You Can’t Detect Me! Using Prompt Engineering to Generate

Undetectable Student Answers

Marie Ernst

, Fabian Rupp

and Katharina Simbeck

Hochschule f

ur Technik und Wirtschaft Berlin, Berlin, Germany

Keywords:

AI-detection-Tools, Readability, Prompt Engineering.

Abstract:

Large Language Models (LLMs) have created the opportunity for students to generate answers to assignments.

While educators rely on detection tools to identify generated content, students can employ prompt engineering

techniques to modify the style of generated outputs and decrease likelihood of detection. In this study, we

analyze the impact of intentional AI obstruction through student prompt variation on detection rate using

three different AI detection tools. In addition, the AI generated answers are analyzed with regards to their

complexity and readability. We found that AI detection tools reliably identiﬁed AI generated text. However,

prompts leading to intentional imperfections, varied sentence structures and a dynamic writing style were able

to reduce recognition rates drastically. We also conﬁrmed that undetected answer were indeed generated in a

less elaborated style, commonly associated with younger learners.

1 INTRODUCTION

With the increasing prevalence of tools using gen-

erative artiﬁcial intelligence (AI) such as ChatGPT,

it has become increasingly challenging in academic

contexts to determine whether submitted work is au-

thored by students. In response to this issue, special-

ized AI detection tools have been developed to distin-

guish between machine-generated and human-written

texts. These tools claim to accurately identify AI-

generated content in a signiﬁcant number of cases.

This claim raises the question of whether and how

easyly these detection algorithms can be outwitted.

This study explores how prompts can be designed to

hinder the correct classiﬁcation of responses as AI

generated. As AI generated answers tend to be more

elaborate and sophisticated, we utilize text readabil-

ity measures to quantitatively describe the impact of

the prompt on the generated text. We have chosen ex-

emplary assignments in a computer science class in a

higher education context.

Previous studies in the ﬁeld of AI detection tools

indicate that the detection accuracy of these tools can

be manipulated (Krishna et al., 2023). For example,

the use of paraphrasing tools such as DIPPER signif-

icantly decreases the detection rate of DetectGPT, re-

ducing it from initial 70.3% to as low as 4.6% (Kr-

https://orcid.org/0009-0009-5963-8539

https://orcid.org/0009-0006-7946-9689

https://orcid.org/0000-0001-6792-461X

ishna et al., 2023; Chaka, 2023; Weber-Wulff et al.,

2023; Kumarage et al., 2023; Flitcroft et al., 2024;

Foster, 2023).

However, such manipulations typically rely on ex-

ternal algorithms or sophisticated techniques. This

study, in contrast, examines whether comparable ef-

fects can be achieved through targeted modiﬁcations

to prompts alone—without the need for additional

software.

Inspired by the work of Weber-Wulff et al. (2023),

who advocate for further research on obfuscation

strategies to manipulate AI recognition tools, includ-

ing the use of machine paraphraser and patch writers,

this study seeks to advance understanding in this area.

By analyzing and optimizing prompts, this research

aims to uncover which linguistic characteristics and

wordings are most likely to be interpreted as ’human

written’ by AI-driven text recognition systems.

The readability of texts is a crucial factor in de-

termining how effectively readers can absorb and un-

derstand information (Wang et al., 2022). A previ-

ous readability study demonstrates that text complex-

ity negatively impacts reading outcomes, particularly

oral reading ﬂuency and recall. More complex texts

impose higher cognitive demands, making compre-

hension more difﬁcult. Therefore, structure and read-

ability of a text are crucial factors for its understand-

ability (Spencer et al., 2019).

The research questions investigated in this study

are:

304

Ernst, M., Rupp, F. and Simbeck, K.

You Can’t Detect Me! Using Prompt Engineering to Generate Undetectable Student Answers.

DOI: 10.5220/0013357500003932

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Computer Supported Education (CSEDU 2025) - Volume 2, pages 304-311

ISBN: 978-989-758-746-7; ISSN: 2184-5026

RQ1: How does intentional prompt variation af-

fect the detection rates of AI-generated content across

different AI detection tools?

RQ2: How does prompt engineering inﬂuence the

complexity and quality of AI-generated responses?

The paper is structured as follows: Section 2 pro-

vides an overview of AI detection tools and their

mechanisms. Section 3 details different prompt

strategies used in this study. Section 4 discusses

text style and readability considerations. Section 5

presents the methodology, including experimental de-

sign and data collection. Section 6 outlines the results.

Finally, Section 7 concludes the paper and provides

directions for future research.

