Prevalence and User Perception of Dark Patterns:

A Case Study on E-Commerce Websites of Bangladesh

Yasin Sazid and Kazi Sakib

Institute of Information Technology, University of Dhaka, Dhaka, Bangladesh

Keywords:

Dark Patterns, Data Analysis, Dataset, Bangladesh, E-Commerce, User Perception, Categorization.

Abstract:

Dark patterns are deceptive design tactics that impact user decision-making. Such tactics can harm user’s

ﬁnances, privacy, and lifestyle. Researchers have explored dark patterns from different perspectives in recent

times. However, most of the work have been conducted in western countries. As dark patterns involve people

and their psychology, explorations in other cultural contexts may generate novel insights. We carried out such

a study in the local context of Bangladesh, focusing on the prevalence and user perception of dark patterns. We

ﬁrst examined Bangladeshi e-commerce websites for dark patterns using a combination of automated methods

such as GPT-3’s in-context learning and a novel segmentation algorithm, along with manual validation to

address any limitations of the automated techniques. We also surveyed Bangladeshi university students about

exposure, awareness, and concern regarding dark patterns. Based on the ﬁndings of both explorations, we

ranked six dark pattern categories and subsequently divided them into two novel groups with distinct traits.

We also found that educational background in technology make users more aware and concerned about dark

patterns. 18.3% of the websites we analyzed in this study contained dark patterns, indicating prevalence of

such design practices in e-commerce websites of Bangladesh.

1 INTRODUCTION

Dark patterns (DP) are user interface design tactics

that deceive users to alter their decision-making by

abusing knowledge of psychology (Gray et al., 2018).

For example, ‘Forced Action’ is such a dark pattern

where users are forced to perform an undesirable ac-

tion to access some other functionality. Such design

tactics can harm users ﬁnancially, privacy-wise, and

even in the form of addiction (Mathur et al., 2021).

As a result, it is imperative to characterize and inves-

tigate these tactics thoroughly.

Existing research on dark patterns and user per-

ception has primarily focused on Europe and the

United States. However, research efforts in other local

contexts may generate novel insights because of cul-

tural and linguistic differences (Hidaka et al., 2023).

Thus, further efforts are needed to analyze dark pat-

tern data as well as user perception in a local context

like Bangladesh.

As a developing country, Bangladesh has rapidly

embraced digitalization. As a result, its e-commerce

sector has experienced signiﬁcant growth in recent

years as well. However, this growth also brings the

risk of dark patterns that can harm the increasing

number of online users. Studying dark patterns in the

e-commerce industry of Bangladesh can help us un-

derstand how much local users are exposed to these

manipulative design tactics. Understanding local user

perception regarding dark patterns is also an impor-

tant area to explore.

Dark patterns can have varying representations in

user interfaces, making it difﬁcult to identify and an-

alyze related data on a large scale. As identiﬁcation

methods of dark patterns are still in the early stages,

there have been only a few efforts in collecting and

analyzing dark pattern data on a large scale. Mathur et

al. conducted the ﬁrst exploration of textual dark pat-

tern data in a large sample of 11,000 shopping web-

sites, extracting 1,818 instances of dark pattern texts

(Mathur et al., 2019). Yada et al. extended their ex-

ploration to extract 1,178 non dark pattern texts from

the same websites (Yada et al., 2022).

Understanding user perception of dark patterns is

another relevant area of research, as these design tac-

tics abuse user psychology. Geronimo et al. found

that users struggle to recognize dark patterns, empha-

sizing the need for improved awareness mechanisms

(Geronimo et al., 2020). Bongard-Blanchy et al. dis-

covered that users can identify dark patterns but are

238

Sazid, Y. and Sakib, K.

Prevalence and User Perception of Dark Patterns: A Case Study on E-Commerce Websites of Bangladesh.

DOI: 10.5220/0012734900003687

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

In Proceedings of the 19th International Conference on Evaluation of Novel Approaches to Software Engineering (ENASE 2024), pages 238-249

ISBN: 978-989-758-696-5; ISSN: 2184-4895

often unaware of their harmful consequences, with

demographic factors inﬂuencing user perception as

well (Bongard-Blanchy et al., 2021). Gray et al. re-

vealed that users can sense manipulation, even with-

out precise terminology to describe such experience

(Gray et al., 2021). Maier found that users may ini-

tially react negatively to dark patterns but eventually

get normalized to them, suggesting awareness alone

might not be enough as a countermeasure (Maier,

2019). Hidaka et al. revealed that Japanese appli-

cations contain fewer instances of dark patterns com-

pared to the United States (Hidaka et al., 2023). This

ﬁnding emphasizes the necessity to analyze dark pat-

tern data and user perception in different local con-

texts. No such exploration has been done yet in the

local context of Bangladesh.

In this study, we analyzed the prevalence of

dark patterns in e-commerce websites of Bangladesh.

Member companies of the E-Commerce Association

of Bangladesh (e-CAB) were selected to be analyzed

to show the prevalence of dark patterns. We were able

to analyze 715 websites and HTML contents from rel-

evant webpages of those websites were extracted. A

novel page segmentation algorithm was used to iden-

tify candidate dark pattern texts from the HTML con-

tents. It was developed with a more relaxed inclu-

sion criteria for candidate dark patterns compared to

the segmentation algorithm developed by Mathur et

al., resulting in improved detection of dark patterns

(Mathur et al., 2019). Detection and classiﬁcation

of dark patterns involved a combination of automated

and manual techniques. In-context learning capabili-

ties of GPT-3 were used to detect dark patterns from

the candidate texts according to our previous work

(Sazid et al., 2023). A manual review was carried out

to remove the false positives and classify the texts into

different dark pattern categories.

