices. There also are potential industrial

applications – dimension reduction enables the

handling of high-throughput demands, from

massive-scale news verification to customer-service

chat filtering, without necessarily losing serious

performance.

Future research could look into the integration of

these techniques in advanced language models while

investigating their capacity to stand against potential

attacks – which a key point in this field. Having the

capacity to detect generated text in a context in

which some user included adversarial tactics in order

to fool potential detectors (e.g. including subtle

changes and aiming for the text to sound more

human) is a very important aspect that should be

further investigated. Another future direction may

examine adversarial robustness to check whether

reduced embeddings are easier or harder to fool.

Finally, hybrid methods integrating statistical

dimensionality reduction techniques with semantic

feature pruning could prove useful in order to further

improve the results.

REFERENCES

Akiba, T., Sano, S., Yanase, T., Ohta, T., Koyama, M.

(2019). Optuna: A Next-generation Hyperparameter

Optimization Framework. In KDD 2019, Proceedings

of the 25th ACM International Conference on

Knowledge Discovery & Data Mining, pages 2623 –

2631. https://doi.org/10.1145/3292500.3330701

Antoun, W., Mouilleron, V., Sagot, B., Seddah, D. (2023).

Towards a Robust Detection of Language Model

Generated Text: Is ChatGPT that Easy to Detect? In

Actes de CORIA-TALN 2023, pages 14 – 27.

Bender, E. M., Gebru, T., McMillan-Major, A.,

Shmitchell, S. (2021). On the Dangers of Stochastic

Parrots: Can Language Models Be Too Big? In FAccT

2021, Proceedings of the 2021 ACM Conference on

Fairness, Accountability, and Transparency, pages

610 – 623. https://doi.org/10.1145/3442188.3445922

Bhattacharjee, A., Liu, H. (2024). Fighting Fire with Fire:

Can ChatGPT Detect AI-generated Text? ACM

SIGKDD Explorations Newsletter 25, pages 14 – 21.

https://doi.org/10.1145/3655103.3655106

Casal, J. E., Kessler, M. (2023). Can linguists distinguish

between ChatGPT/AI and human writing? A study of

research ethics and academic publishing. Research

Methods in Applied Linguistics 2, 100068.

https://doi.org/10.1016/j.rmal.2023.100068

Dou, Y., Forbes, M., Koncel-Kedziorski, R., Smith, N.,

Choi, Y. (2022). Is GPT-3 Text Indistinguishable from

Human Text? Scarecrow: A Framework for

Scrutinizing Machine Text. In Proceedings of the 60th

Annual Meeting of the Association for Computational

Linguistics (Volume 1: Long Papers), pages 7250 –

7274. https://doi.org/10.18653/v1/2022.acl-long.501

Guo, B., Zhang, X., Wang, Z., Jiang, M., Nie, J., Ding, Y.,

Yue, J., Wu, Y. (2023). How Close is ChatGPT to

Human Experts? Comparison Corpus, Evaluation, and

Detection. Preprint arXiv: 2301.07597.

Hu, X., Chen, P.-Y., Ho, T.-Y. (2023). RADAR: Robust

AI-Text Detection via Adversarial Learning. Preprint

arXiv: 2307.03838.

Ignat, O., Xu, X., Mihalcea, R. (2024). MAiDE-up:

Multilingual Deception Detection of GPT-generated

Hotel Reviews. Preprint arXiv:2404.12938.

Ippolito, D., Duckworth, D., Callison-Burch, C., Eck, D.

(2020). Automatic Detection of Generated Text is

Easiest when Humans are Fooled. In Proceedings of

On Improving the Efﬁciency of AI-Generated Text Detection

737

the 58th Annual Meeting of the Association for

Computational Linguistics, pages 1808 – 1822.

https://doi.org/10.18653/v1/2020.acl-main.164

Jakesch, M., Hancock, J. T., Naaman, M. (2023). Human

heuristics for AI-generated language are flawed.