2 AI DETECTION TOOLS

With the development of advanced AI models such

as GPT-4, the identiﬁcation of AI-generated texts has

become a quality assurance step in education and

research. AI detection tools use various methods

to differentiate between human-written and machine-

generated content. Current research shows that the

effectiveness of these systems is increasingly chal-

lenged (Chaka, 2023; Weber-Wulff et al., 2023).

Weber-Wulff et al. (2023) show in comprehensive

tests that the recognition rate varies greatly depending

on the tool used. The work underlines the challenge

of establishing consistent standards for AI detection

(Chaka, 2023; Weber-Wulff et al., 2023).

Anderson et al. (2023) show that the use of para-

phrasing tools signiﬁcantly changes the AI recogni-

tion rate. In an example, the ”real” score of a text by

the GPT-2 Output detector increased from 0.02% to

99.52%.

New developments in AI detection tools such as

Fast-DetectGPT rely on the curvature of conditional

probabilities to recognize machine-generated texts by

choosing more probable words (Bao et al., 2023).

This method exploits the discrepancy between collec-

tive AI spelling and individual human spelling and

improves the efﬁciency of recognition by requiring

fewer model calls (Bao et al., 2023). Research pub-

lished in BMJ Open SEM (2023) further emphasizes

the importance of developing robust detection frame-

works to address the growing sophistication of AI text

generation systems (Anderson et al., 2023). Another

approach is the Multiscale Positive-Unlabeled (MPU)

framework, which uses length-sensitive probabilities

to accurately analyze variable-length text. It increases

recognition accuracy, especially in scenarios where

classical methods for AI detection fail due to short

texts (Chaka, 2023; Sadasivan et al., 2023).

Chakraborty et al. (2023) show that as the qual-

ity of machine-generated texts increases, the sample

size required for reliable recognition increases. Using

theoretical and empirical analyses (e.g., with datasets

such as Xsum and IMDb), they demonstrate that im-

proved recognition methods are feasible (Chakraborty

et al., 2023).

Overall, it is clear that the detection of AI-

generated texts remains a complex technical chal-

lenge that requires continuous research and further

development (Dalalah and Dalalah, 2023; Foster,

2023). To illustrate the strengths and weaknesses

of current detection methods, three commonly used

tools are examined: ZeroGPT, GPTZero, and Copy-

leaks. These tools were selected because they rep-

resent different approaches—probabilistic modeling,

statistical analysis, and hybrid AI-rule-based detec-

tion.

• ZeroGPT. This tool uses probabilistic models,

especially log-likelihood calculations, to distin-

guish between human- and AI-generated texts

(ZeroGPT, 2024). By analyzing token probabili-

ties in context, it identiﬁes patterns typical of each

(ZeroGPT, 2024). Texts with uniform probabili-

ties and low token variability are ﬂagged as AI-

generated (ZeroGPT, 2024). The tool also detects

machine-like traits, such as repetitive structures

and predictable word sequences, without needing

extensive training data (ZeroGPT, 2024). Studies

show ZeroGPT excels at spotting the consistent

styles of AI-generated writing (Kumarage et al.,

2023; Taguchi et al., 2024).

• GPTZero. This tool relies on statistical and

dynamic features, such as text length, syntac-

tic complexity, and token perplexity, to de-

tect AI-generated content (Tian and Cui, 2024;

GPTZero, 2024). Human-written texts typically

show higher perplexity due to idiomatic expres-

sions and grammatical variability (Tian and Cui,

2024). GPTZero leverages pre-trained models

like RoBERTa to spot syntactic and semantic ir-

regularities, common in AI-generated texts with

excessive coherence or complexity (Tian and Cui,

2024). It also analyzes how text traits change

with varying prompts, improving adaptability and

resilience against manipulation (Kumarage et al.,

2023; Park et al., 2024).

• Copyleaks Combining rule-based methods with

AI-driven algorithms, Copyleaks employs Detect-

GPT, which evaluates probabilistic differences be-

tween original and slightly altered texts (Copy-

leaks, 2024). Machine-generated texts are more

sensitive to such changes, as AI models fa-

vor high-probability outputs (Copyleaks, 2024).

You Can’t Detect Me! Using Prompt Engineering to Generate Undetectable Student Answers

305

Copyleaks leverages these deviations to detect AI-

generated content reliably (Copyleaks, 2024). It

also identiﬁes advanced manipulations like para-

phrasing and stylistic tweaks, making it highly

effective in academic settings (Copyleaks, 2024;

Park et al., 2024; Taguchi et al., 2024).