We complemented our work on the prevalence of

dark patterns with an empirical study on user per-

ception of dark patterns. A survey was conducted

with Bangladeshi university students of two groups -

one with a background in technology education and

the other without such a background. Participants

were presented with user interface design instances

of different dark pattern categories and then expo-

sure, awareness, and concern ratings were collected

for each dark pattern category.

We found 931 instances of dark patterns, 99 of

which were unique instances. These dark pattern in-

stances were classiﬁed into six common dark pat-

tern categories, including ‘Misdirection’, ‘Urgency’,

‘Scarcity’, ‘Social Proof’, ‘Obstruction’, and ‘Forced

Action’. The 931 dark pattern instances belonged to

131 different websites, which is roughly 18.3% of the

715 e-commerce websites analyzed in this study. Av-

erage ratings of awareness and concern across all dark

pattern categories show that users are aware of the

inﬂuences of dark patterns and are concerned about

such design tactics in general. Based on the ﬁnd-

ings from the user perception survey and the preva-

lence data across dark pattern categories, we divided

the categories into two distinct groups - ‘Passive Dark

Patterns’ containing ‘Social Proof’, ‘Urgency’, and

‘Scarcity’ categories and ‘Active Dark Patterns’ con-

taining ‘Forced Action’, ‘Obstruction’, and ‘Misdi-

rection’ categories. We found distinct characteris-

tics of these two groups in terms of prevalence, user

awareness, and concern. User perception survey also

revealed that users with a background in technology

education were more aware and concerned about dark

patterns than users with no such background.

The main contributions of this study can be sum-

marized as follows:

• Analysis of the prevalence of dark patterns in e-

commerce websites of Bangladesh

• Novel page segmentation algorithm that extracts

candidate dark pattern texts from HTML contents

of webpages

• Collection of a dark pattern dataset containing 99

unique text instances labelled with different dark

pattern categories

• Ranking dark pattern categories in terms of user

exposure, awareness, and concern

• Novel grouping of dark pattern categories accord-

ing to their approach of abusing user psychology

• Revealing the inﬂuence of technology education

(i.e., computer science or related ﬁelds) on user

awareness and concern regarding dark patterns

2 RELATED WORK

Dark patterns have attracted a lot of attention from re-

searchers in recent times. Researchers have worked

on the deﬁnition and classiﬁcation of dark patterns

(Gray et al., 2018) (Mathur et al., 2019), explored

legal and policy-making perspectives (Luguri and

Strahilevitz, 2019), and studied user perception of

dark patterns (Geronimo et al., 2020). Detection tech-

niques (Mansur et al., 2023) (Stavrakakis et al., 2021)

and data collection efforts (Mathur et al., 2019) re-

garding dark patterns have also been explored. This

section provides an overview of past literature on tax-

onomy, data collection and analysis, and user percep-

tion of dark patterns.

Prevalence and User Perception of Dark Patterns: A Case Study on E-Commerce Websites of Bangladesh

239

2.1 Taxonomy of Dark Patterns

Dark patterns have versatile representations in user

interfaces, making it very challenging to deﬁne and

classify them under a holistic taxonomy. Brignull et

al. ﬁrst deﬁned the term ‘dark patterns’ as a set of

“tricks used in websites that make a user do things that

the user did not mean to” (Brignull et al., 2023). Re-

searchers since have tried to deﬁne and classify dark

patterns in various ways. Gray et al. classiﬁed dark

patterns into ﬁve distinct categories based on user ex-

perience - ‘Nagging’, ‘Obstruction’, ‘Sneaking’, ‘In-

terface Interference’ and ‘Forced Action’ (Gray et al.,

2018). Some researchers have classiﬁed dark patterns

speciﬁcally based on privacy (Jarovsky, 2022) (B

osch

et al., 2016). Other researchers have also studied dark

patterns found on social media (Roffarello and Rus-

sis, 2022) and video streaming platforms (Chaudhary

et al., 2022).

Mathur et al. mentioned seven broad categories

of dark patterns in their work (Mathur et al., 2019).

These dark pattern categories are brieﬂy discussed be-

low:

• Urgency refers to the category of dark patterns

that enforce a time limit on a sale or offer,

prompting users to make decisions and purchases

quickly.

• Misdirection refers to the category of dark pat-

terns that employ visuals, language, and emotions

to inﬂuence users towards or away from speciﬁc

choices.

• Social Proof refers to the category of dark pat-

terns that accelerate user decision-making and

purchases, by showing the actions and behaviours

of other users.

• Scarcity refers to the category of dark patterns

that indicate the limited availability or high de-

mand of a product, thereby increasing its per-

ceived value and desirability.

• Obstruction refers to the category of dark pat-

terns that intentionally complicate a speciﬁc ac-

tion beyond what is necessary to discourage users

from taking that action.

• Forced Action refers to the category of dark pat-

terns that require users to perform an additional

action, which is often undesirable, to accomplish

some other tasks.

• Sneaking refers to the category of dark patterns

aimed at misrepresenting user actions or conceal-

ing/delaying information that users would proba-

bly oppose if they were aware of it.

2.2 Dark Pattern Related Data

Dark pattern data collection efforts rely heavily on

the successful detection of such deceptive tactics in

websites. However, it is a challenging task, especially

without a holistic deﬁnition and classiﬁcation of dark

patterns. This is also a relatively new ﬁeld of research,

and detection methods are yet to be matured. As a re-

sult, there have been very few efforts to collect dark

pattern data on a large scale. Another challenge in

such data collection efforts is determining the appro-

priate format (image, video, text, etc.) to represent

dark patterns.