Proceedings of the National Academy of Sciences 120,

e2208839120. https://doi.org/10.1073/

pnas.2208839120

Jawahar, G., Abdul-Mageed, M., Lakshmanan, L. V. S.

(2020). Automatic Detection of Machine Generated

Text: A Critical Survey. In Proceedings of the 28th

International Conference on Computational

Linguistics, pages 2296 – 2309.

Khan, W., Turab, M., Ahmad, W., Ahmad, S. H., Kumar,

K., Luo, B. (2022). Data Dimension Reduction makes

ML Algorithms efficient. In Proceedings of the 2022

International Conference on Emerging Technologies

in Electronics, Computing and Communication

(ICETECC), pages 1 – 7. https://doi.org/10.1109/

ICETECC56662.2022.10069527

Koike, R., Kaneko, M., Okazaki, N. (2024). OUTFOX:

LLM-Generated Essay Detection Through In-Context

Learning with Adversarially Generated Examples. In

AAAI 2024, Proceedings of 38th AAAI Conference on

Artificial Intelligence, pages 21259 – 21266.

https://doi.org/10.1609/aaai.v38i19.30120

Kumarage, T., Garland, J., Bhattacharjee, A.,

Trapeznikov, K., Ruston, S., Liu, H. (2023).

Stylometric Detection of AI-Generated Text in Twitter

Timelines. Preprint arXiv: 2303.03697.

Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D.,

Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.

(2019). RoBERTa: A Robustly Optimized BERT

Pretraining Approach. Preprint arXiv:1907.11692.

Moulik, R., Phutela, A., Sheoran, S., & Bhattacharya, S.

(2023). Accelerated Neural Network Training through

Dimensionality Reduction for High-Throughput

Screening of Topological Materials. Preprint arXiv:

arXiv:2308.12722.

Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J.,

Chanan, G., Killeen, T., Lin, Z., Gimelshein, N.,

Antiga, L., Desmaison, A., Kopf, A., Yang, E.,

DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S.,

Steiner, B., Fang, L., Chintala, S. (2019). PyTorch: An

Imperative Style, High-Performance Deep Learning

Library. Advances in Neural Information Processing

Systems 32, pages 8024 – 8035.

Patel, P., Choukse, E., Zhang, C., Shah, A., Goiri, I.,

Maleki, S., Bianchini, R. (2024). Splitwise: Efficient

Generative LLM Inference Using Phase Splitting. In

ISCA 2024, Proceedings of the 51st ACM/IEEE

Annual International Symposium on Computer

Architecture, pages 118–132.

https://doi.org/10.1109/ISCA59077.2024.00019

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V.,

Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P.,

Weiss, R., Dubourg, V., Vanderplas, J., Passos, A.,

Cournapeau, D., Brucher, M., Perrot, M., &

Duchesnay, É. (2011). Scikit-learn: Machine Learning

in Python. Journal of Machine Learning Research 12,

2825 – 2830.

Rojas-Simón, J., Ledeneva, Y., García-Hernández, R. A.

(2024). A Dimensionality Reduction Approach for

Text Vectorization in Detecting Human and Machine-

generated Texts. Computación y Sistemas 28, pages

1919 – 1929.

https://doi.org/10.13053/cys-28-4-5214

Singh, K. N., Devi, S. D., Devi, H. M., Mahanta, A. K.

(2022). A novel approach for dimension reduction

using word embedding: An enhanced text

classification approach. International Journal of

Information Management Data Insights 2, 100061.

https://doi.org/10.1016/J.JJIMEI.2022.100061

Tang, R., Chuang, Y.-N., Hu, X. (2024). The Science of

Detecting LLM-Generated Texts. Communications of

the ACM 67, pages 50 – 59.

The Jupyter Development Team. (2015). Project Jupyter.

Jupyter Notebook. Available at https://jupyter.org/.

ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering

738