3 PROMPT STRATEGIES

Optimizing prompts is a promising strategy to en-

sure texts are classiﬁed as human-written (Kumarage

et al., 2023). Even simple adjustments, such as alter-

ing writing perspective or sentence structure, can sig-

niﬁcantly inﬂuence classiﬁcation results (Kumarage

et al., 2023). Variations in length, syntax, and lexi-

cal diversity provide greater control over text output

(Park et al., 2024).

Strategies to manipulate AI detection tools in-

clude introducing deliberate imperfections, such as

grammatical errors or inconsistent sentence structures

(Park et al., 2024; Foster, 2023). Alternating short

and long sentences or using idiomatic expressions

can improve human readability and reduce detectabil-

ity (Weber-Wulff et al., 2023). Authentic language

styles, like those mimicking a master’s student, fur-

ther enhance text authenticity (Kumarage et al., 2023;

Flitcroft et al., 2024). Avoiding typical AI patterns,

such as overly regular structures or excessive gram-

matical accuracy, is another common approach (Park

et al., 2024; Foster, 2023).

To investigate the impact of prompt-speciﬁc ab-

breviations in the recognition of AI-generated texts

Park et al. (2024) developed a new attack method

FAILOpt. FAILOpt uses feedback to optimize in-

structions that speciﬁcally degrade recognition per-

formance (Park et al., 2024). The study shows that the

FAILOpt method can signiﬁcantly impair the perfor-

mance of AI text detectors and that detectors trained

on limited input prompts could easily be fooled by

speciﬁc instructions (Park et al., 2024).

Foster (2023) highlights that well-crafted prompts

can enable GPT-4 to create texts classiﬁed as human

by advanced systems such as Turnitin (Foster, 2023).

Foster emphasizes that variations in text structure and

semantic depth are particularly inﬂuential in evading

detection (Foster, 2023).

Researchers argue that the detection of AI-

generated texts becomes problematic in the long

term, as the distinction between AI and human text

distributions is made more difﬁcult by total varia-

tion distance (Dalalah and Dalalah, 2023; Sadasivan

et al., 2023). This could result in recognition accu-

racy barely exceeding random decisions (Dalalah and

Dalalah, 2023; Sadasivan et al., 2023). Chaka (2023)

points out that even embedded watermarks or para-

phrasing tools can make detection almost impossi-

ble, as the similarity between AI-generated and hu-

man texts is further increased (Chaka, 2023). The

challenges of detection highlight the need for rigorous

evaluations of the systems in terms of their reliability

and robustness against tampering attempts (Weber-

Wulff et al., 2023; Sadasivan et al., 2023).

Despite their success, these techniques face chal-

lenges. Advanced methods like feedback-based opti-

mization or adversarial prompts often target speciﬁc

weaknesses of individual tools and lack universal ap-

plicability (Park et al., 2024). Moreover, such strate-

gies can reduce text readability, especially in aca-

demic settings (Foster, 2023).

While prompt design has proven effective, few

studies explore the interplay between prompt opti-

mization and text style (Flitcroft et al., 2024). Further

research is needed to assess how optimized prompts

impact both detectability and content quality (DuBay,

2007).

4 TEXT STYLE AND

READIBILITY

Readability is the ease with which a text can be

understood, inﬂuenced by its content, style, design,

and structure, and how well these align with the

reader’s background, abilities, interests, and motiva-

tion (DuBay, 2007). It is not the same as legibility,

which is about how clear and visually easy the text is

to see, such as the font and layout (Dubay, 2004). The

main idea is to help adjust the difﬁculty of written ma-

terial to match the reader’s ability, thereby enhancing

communication and learning (Zakaluk and Samuels,

1988). Edgar Dale and Jeanne Chall (1949) described

readability as the combination of factors in a text that

determine how successfully readers can understand it,

read it efﬁciently, and ﬁnd it engaging or interesting

(Dubay, 2004). Sentence construction impacts read-

ability with shorter or simpler sentences often enhanc-

ing readability while maintaining a balance of sen-

tence lengths for style (Klare, 2000). Shorter words

are more frequent and versatile in meaning, while

longer words are often less familiar; long sentences,

with complex syntactic structures place greater cog-

nitive demands on the reader (Tekﬁ, 1987). Several

readability formulas have been developed to evaluate

the difﬁculty of written text. These formulas typically

focus on two key aspects: (1) the complexity of sen-

tences, often measured by their length, and (2) the dif-

ﬁculty of words used in the text (Thomas et al., 1975).

CSEDU 2025 - 17th International Conference on Computer Supported Education

306

The Flesch Reading Ease Score and the Gunning-

Fog Index are well-established Formulas for measur-

ing text readability (Flesch, 1948). The Flesch score

considers the average sentence length and the av-

erage number of syllables per word, favoring texts

with clear and simple language (Flesch, 1948). The

Gunning-Fog Index, on the other hand, evaluates

readability by analyzing sentence length and the pro-

portion of complex words, with complex words de-

ﬁned as those with three or more syllables (Gunning,

1952).