Researchers have collected dark patterns in var-

ious formats, including video recordings (Geronimo

et al., 2020), screenshots (Mansur et al., 2023), ex-

tracted features (Soe et al., 2022), and textual data

(Yada et al., 2022) from user interfaces, each with

its own advantages and disadvantages. Geronimo et

al. prepared a comprehensive video dataset com-

prising screen recordings of mobile application us-

age (Geronimo et al., 2020). The dataset contains

15 videos, accompanied by classiﬁcations of identi-

ﬁed dark patterns at different timestamps within the

videos. Mansur et al. created a dark pattern dataset

containing 501 instances by manually labelling user

interface screenshots collected from different sources

(Mansur et al., 2023). Soe et al. analyzed dark pat-

terns in cookie banners of 300 news websites (Soe

et al., 2020). For each cookie banner, they manually

extracted a set of features (location within the page,

clarity of options, site works after rejecting cookies

or not, number of clicks needed to reject all cook-

ies, etc.) in terms of dark patterns. However, data

collection and analysis efforts using video recordings,

screenshots, or manually extracted features are quite

challenging to carry out on a large scale. Alterna-

tively, textual data collection and analysis are more

convenient in terms of scalability.

Mathur et al. collected the ﬁrst large-scale textual

dark pattern data (Mathur et al., 2019). They analyzed

more than 11,000 e-commerce websites and extracted

meaningful text segments using their page segmenta-

tion algorithm. Even though they collected text seg-

ments using automated techniques, dark pattern de-

tection and classiﬁcation were carried out manually,

which is a time-consuming task. To ease the man-

ual labelling of a large number of text segments, they

used hierarchical clustering in their work. They found

1,818 instances of dark pattern texts from 1,254 web-

sites, which was roughly 11.1% of all analyzed web-

sites in their work. Yada et al. extended this textual

dataset of dark patterns by adding 1,178 non dark pat-

tern texts from the same websites (Yada et al., 2022).

ENASE 2024 - 19th International Conference on Evaluation of Novel Approaches to Software Engineering

240

Figure 1: Overview of Dark Pattern Data Collection.

2.3 User Perception of Dark Patterns

As dark patterns abuse user psychology and manip-

ulate their decision-making, user perception regard-

ing the phenomenon is a signiﬁcant area of research.

Geronimo et al. studied user awareness and their abil-

ity to recognize dark patterns found within popular

mobile applications (Geronimo et al., 2020). Their

efforts revealed that users often struggle to identify

dark patterns, highlighting the necessity to improve

user awareness mechanisms. Bongard-Blanchy et al.

investigated user awareness, and their capability to

recognize and resist dark patterns (Bongard-Blanchy

et al., 2021). They found that users can recognize dark

patterns even though they are often unaware of the

actual harmful consequences. Their study also iden-

tiﬁed that demographic factors such as age and edu-

cational background can inﬂuence user perception of

dark patterns.

Gray et al. conducted a survey on English and

Mandarin Chinese-speaking participants to explore

users’ experiences of feeling manipulated by digital

technologies, their emotional responses to such expe-

riences, and their awareness of manipulative design

practices (Gray et al., 2021). Their ﬁndings indicate

that users often have a sense that something is “off”

or manipulative, even when they lack the precise lan-

guage to describe it. Maier investigated how users

perceive, experience, and respond to dark patterns

(Maier, 2019). His ﬁndings suggest that while users

may initially be angered by dark patterns, they often

become accustomed to them over time. This leads to

the normalization of these deceptive design practices.

His work also highlights that awareness mechanisms

alone may not be sufﬁcient to safeguard users from

dark patterns.

Most research efforts in analyzing dark patterns

and related user perception have been mainly con-

ducted in Europe and the United States. Hidaka et

al. termed these works as coming from the ‘Western

context’ and argued that the ‘Eastern context’ might

be different because of cultural and language differ-

ences (Hidaka et al., 2023). Their work also revealed

fewer dark patterns in Japanese applications (i.e., the

Eastern context) compared to the Western context.

Thus, further efforts are needed to analyze dark pat-

tern data as well as user perception in a local context

like Bangladesh. Such efforts are especially crucial

for local policy-making that can protect users from

the harmful consequences of dark patterns.

3 DARK PATTERNS IN

E-COMMERCE OF

BANGLADESH

In this study, we extracted dark patterns from e-

commerce websites of Bangladesh to analyze the lo-

cal context and show the prevalence of such design

tactics in local e-commerce websites. We started our

exploration by deﬁning a sampling frame for local e-

commerce websites, followed by automated crawling

of relevant website data. The crawled textual data was

then processed using a combination of automated and

manual methods to retrieve dark pattern texts across

different categories. Figure 1 provides an overview

of the different stages of data collection carried out

for this purpose. The resulting dataset of dark pat-

terns and other technical artifacts associated with this

study are publicly available on GitHub

https://github.com/bsse1006/dark-patterns-bangladesh

Prevalence and User Perception of Dark Patterns: A Case Study on E-Commerce Websites of Bangladesh

241

3.1 Sampling Procedure

The sampling frame for this study is derived from

the membership list of the E-Commerce Association

of Bangladesh (e-CAB)

, which is a Bangladeshi lo-

cal organization that has 2,000+ e-commerce com-

panies as members. The members’ list of e-CAB

is the largest available representative sample of

Bangladeshi e-commerce companies. Thus, we con-

sidered this list as our sampling frame for this study.

Upon inspecting the members’ list of e-CAB

available online at ‘https://e-cab.net/member-list’, we

found a general pattern in webpage URLs associated

with each e-CAB member proﬁle - ‘https://e-

cab.net/company-proﬁle/<member id>’, where

<member id> refers to the numbers ‘0001’ to ‘2238’

which are distinct member id numbers associated

with each e-CAB member companies. We used the

Javascript library ‘puppeteer’ to crawl the ‘company

name’ and ‘website URL’ data for each e-CAB mem-

ber company from their associated member proﬁle

pages using this general URL pattern. As some

member proﬁle pages were unavailable, we were

able to collect data for 2,216 member companies of

e-CAB. Two exclusion criteria were applied to these

2,216 instances of data - (1) 104 companies were

removed because their associated ‘website URL’ data

contained URLs of Facebook pages, not of company

websites, and (2) 57 companies were removed for

not having any speciﬁed ‘website URL’ at all. After

excluding these companies, we considered a total of

2,055 e-commerce websites for further inspection.

3.2 Crawling Website Data

We crawled the e-commerce websites to extract text

segments from them. We ﬁrst manually examined

all the websites to collect and store relevant URLs,

followed by automatically crawling the HTML con-

tents from those URLs. Subsequently, we applied a

page segmentation algorithm to the HTML contents

to identify text segments for further inspection.

3.2.1 URL Retrieval

We tried to manually visit the home or landing page

of all of the 2,055 company websites. However, 650

websites were unreachable. We inspected the remain-

ing 1,405 websites to collect two additional URLs

alongside the home page URL from each website. For

product-based companies, we navigated to one cate-

gory page and one product page per website. And for

service-based companies, we navigated to the offered

https://e-cab.net

services page and the portfolio page. In this way, we

manually inspected all the websites and tried to col-

lect URLs of the home page, one category/services

page, and one product/portfolio page. Two exclusion

criteria were applied - (1) websites that were static

pages with no further navigation available, and (2)

websites that lacked relevant information. These ex-

clusion criteria reduced the number of websites to 715

and 2,119 URLs were collected from these websites.

3.2.2 HTML Data Retrieval

We again used the Javascript library ‘puppeteer’ to

crawl the 2,119 URLs and retrieve the whole HTML

content associated with each URL. Some websites de-

tected our ‘puppeteer’ crawler as a bot and restricted it

from accessing the original HTML content. We used

the ‘puppeteer-extra-plugin-stealth’ library to bypass

this bot detection. This library conceals certain prop-

erties from outgoing HTTP requests that can ﬂag the

requests as coming from a bot. HTML contents of 51

URLs could not be retrieved because of connection

and navigation problems. We were able to collect the

HTML contents of the remaining 2,068 URLs.

3.2.3 Candidate Dark Pattern Extraction

Text segments need to be extracted from HTML con-

tents that we call ‘Candidate Dark Patterns’. For this

purpose, we needed a segmentation algorithm to ana-

lyze webpages and identify candidate dark patterns.

Mathur et al. proposed such a segmentation algo-

rithm, where segments are deﬁned as visible HTML

elements that contain no other block-level elements

and at least one text element (Mathur et al., 2019).

Using this algorithm led to the discovery of very few

dark patterns in the HTML contents. We explored the

causes and found that the algorithm was too restric-

tive, discarding potential dark pattern texts.

We modiﬁed the segmentation algorithm to make

the inclusion criteria as relaxed as possible. We de-

ﬁned segments as any visible HTML element that is

a leaf (i.e., does not contain any child elements) and

contains text content. However, we exclude paragraph

elements from this deﬁnition, as they usually repre-

sent longer blocks of descriptive text that are not rele-

vant to our analysis. Details of our page segmentation

algorithm are illustrated in Algorithm 1.

This version of the segmentation algorithm re-

sulted in the discovery of both, increased candidate

dark patterns and subsequently increased dark pat-

terns. So we used this algorithm to extract candidate

dark pattern texts from the HTML contents obtained

in the previous step. 770,570 texts was extracted from

the 2,068 HTML contents under analysis.

ENASE 2024 - 19th International Conference on Evaluation of Novel Approaches to Software Engineering

242

Algorithm 1: Page Segmentation.

Input: HTML Document

Output: List of Text Segments

1 Function extractTextSegments(page):

2 ignoredElements ← [“p”, “script”,

“style”, “noscript”, “br”, “hr”];

3 segmentsList ← [];

4 foreach element in page do

5 if element is not in ignoredElements

and element is visible and element

does not contain any child elements

then

6 innerText ← element.innerText;

7 segmentsList.append(innerText);

8 end

9 end

10 return segmentsList;

3.3 Detection and Classiﬁcation

It is a time-consuming task to detect dark patterns

by manually reviewing the large set of 770,570 texts.

Thus, we employed a combination of automated and

manual approaches to detect and classify dark pattern

texts from the candidate texts. Dark pattern texts were

ﬁrst detected using an automated approach leverag-

ing machine learning, followed by a manual review

of each text and associated URLs to remove the false

positives and classify the remaining texts into differ-

ent dark pattern categories.

3.3.1 Data Preprocessing

We removed candidate dark pattern texts that con-

sisted of only numbers or a single word. We also re-

moved price tag texts containing the local currency

sign, using regular expressions. This preprocessing

reduced the set of candidate dark pattern texts by 43%

to 435,536 texts. Removing the duplicates would re-

duce the set even further. However, multiple web-

sites or multiple webpages within the same website

can contain the same dark pattern text. Thus, remov-

ing duplicates would deny us the real scenario of dark

patterns in the e-commerce websites of Bangladesh.

As a result, we opted not to remove the duplicates in

this step.

3.3.2 Automated Dark Pattern Detection

We used machine learning to automatically classify

the candidate texts as ‘dark pattern’ or ‘not dark pat-

tern’. Yada et al. provided some baseline machine

learning models that could be used for this purpose

(Yada et al., 2022). Initially, we used their best-

performing model (RoBERTa

large

), trained on 1,178

dark pattern and the same number of non dark pattern

texts provided in their dataset. The hyper-parameters

were set according to their work. However, this ap-

proach resulted in a large number of false positives.

As a result, the subsequent manual veriﬁcation would

be time-consuming, undermining the purpose of auto-

mated detection. Upon inspecting for causes behind

high false positives, we found the model to be vulner-

able to data overﬁtting.

Our previous work developed an automated detec-

tion technique that could detect dark patterns without

the need for training data, thus eliminating the con-

cern regarding data overﬁtting (Sazid et al., 2023).