The Wiener Sachtextformel (WSF) evaluates text

complexity by analyzing the proportion of words with

three or more syllables, words with over six let-

ters, monosyllabic words, and the average sentence

length (Dunkl, 2015). Speciﬁcally designed for Ger-

man texts, it evaluates readability by analyzing fac-

tors like sentence length, the proportion of mono-

syllabic and polysyllabic words, and word length.

Lower scores represent simpler texts (Bamberger and

Vanacek, 1984).

WSF = 0.1935 · ASL + 0.1672 · ASW

+ 0.1297 · PSW − 0.0327 · I − 0.875

The formula uses the average sentence length (ASL),

the number of syllables per word (ASW), the pro-

portion of polysyllabic words (PSW) and the pro-

portion of personal pronouns (I). These factors in-

ﬂuence the comprehensibility of the text (Bamberger

and Vanacek, 1984).

5 METHOD

This study investigates whether AI detection can be

outwitted through prompt engineering and which text

properties cause tools to fail. All analyzed texts were

created with GPT-4 using the default settings (Chat-

GPT, 2024). This ensures that the generated output

corresponds to those of standard users. The AI recog-

nition tools ZeroGPT, GPTZero and Copyleaks clas-

sify the previously generated texts (ZeroGPT, 2024;

GPTZero, 2024; Copyleaks, 2024). The free versions

of the tools are utilized and the default settings are re-

tained. The selection of these tools is based on two

primary criteria: ﬁrst, their accessibility due to being

free of charge, and second, their demonstrated per-

formance in previous studies (Singh, 2023; Chaka,

2023; Flitcroft et al., 2024; Weber-Wulff et al., 2023).

All generated texts are copied from ChatGPT with the

help of the copy key combination and pasted into the

text ﬁelds of the three AI recognition tools using the

paste key combination. Finally, the texts are classiﬁed

by the tools. The prompts and assignments used can

be found under the following link: https://iug.htw-

berlin.de/you-cant-detect-me/. All prompts, assign-

ments, and resulting texts are in German.

5.1 Assignment Questions

The tasks are set in the context of the business

computing course Enterprise Content Management

(ECM) on master degree. A total of 15 tasks are used,

covering a range of difﬁculty levels and subject ar-

eas and requiring text-based answers. The ﬁrst ﬁve

tasks (A1-A5) originate from actual examinations in

the master’s program in business computing at HTW

Berlin. The other ten tasks (B1-B10) were generated

using ChatGPT. To ensure a balanced selection, these

tasks are categorized into ﬁve levels of difﬁculty: Ba-

sic, intermediate, advanced, expert and strategic and

future-oriented tasks. Each category includes two

tasks designed to vary in technical depth and the de-

gree of abstraction required in the answers. This com-

bination of real-world and AI-generated tasks enables

a well-founded analysis of the prompts across varying

levels of difﬁculty and application scenarios.

5.2 Prompt Design

The development and optimization of prompts oc-

curs in iterative steps to identify which prompt el-

ements are most likely to cause misclassiﬁcation by

AI recognition tools. The design process is based on

the ﬁndings of previous work in this area (Kumarage

et al., 2023; Park et al., 2024; Foster, 2023). The

process begins with the basic prompt 1 that instructs

the model to directly answers the task. In the next

step, the prompt is expanded by specifying a writing

style (prompt 2). The prompting instructs the model

to write in the style of a Master’s student in busi-

ness computing in their mid-twenties with a Bache-

lor’s degree. The goal is to create an authentic yet

academic language. Additionally, the prompt empha-

sizes to create texts that AI recognition tools cannot

identify as machine-generated. Another approach in-

volves revising texts previously generated by Chat-

GPT (prompt 3 and 4). The revisions aim to elim-

inate features typically associated with AI-generated

content. Key indicators such as consistent sentence

structures, overly coherent word choices, and ﬂaw-

less transitions were found to increase the likelihood

of classiﬁcation as AI-generated (Park et al., 2024;

Foster, 2023). To counteract this, minor grammati-

cal errors and a less rigid structure should make the

text appear more human (Kumarage et al., 2023; Fos-

ter, 2023). In addition, introductions and summaries

are omitted to focus on answering the question short

You Can’t Detect Me! Using Prompt Engineering to Generate Undetectable Student Answers

307

Table 1: Prompt-characteristics used to generate texts.

1 2 3 4 5 6 7 8 9

Scientiﬁc language x

Avoid AI patterns x x x x

Mistakes x x x x

Explicit naming of AI patterns to be avoided x x x x x

Structure x x x x x x x

Student perspective x x x x x

Stylistic devices x x

Continuous text x x x x x

Short text x x

and to the point, as shorter texts are harder for AI

detection tools to classify (Chaka, 2023; Sadasivan

et al., 2023). In order to determine whether the ex-

plicit naming of the patterns to be avoided makes a

difference, the revision was tested in two scenarios:

emphasizing to create texts that AI recognition tools

cannot identify as machine-generated (prompt 4) and

on the other hand with explicit naming of the AI pat-

terns (prompt 3). In contrast to prompt 1, 2 and 4,

prompt 3 contains the typical AI patterns and it is en-

sured that these are avoided. The typical AI patterns

to be avoided were additionally tested within three

scenarios: executed to revise the previously generated

texts (prompt 3), directly in connection with the task

(prompt 5) and in combination with the word “brieﬂy”

in front of the respective task (prompt 6), e.g. “brieﬂy

describe what coded and non-coded information is”.

Furthermore, advanced rephrasing strategies are em-

ployed (prompt 7-9). These include alternating short

and long sentences, using idiomatic expressions, and

adding occasional digressions for a more dynamic and

engaging tone. Stylistic devices like comparisons,

metaphors, and rhetorical questions further enrich the

text, making it vivid and varied. The prompts were

not executed multiple times per task. An overview of

the different prompt-characteristics can be found in

table 1.

6 RESULTS

6.1 AI Detection Tools

Prompt design impact the classiﬁcation of text as

AI genereated or written by human. The effective-

ness of prompt changes between detection tools. An

overview can be found in table 2. The simplest

prompt, prompt 1, resulted in the highest likelihood

of texts being classiﬁed as AI-generated. Across all

three tools 95% of the texts were classiﬁed as AI-

generated, while only 4% were identiﬁed as human.

Only GPTZero classiﬁed two text as human-written.

Adapting the writing style in prompt 2, to resem-

ble that of a master student in business informat-

ics and avoiding AI patterns, led to slight improve-

ments. With this approach, 93% of the texts were

still recognized as AI-generated and 7% as human.

The tools again largely converged in their classiﬁca-

tions. A targeted revision in prompt 3 of the texts

created by prompt 2 improved the results. The pro-

portion of texts classiﬁed as AI-generated dropped to

29%, while 71% were classiﬁed as human. This un-

derscores the importance of explicitly addressing typ-

ical AI patterns, such as uniform sentence structures

and grammatical perfection, in the prompt. Also the

instruction to focus only on the essential points to an-

swer the question seems to have a proactive inﬂuence.

However, the tools varied in their responsiveness to

this prompt. Prompt 4 mimicks a master student

and subsequent revisions yielded lower-than-expected

success. Although 29% of texts were classiﬁed as hu-

man, this approach was less effective than the previ-

ous revision. This leads to the conclusion that enu-

merating the typical AI patterns to avoid and to fo-

cus only on the essential points to answer the question

probably has an inﬂuence on the effectiveness. This

suggests that these instructions are important to en-

sure that texts are predominantly classiﬁed as written

by humans. The highest success rate was achieved

with prompt 6 generating an lively, dynamic and de-

liberately imperfect text with varied sentence struc-

ture, varied word choice and occasionally faulty tran-

sitions. This leads to an rise of human classiﬁcations

to 86%. The human classiﬁcation rate is similar for

all 3 detection tools. Adding the term “brieﬂy” leads

to good results as well, but at 64% human classiﬁca-

tions fails to match prior results. Contrary to expecta-

tions, advanced reformulation strategies, in prompts

8 and 9, incorporating idiomatic expressions, varied

sentence structures, and occasional digressions do not

yield meaningful improvements. With 93% AI clas-

siﬁcations using prompt 8 and only 80% for prompt

9, this approach fell far short of expectations. All

three tools exhibit similar results. This suggests that

CSEDU 2025 - 17th International Conference on Computer Supported Education

308

Table 2: Comparison of AI Detection Tools: Percentage of Texts Classiﬁed as Human-Written and mean, standard deviation

and polysyllable count of WSF for the individual results.

Prompt ZeroGPT GPTZero Copyleaks Total avg.

prompt

average

score

Standard

deviation

mean

polysyll.

1 0% 13% 0% 4% 15.3 1.5 206

2 6% 13% 0% 7% 13.9 1.5 226

3 87% 73% 67% 71% 9.2 1.3 95

4 20% 67% 0% 29% 9.4 1.3 107

5 0% 13% 0% 4% 14.8 1.7 193

6 93% 80% 87% 87% 8.6 1.1 137

7 100% 40% 53% 65% 7.8 0.9 62

8 7% 13% 7% 7% 12.5 1.2 117

9 27% 13% 27% 20% 12.6 2.7 114

Total avg.