This approach utilized the in-context learning capa-

bilities of GPT-3 to detect dark pattern texts. The best-

performing model from this work was the few-shot

model where two example texts per category were

used along with the category deﬁnitions to engineer

prompts for GPT-3. This approach was able to detect

six common categories of dark pattern texts success-

fully - ‘Misdirection’, ‘Urgency’, ‘Scarcity’, ‘Social

Proof’, ‘Obstruction’, and ‘Forced Action’. As the

approach using in-context learning resulted in a very

low rate of false negatives (under 10%) in our previ-

ous work, we reused this approach to detect dark pat-

tern texts from the candidate texts obtained in the pre-

vious step. To optimize resource usage, we only con-

sidered the unique texts in this step. We also treated

texts with numerical variations (e.g., ‘Only 4 left in

stock’ and ‘Only 1 left in stock’) as duplicates, en-

suring that similar texts are not processed by GPT-

3 separately. There were 91,585 such unique texts

among the set of 435,536 candidate texts. Our auto-

mated approach detected 2,185 out of those 91,585

texts as dark patterns.

3.3.3 Manual Dark Pattern Classiﬁcation

We manually reviewed each of the 2,185 dark pattern

texts to check whether they were really instances of

dark patterns or not. We discarded 2,086 instances

as they did not belong to any dark pattern category

considered in this study. We reviewed each of the re-

maining 99 instances to classify them into the six dark

pattern categories. During the review, if the textual

data was insufﬁcient for labelling, we manually vis-

ited the associated URL to gather context about the

text. This resulted in a dataset of dark patterns con-

taining 99 unique text instances labelled with their re-

spective dark pattern categories. As we only consid-

ered the unique texts in the previous step, we revis-

ited the set of 435,536 candidate texts to extract and

label the duplicates matching the 99 unique dark pat-

tern texts. Numerical variations were also treated as

Prevalence and User Perception of Dark Patterns: A Case Study on E-Commerce Websites of Bangladesh

243

similar texts in the previous step. Thus, we labelled all

such groups of texts with the same dark pattern cate-

gory. After considering the duplicates and numerical

variations, the ﬁnal set of dark patterns consisted of

931 instances.

4 EMPIRICAL STUDY ON USER

PERCEPTION

We conducted an empirical study to analyze user per-

ception regarding the presence of dark patterns in e-

commerce websites of Bangladesh. The goal was to

analyze and rank the dark pattern categories accord-

ing to users’ exposure, awareness about inﬂuence, and

concern. We also wanted to see if there is any differ-

ence in user perception based on whether they had an

educational background in technology or not.

4.1 Study Design

An open survey was conducted, directed primar-

ily at Bangladeshi university students. Participants

were ﬁrst asked about their general understanding

of the ability of user interface designs to inﬂuence

user’s decision-making, followed by the frequency of

their usage of e-commerce websites and applications.

Later, participants were presented with six user inter-

face design instances containing dark patterns from

the six categories mentioned in the previous section.

Presenting only dark pattern instances could bias the

participants in favour of a negative perception regard-

ing the design instances. To counteract that possibil-

ity, four non dark pattern design instances were also

presented in the survey. The ten user interface de-

sign instances presented in this survey are illustrated

in Figure 2. Participants were presented with these

design instances in a random order and three ratings

were measured for each design. ‘Exposure’ rating

was measured based on whether the participant had

ever encountered such a design or not (0 = never en-

countered, 1 = encountered). To measure ‘Awareness’

and ‘Concern’ ratings, the following two statements

were presented respectively:

• This kind of design can inﬂuence users to do

something that users normally would not do.

• I am worried about the inﬂuence of this kind of

design on users’ decision-making and choices.

Participants were instructed to rate their agree-

ment on a 5-point Likert scale (from 1 = strongly dis-

agree to 5 = strongly agree). In addition, demographic

information (age, sex, and educational background)

was collected from the participants. All questions in

the survey were mandatory.

4.2 Participants

We collected responses from 68 participants. The

demographics of the participants were as follows:

45 male, 23 female. Their age ranged from 19 to

27 years (mean 22.57, SD 1.47). Participants were

divided into two groups based on their educational

background - 36 participants had an educational back-

ground in computer science or related ﬁelds and the

other 32 had no such background in technology edu-

cation. Our thinking behind such grouping is that par-

ticipants with an educational background in technol-

ogy might have a different perception regarding dark

patterns than the ones without such a background.

5 RESULTS

This section analyzes and discusses our ﬁndings

regarding the prevalence of dark patterns in e-

commerce websites of Bangladesh as well as the user

perception survey.

5.1 Prevalence of Dark Patterns

We analyzed the collected 931 dark pattern texts

to showcase the prevalence of dark patterns in e-

commerce websites of Bangladesh. These dark pat-

tern texts were found across 131 websites, consti-

tuting approximately 18.3% of the 715 analyzed e-

commerce websites. This is a much higher rate com-

pared to the work of Mathur et al., who found dark

patterns in 11.1% of their analyzed websites (Mathur

et al., 2019). The increased rate can be credited to

our use of a more relaxed page segmentation algo-

rithm than theirs, resulting in improved detection of

dark patterns. It is important to note that we exclu-

sively analyzed the home page, one category/services

page, and one product/portfolio page per website and

our analysis only focused on the texts of these pages.

Given the scope of our exploration, this ﬁnding serves

as a conservative estimate. As a result, it can be said

that the actual prevalence of dark patterns in the e-

commerce websites of Bangladesh is even higher.

Figure 3 illustrates the frequency of each dark pat-

tern category. ‘Social Proof’ and ‘Urgency’ had the

highest number of instances among all the categories

with 466 and 271 occurrences respectively. 165 in-

stances belonged to the ‘Scarcity’ category while

‘Forced Action’ and ‘Misdirection’ had 23 and 6 oc-

currences respectively.

ENASE 2024 - 19th International Conference on Evaluation of Novel Approaches to Software Engineering

244

(a) Forced Action. (b) Misdirection.

(e) Scarcity. (f) Urgency.

(g) Not Dark Pattern. (h) Not Dark Pattern.

(i) Not Dark Pattern. (j) Not Dark Pattern.

Figure 2: User Interface Design Instances Used in This Study.

Prevalence and User Perception of Dark Patterns: A Case Study on E-Commerce Websites of Bangladesh

245

Figure 3: Instances Across DP Categories.

Figure 4 illustrates the number of unique instances

across each dark pattern category. ‘Urgency’ tops

the other categories in this regard, with 41 unique

instances. ‘Social Proof’ and ‘Scarcity’ had 26 and

24 unique instances respectively while ‘Forced Ac-

tion’ and ‘Misdirection’ had only 6 and 2 respec-

tively. No instance of the ‘Obstruction’ category was

found in the study. We compared this prevalence of

unique dark pattern texts across categories with the

ﬁndings of Mathur et al., who also gathered unique

instances of dark pattern texts in their work (Mathur

et al., 2019). While the rankings of dark pattern cat-

egories in terms of prevalence do not perfectly align

between the two studies, two distinct groups emerge.

Both analyses indicate that ‘Urgency’, ‘Social Proof’,

and ‘Scarcity’ (Group A) had the highest number

of unique instances, while ‘Forced Action’, ‘Misdi-

rection’, and ‘Obstruction’ (Group B) had relatively

fewer unique instances. As a result, we can say that

Group A dark patterns were more prevalent in our

analysis than Group B dark patterns.

Figure 5 also conﬁrms such a grouping of dark

pattern categories in terms of prevalence across web-

sites. ‘Urgency’, ‘Scarcity’, and ‘Social Proof’ had

the more prominent presence, with occurrences on

78, 28, and 26 websites respectively. ‘Forced Ac-

tion’, ‘Misdirection’, and ‘Obstruction’ were much

less prominent, with occurrences on only 7, 2, and

0 websites respectively. Interestingly, 121 of the 131

websites were associated with a single dark pattern

category, with the remaining 10 being associated with

at most two dark pattern categories.

Figure 6 illustrates the frequency of dark patterns

across the 131 websites. Most of the websites present

only 1, 2, or 3 unique instances of dark patterns,

whereas very few websites present larger quantities.

The highest number of unique dark pattern texts on a

single website is 10.

Figure 4: Unique Instances Across DP Categories.

Figure 5: Websites Across DP Categories.

Figure 6: Unique Instances Across Websites.

5.2 User Perception of Dark Patterns

Based on the survey responses, we ranked and

grouped the dark pattern categories in terms of ex-

posure, awareness, and concern. We also carried out

Mann-Whitney U test to see if there was any differ-

ence in user perception based on educational back-

ground in technology.

ENASE 2024 - 19th International Conference on Evaluation of Novel Approaches to Software Engineering

246

Table 1: Dark Pattern Category Rankings Based on User Perception.

Rank

Exposure Awareness Concern

DP Category Mean Rating (%) DP Category Mean Rating (1 to 5) DP Category Mean Rating (1 to 5)

1 Forced Action 94.12 Urgency 4.21 Obstruction 3.81

2 Urgency 88.24 Scarcity 3.99 Forced Action 3.78

3 Scarcity 86.76 Social Proof 3.96 Urgency 3.56

4 Social Proof 77.94 Obstruction 3.96 Scarcity 3.44

5 Misdirection 73.53 Forced Action 3.81 Social Proof 3.22

6 Obstruction 30.88 Misdirection 3.65 Misdirection 3.17

5.2.1 Ranking Dark Pattern Categories

Survey data was analyzed to rank different categories

of dark patterns based on users’ ratings of exposure,

awareness of their inﬂuence, and concern regarding

them. The rankings were determined by calculating

the mean ratings of exposure, awareness, and con-

cern for each dark pattern category by averaging all

participant responses. The resulting rankings are pro-

vided in Table 1. Notably, participants consistently

rated their awareness and concern regarding dark pat-

terns above the neutral value of 3, indicating a gen-

eral awareness of the potential inﬂuences of dark pat-

tern design instances and signiﬁcant concern about

them. However, the category rankings of awareness

and concern do not perfectly align with each other,

suggesting that participants were not necessarily more

concerned about a design simply because they per-

ceived it to be more inﬂuential, and vice versa.

5.2.2 Grouping Dark Pattern Categories

Exposure, awareness, and concern ratings produce

further evidence in favour of our grouping of dark

pattern categories in the previous section. There is a

clear distinction in the ratings between the two groups

of dark patterns. Participants were more exposed

to Group A dark patterns than Group B, except for

‘Forced Action’. Such a contrast in the exposure rat-

ings makes more sense as they align directly with the

prevalence of these two groups explained in the pre-

vious section. Participants found Group A containing

‘Urgency’, ‘Scarcity’, and ‘Social Proof’ categories

more inﬂuential than Group B containing ‘Misdirec-

tion’, ‘Obstruction’, and ‘Forced Action’. The reverse

tendency is found in concern ratings as participants

found Group B more concerning than Group A, ex-

cept for ‘Misdirection’.