Tool

13% 12% 9% 34%

stylistic sophistication alone, without explicit imper-

fection, is insufﬁcient to inﬂuence classiﬁcation out-

comes. ZeroGPT achieves the most human classiﬁca-

tions in Prompts 7. GPTZero showed its best perfor-

mance Prompt 6. Copyleaks, the strictest tool with the

most AI classiﬁcations, responded well to Prompt 6.

The 3 tools have similar total AI classiﬁcation rates.

The study shows that targeted variations in prompt de-

sign inﬂuence the recognition rates of AI-generated

texts. With regard to research question RQ1, it can be

stated that prompts that incorporate intentional imper-

fections such as grammatical mistakes, irregular sen-

tence structures and dynamic writing styles reduce the

recognition rate, while advanced reformulations with-

out deliberate deviations were less effective. Prompts

that incorporate typical AI patterns, which should be

avoided, make detection by current tools more chal-

lenging. It is important to note that repeated execu-

tions of the same prompts can generate different texts,

potentially leading to variability in results.

6.2 Text Style and Detection

In examining the impact of text style on AI detec-

tion, the readability and complexity of texts gener-

ated by different prompts were analyzed. The dif-

ferent prompts yielded texts that differ strongly in

stylistic complexity, measured by WSF-score (Table

3). WSF was chosen because it is speciﬁcally de-

veloped for the German language and takes sentence

length, word complexity into account. WSF values

typically range from 4 to 15, where 4 indicates very

easy texts suitable for younger students, and 15 indi-

cates very difﬁcult texts suitable for advanced readers

on an academic level. The analysis revealed that some

texts were evaluated as extremely complex, due to

WSF Score ( >14), while others were deemed easily

readable for ninth-grade students (ages 14-15). WSF

scores and AI detection results show a notable cor-

relation. It aligns closely with the WSF scores, sug-

gesting it is well-suited for evaluating the readability

and complexity of German texts. For instance, texts

generated by ChatGPT that are typically at a mas-

ter’s level are often recognized as AI-generated. In

contrast, texts not recognized as AI-generated tend to

be at a high school level, suitable for students aged

14-15. This indicates that simpler texts with lower

readability scores are more likely to be classiﬁed as

human-written. Prompt 1 has no speciﬁc features

to avoid AI patterns or include mistakes and has the

highest WSF score (15.3), indicating very complex

texts that are difﬁcult to understand. Prompt 3 in-

cludes several features such as avoiding AI patterns

and incorporating mistakes, resulting in a lower WSF

score (9.2), indicating simpler and more understand-

able texts. Prompt 6 also has many features to avoid

AI patterns and include mistakes, leading to one of

the lowest WSF scores (8.6). Prompt 9 contains sci-

entiﬁc language and stylistic devices, resulting in a

higher WSF score (12.6) and a larger standard devia-

tion (2.7), indicating greater variability in text com-

plexity. Prompts that explicitly avoid AI patterns

and include mistakes result in lower WSF scores and

higher rates of human classiﬁcation. For example,

Prompt 3 and Prompt 6, which incorporate these fea-

tures, have lower WSF scores (9.2 and 8.6) and higher

human classiﬁcation rates (71% and 87%). In con-

trast, Prompt 1, with no special features and a high

WSF score (15.3), has a low human classiﬁcation rate

(4%). This indicates that simpler, less complex texts

are more likely to be recognized as human-written.

Lower WSF scores (indicating simpler texts) correlate

with higher human classiﬁcation rates. For example,

Prompt 6 has a low average WSF score of 8.6 and a

high human classiﬁcation rate of 87%. Higher WSF

scores (indicating more complex texts) correlate with

You Can’t Detect Me! Using Prompt Engineering to Generate Undetectable Student Answers

309

lower human classiﬁcation rates. Prompt 1, with an

average WSF score of 15.3, has a low human clas-

siﬁcation rate of 4%. In addition, further readability

properties were examined. Long sentences and poly-

syllabic words impact the readability of texts, mak-

ing them more challenging to understand. High sylla-

ble and lexicon counts generally indicate a more de-

tailed and complex text. Conversely, texts with more

monosyllabic words and shorter sentences promote

higher readability, resulting in better comprehension

and lower readability index scores.