The contrast between awareness and concern rat-

ings across the two groups of dark pattern categories

can be explained by how these design instances man-

ifest themselves in user interfaces. Group A dark pat-

terns get users’ attention to certain actions using the

attraction of limited sales or discounts, and the rar-

ity or credibility of a product. Even though they can

create a sense of hurry among users, the user expe-

rience is not necessarily unpleasant, resulting in less

concern. On the other hand, Group B dark patterns

make users feel uncomfortable as they get controlled,

forced, or deceived by the user interface. As a result,

users are more concerned about this group of dark pat-

tern categories. According to this ﬁnding, we name

Group A dark patterns as ‘Passive Dark Patterns’ and

Group B dark patterns as ‘Active Dark Patterns’:

Passive Dark Patterns: This group of dark patterns

consists of ‘Urgency’, ‘Social Proof’, and ‘Scarcity’

categories. All of these techniques use users’ ‘fear of

missing out’ (FOMO) and create a sense of attraction

in users to take certain actions. They rely on social

conformity and the perception of limited availabil-

ity (time or quantity) to motivate users to take action

quickly. The main identifying factor of these dark pat-

terns is that they entice and attract users, rather than

forcing them. Because of their subtle presentation in

user interfaces, they do not seem abnormal to users,

even when users think they have a high inﬂuence on

their decision-making. As passive dark patterns have

been ‘normalized’ to users, ﬁnding countermeasures

for this group of dark patterns is more difﬁcult and

necessitates legal and policy conversations regarding

marketing and advertisement strategies in general.

Active Dark Patterns: This group consists of

‘Forced Action’, ‘Misdirection’, and ‘Obstruction’

categories. These techniques manipulate user behav-

ior through deception, obstacles, or coercion. These

design instances forcefully drive users to act in ways

that are against their wishes or interests. The main

identifying factor of these dark patterns is that they

mislead or force users to make certain choices. As a

result, users often feel uncomfortable while encoun-

tering these dark patterns. These techniques under-

mine user autonomy and can lead to a negative user

Prevalence and User Perception of Dark Patterns: A Case Study on E-Commerce Websites of Bangladesh

247

Table 2: Results of Mann-Whitney U Test (n1=36, n2=32).

Variable Mann-Whitney U Statistic P Value H

Rejected

Exposure 614 0.3184 No

Awareness 790 0.0042 Yes

Concern 739 0.0225 Yes

experience. Thus, this group of dark patterns does not

fall into traditional marketing or advertisement strate-

gies. Legal and policy regulations over design princi-

ples of online platforms can have a signiﬁcant impact

on counteracting these dark patterns.

Two deviations are visible in Table 1 from our

analysis regarding the grouping of dark pattern cat-

egories - ‘Forced Action’ in the case of exposure rat-

ings and ‘Misdirection’ in the case of concern ratings.

We argue that these deviations are due to participant

bias or misunderstanding about the particular design

instances used in both cases. As shown in Figure 2a,

participants were presented with a cookie banner fea-

turing solely an ‘Accept’ button without a correspond-

ing ‘Reject’ button, as an example of ‘Forced Ac-

tion’. The heightened frequency of responses indicat-

ing high exposure to ‘Forced Action’ could stem from

participants’ familiarity with cookie banners rather

than the dark pattern design featuring only an ‘Ac-

cept’ button and lacking a ‘Reject’ button. As shown

in Figure 2b, participants were presented with an in-

staller interface containing two choices where one

was recommended and the other was blurred, as an

example of ‘Misdirection’. Participants could have

perceived such a recommendation as a helpful thing

and not considered situations where such preselection

and design interference could go against their inter-

ests; leading to less concern about this category.

5.2.3 Impact of Technology Education on User

Perception of Dark Patterns

Mann-Whitney U test was used to see if there was

any difference in user perception between participants

who had an educational background in technology

and participants who did not have such a background.

For this purpose, ratings of exposure, awareness, and

concern across all dark pattern categories were aver-

aged to get mean exposure, awareness, and concern

ratings for each participant, which were the three vari-

ables used in hypothesis testing. For each one of

the three variables, the null and alternative hypothe-

ses were as follows (CS = participants with an educa-

tional background in technology, NCS = participants

with no educational background in technology):

• H

: Ratings from CS participants are lower than

or equal to that of NCS participants

• H

: Ratings from CS participants are greater than

that of NCS participants

Table 2 presents the results from the one-tailed

Mann-Whitney U test (level of signiﬁcance, α =

0.05). While the null hypothesis could not be rejected

for exposure, it was rejected for awareness and con-

cern; indicating statistically signiﬁcant evidence that

participants with an educational background in tech-

nology are more aware and concerned about dark pat-

terns than participants without such a background.

6 THREATS TO VALIDITY

This section discusses the potential threats that may

affect the validity of the study.

Threats to External Validity: The selection of the

e-CAB members’ list as the sampling frame poses an

external validity threat as not all e-commerce web-

sites of Bangladesh are members of e-CAB. However,

this is the largest possible representative sample of e-

commerce websites originated in Bangladesh. Ana-

lyzing a few URLs per website also poses a threat, as

other URLs may produce different results. However,

e-commerce websites generally have similar user in-

terface designs across category/services pages and

product/portfolio pages. That is why we considered

one instance of each of these webpage types, along

with the homepage. The survey participants in our

empirical study were university students. Thus, the

ﬁndings may differ for other age groups and educa-

tional levels.

Threats to Internal Validity: Survey participants

could have made assumptions about which responses

were socially desirable for this study, posing a threat

to the internal validity. However, we mitigated this

bias by communicating to the participants that there

were no ‘right’ answers. Random ordering of the user

interface design instances also poses a threat to the

internal validity.

Threats to Construct Validity: The automated de-

tection and manual classiﬁcation techniques used in

the study may affect the construct validity. Our au-

tomated detection approach may have missed some

dark pattern data as false negatives. However, we car-

ried out the manual classiﬁcation with great care to re-

duce the remaining threats as much as possible. There

would always be a possibility for survey participants

to misunderstand questions or appropriate contexts.

To mitigate this threat, we carried out a physical sur-

vey instead of a remote one. As a result, we were able

to provide clear explanations and instructions about

the survey to the participants.