7 CONCLUSIONS

The results of this study show that AI recognition

tools can be manipulated by strategic prompt de-

sign. Introducing human-like imperfections, alternat-

ing sentence structures and thus avoiding typical AI

patterns increases the likelihood of AI-generated texts

being categorized as written by humans. These strate-

gies were also very effective when combined with fo-

cused and concise responses to the task. Furthermore,

we show the varying effectiveness of the recognition

tools, with GPTZero showing the highest sensitivity

to prompt adaptation and Copyleaks the lowest.

The readability and complexity of texts generated

by different prompts were analyzed using the WSF-

readability score. Prompts that explicitly avoided AI

patterns and included mistakes resulted in lower WSF

scores and higher rates of human classiﬁcation. This

indicates that simpler, less complex texts are more

likely to be recognized as human-written.

The results of this study conﬁrm studies, such as

those by Krishna et al. (2023) and Weber-Wulff et al.

(2023), who have demonstrated that AI detection ac-

curacy can be manipulated through paraphrasing and

other external tools. This study contributes to the

ﬁeld by showing that similar effects can be achieved

through strategic prompt design alone, without the

need for additional software.

Our ﬁndings also resonate with the work of An-

derson et al. (2023), who showed that paraphrasing

tools could signiﬁcantly alter AI recognition rates.

Similarly, our study demonstrates that prompt modiﬁ-

cations can achieve comparable results. Additionally,

the research by Foster (2023) on the impact of text

structure and semantic depth on detection aligns with

our ﬁndings that dynamic and varied writing styles

reduce AI detection rates.

Moreover, the inclusion of readability analysis us-

ing the WSF formula provides a novel perspective.

While prior research has focused on the technical

manipulation of text to evade detection, our ﬁndings

highlight the importance of text readability and com-

plexity. Texts with lower readability scores, indicat-

ing simpler language, are more likely to be classi-

ﬁed as human-written. This suggests that readabil-

ity metrics can be a valuable tool in understanding

and improving the effectiveness of prompt engineer-

ing strategies.

This research not only conﬁrms the manipula-

bility of current detection systems but also provides

a framework for future studies to explore the inter-

play between readability and AI detection. Our ﬁnd-

ings highlight the need for improved detection al-

gorithms capable of recognizing prompt engineering

tactics. Further research could explore dynamic de-

tection models that adapt to evolving manipulation

strategies and ensure more robust systems for iden-

tifying AI-generated content. As AI detection tools

continue to be unreliable, educators need to consider

either controlling for AI use in in-classroom tests or

increasing task difﬁculty while allowing use of AI

tools.

REFERENCES

Anderson, N., Belavy, D. L., Perle, S. M., Hendricks,

S., Hespanhol, L., Verhagen, E., and Memon, A. R.

(2023). Ai did not write this manuscript, or did it? can

we trick the ai text detector into generated texts? the

potential future of chatgpt and ai in sports & exercise

medicine manuscript generation.

Bamberger, R. and Vanacek, E. (1984). Lesen-Verstehen-

Lernen-Schreiben. Diesterweg.

Bao, G., Zhao, Y., Teng, Z., Yang, L., and Zhang, Y.

(2023). Fast-detectgpt: Efﬁcient zero-shot detection

of machine-generated text via conditional probability

curvature. arXiv preprint arXiv:2310.05130.

Chaka, C. (2023). Detecting ai content in responses gen-

erated by chatgpt, youchat, and chatsonic: The case

of ﬁve ai content detection tools. Journal of Applied

Learning and Teaching, 6(2).

Chakraborty, S., Bedi, A. S., Zhu, S., An, B., Manocha,

D., and Huang, F. (2023). On the possibili-

ties of ai-generated text detection. arXiv preprint

arXiv:2304.04736.

ChatGPT (2024). https://chatgpt.com.

Copyleaks (2024). https://help.

copyleaks.com/hc/en-us/articles/

23768610748301-How-does-Copyleaks-work.

Dalalah, D. and Dalalah, O. M. (2023). The false positives

and false negatives of generative ai detection tools in

education and academic research: The case of chatgpt.

The International Journal of Management Education,

21(2):100822.

Dubay, W. (2004). The principles of readability. CA,

92627949:631–3309.

CSEDU 2025 - 17th International Conference on Computer Supported Education

310

DuBay, W. H. (2007). Smart Language: Readers, Read-

ability, and the Grading of Text. ERIC.

Dunkl, M. (2015). Verst

andlichkeit, pages 41–88. Springer

Fachmedien Wiesbaden, Wiesbaden.

Flesch, R. F. (1948). A new readability yardstick. The Jour-

nal of applied psychology, 32 3:221–33.

Flitcroft, M. A., Sheriff, S. A., Wolfrath, N., Maddula, R.,

McConnell, L., Xing, Y., Haines, K. L., Wong, S. L.,

and Kothari, A. N. (2024). Performance of artiﬁcial

intelligence content detectors using human and artiﬁ-

cial intelligence-generated scientiﬁc writing. Annals

of Surgical Oncology, pages 1–7.

Foster, A. (2023). Can gpt-4 fool turnitin? testing the limits

of ai detection with prompt engineering.

GPTZero (2024). https://gptzero.me.

Gunning, R. (1952). The Technique of Clear Writing.

McGraw-Hill.

Klare, G. R. (2000). The measurement of readability: use-

ful information for communicators. ACM J. Comput.

Doc., 24(3):107–121.

Krishna, K., Song, Y., Karpinska, M., Wieting, J., and

Iyyer, M. (2023). Paraphrasing evades detectors of

ai-generated text, but retrieval is an effective defense.

In Oh, A., Naumann, T., Globerson, A., Saenko, K.,

Hardt, M., and Levine, S., editors, Advances in Neu-

ral Information Processing Systems, volume 36, pages

27469–27500. Curran Associates, Inc.

Kumarage, T., Sheth, P., Moraffah, R., Garland, J., and

Liu, H. (2023). How reliable are ai-generated-text de-

tectors? an assessment framework using evasive soft

prompts. arXiv preprint arXiv:2310.05095.

Park, C., Kim, H. J., Kim, J., Kim, Y., Kim, T., Cho, H., Jo,

H., Lee, S.-g., and Yoo, K. M. (2024). Investigating

the inﬂuence of prompt-speciﬁc shortcuts in ai gener-

ated text detection. arXiv preprint arXiv:2406.16275.

Sadasivan, V. S., Kumar, A., Balasubramanian, S., Wang,

W., and Feizi, S. (2023). Can ai-generated text be re-

liably detected? arXiv preprint arXiv:2303.11156.

Singh, A. (2023). A comparison study on ai language de-

tector. In 2023 IEEE 13th Annual Computing and

Communication Workshop and Conference (CCWC),

pages 0489–0493. IEEE.

Spencer, M., Gilmour, A. F., Miller, A. C., Emerson, A. M.,

Saha, N. M., and Cutting, L. E. (2019). Understanding

the inﬂuence of text complexity and question type on

reading outcomes. Reading and Writing, 32:603–637.

Taguchi, K., Gu, Y., and Sakurai, K. (2024). The impact of

prompts on zero-shot detection of ai-generated text.

arXiv preprint arXiv:2403.20127.

Tekﬁ, C. (1987). Readability formulas: An overview. Jour-

nal of documentation, 43(3):261–273.

Thomas, D. G., Hartley, R. D., and Kincaid, J. P. (1975).

Test-retest and inter-analyst reliability of the auto-

mated readability index, ﬂesch reading ease score, and

the fog count. Journal of Reading Behavior, 7(2):149–

154.

Tian, E. and Cui, A. (2024). Gptzero: Towards detection

of ai-generated text using zero-shot and supervised

methods.

Wang, S., Liu, X., and Zhou, J. (2022). Readability is de-

creasing in language and linguistics. Scientometrics,

127:4697–4729.

Weber-Wulff, D., Anohina-Naumeca, A., Bjelobaba, S.,

Folt

ynek, T., Guerrero-Dib, J., Popoola, O.,

Sigut, P.,

and Waddington, L. (2023). Testing of detection tools

for ai-generated text (arxiv: 2306.15666). arxiv.

Zakaluk, B. L. and Samuels, S. J. (1988). Readability: Its

Past, Present, and Future. ERIC.

ZeroGPT (2024). https://www.zerogpt.com.

APPENDIX

Table 3: Used wordings for the prompt characteristics.

Characteristic Wording

Scientiﬁc

language

Write the text in neutral and

factual language

Avoid AI

patterns

generated text is not recognized

by AI detectors

Mistakes insert a few grammatical errors

Explicit

naming of

AI patterns

to be avoided

Avoid typical AI patterns such as

uniform sentence structure,

consistent word choice, perfectly

ﬂowing transitions and

grammatical correctness. Make

your text varied, “imperfect” and a

little less stringent.

Structure omit all headings, introduction

and conclusion/summary

Student

perspective

Master’s student, using natural

language as a person in their

mid-twenties with a Bachelor’s

degree

Stylistic

devices

Occasionally use stylistic devices

such as rhetorical questions,

comparisons or metaphors

Continuous

text

Write a continuous text

Short text Focus only on necessary

information to answer the question

and leave out everything else

You Can’t Detect Me! Using Prompt Engineering to Generate Undetectable Student Answers

311