ENASE 2024 - 19th International Conference on Evaluation of Novel Approaches to Software Engineering

248

7 CONCLUSION

This study marks the ﬁrst exploration into dark pat-

terns in the local context of Bangladesh. We com-

bined automated and manual methods to collect tex-

tual dark pattern data from the websites of mem-

ber companies of the E-Commerce Association of

Bangladesh (e-CAB). We analyzed 715 websites and

exposed dark patterns in 18.3% of those websites,

which is 7.2% higher than in previous research us-

ing a similar approach. We also surveyed 68 univer-

sity students about their perception of dark patterns.

Based on our ﬁndings from both explorations, we di-

vided dark pattern categories into two distinct groups

- ‘Passive Dark Patterns’ and ‘Active Dark Patterns’.

Most of the dark pattern instances found in this study

belonged to ‘Passive Dark Patterns’. Our analysis also

revealed that users with a background in technology

education are more aware and concerned about dark

patterns than other users. Further research is needed

to analyze the reasons and consequences of dark pat-

terns, local developer perspectives, and potential reg-

ulatory frameworks in the context of Bangladesh.

ACKNOWLEDGEMENTS

This study is supported by the fellowship from the

ICT Division, Government of Bangladesh – No.:

56.00.0000.052.33.001.23-09, date: 04.02.2024.

REFERENCES

Bongard-Blanchy, K., Rossi, A., Rivas, S., Doublet, S.,

Koenig, V., and Lenzini, G. (2021). ” i am deﬁ-

nitely manipulated, even when i am aware of it. it’s

ridiculous!”-dark patterns from the end-user perspec-

tive. In Designing Interactive Systems Conference

2021, pages 763–776.

osch, C., Erb, B., Kargl, F., Kopp, H., and Pfattheicher, S.

(2016). Tales from the dark side: privacy dark strate-

gies and privacy dark patterns. Proc. Priv. Enhancing

Technol., 2016(4):237–254.

Brignull, H., Leiser, M., Santos, C., and Doshi, K. (2023).

Deceptive patterns – user interfaces designed to trick

you. https://www.deceptive.design/.

Chaudhary, A., Saroha, J., Monteiro, K., Forbes, A. G., and

Parnami, A. (2022). “are you still watching?”: Ex-

ploring unintended user behaviors and dark patterns

on video streaming platforms. In Designing Interac-

tive Systems Conference. ACM.

Geronimo, L. D., Braz, L., Fregnan, E., Palomba, F., and

Bacchelli, A. (2020). UI dark patterns and where to

ﬁnd them. In Proceedings of the 2020 CHI Conference

on Human Factors in Computing Systems. ACM.

Gray, C. M., Chen, J., Chivukula, S. S., and Qu, L. (2021).

End user accounts of dark patterns as felt manipula-

tion. Proceedings of the ACM on Human-Computer

Interaction, 5(CSCW2):1–25.

Gray, C. M., Kou, Y., Battles, B., Hoggatt, J., and Toombs,

A. L. (2018). The dark (patterns) side of UX design. In

Proceedings of the 2018 CHI Conference on Human

Factors in Computing Systems. ACM.

Hidaka, S., Kobuki, S., Watanabe, M., and Seaborn, K.

(2023). Linguistic dead-ends and alphabet soup: Find-

ing dark patterns in japanese apps. In Proceedings of

the 2023 CHI Conference on Human Factors in Com-

puting Systems, pages 1–13.

Jarovsky, L. (2022). Dark patterns in personal data col-

lection: Deﬁnition, taxonomy and lawfulness. SSRN

Electronic Journal.

Luguri, J. and Strahilevitz, L. (2019). Shining a light on

dark patterns. SSRN Electronic Journal.

Maier, M. (2019). Dark patterns–an end user perspective.

Mansur, S. M. H., Salma, S., Awoﬁsayo, D., and Moran,

K. (2023). AidUI: Toward automated recognition of

dark patterns in user interfaces. In 2023 IEEE/ACM

45th International Conference on Software Engineer-

ing (ICSE). IEEE.

Mathur, A., Acar, G., Friedman, M. J., Lucherini, E.,

Mayer, J., Chetty, M., and Narayanan, A. (2019). Dark

patterns at scale. Proceedings of the ACM on Human-

Computer Interaction, 3(CSCW):1–32.

Mathur, A., Kshirsagar, M., and Mayer, J. (2021). What

makes a dark pattern... dark? In Proceedings of the

2021 CHI Conference on Human Factors in Comput-

ing Systems. ACM.

Roffarello, A. M. and Russis, L. D. (2022). Towards un-

derstanding the dark patterns that steal our attention.

In CHI Conference on Human Factors in Computing

Systems Extended Abstracts. ACM.

Sazid, Y., Fuad, M. M. N., and Sakib, K. (2023). Automated

detection of dark patterns using in-context learning ca-

pabilities of gpt-3. In 2023 30th Asia-Paciﬁc Soft-

ware Engineering Conference (APSEC), pages 569–

573. IEEE.

Soe, T. H., Nordberg, O. E., Guribye, F., and Slavkovik,

M. (2020). Circumvention by design-dark patterns

in cookie consent for online news outlets. In Pro-

ceedings of the 11th nordic conference on human-

computer interaction: Shaping experiences, shaping

society, pages 1–12.

Soe, T. H., Santos, C. T., and Slavkovik, M. (2022). Au-

tomated detection of dark patterns in cookie banners:

how to do it poorly and why it is hard to do it any other

way. arXiv preprint arXiv:2204.11836.

Stavrakakis, I., Curley, A., O’Sullivan, D., Gordon, D., and

Tierney, B. (2021). A framework of web-based dark

patterns that can be detected manually or automati-

cally.

Yada, Y., Feng, J., Matsumoto, T., Fukushima, N., Kido, F.,

and Yamana, H. (2022). Dark patterns in e-commerce:

a dataset and its baseline evaluations. In 2022 IEEE

International Conference on Big Data (Big Data).

IEEE.

Prevalence and User Perception of Dark Patterns: A Case Study on E-Commerce Websites of Bangladesh

249