Deceptive Patterns · Case Study
750
Component
Timeline
Nov 2025 – May 2026
Type
HCI Research
Role
UX Research | UX Design
Interface designs that steer users toward choices they did not intend to make. [1]
UI that lets people interact with computers using natural human language via text or voice.

Deceptive Design is now happening through Conversational UI, too.

For my Master Thesis in HCI, I did a 7‑month research on Conversational UI (typically: chatbots) vs. Graphical UI (typically: websites), exploring how users trust these modalities and how they react to deceptive design in each, in the domain of eCommerce.
How Significant?

Everywhere and illegal.

The numbers tell the story quickly:

97%
of top EU sites use deceptive design [2]
1,818
instances across a crawl of 11K sites [3]
$72B
projected chatbot retail by 2028 [4]
The $72B figure led me to focus on e‑commerce. Beyond its scale, financial contexts matter because users are more cautious when money is involved, and their trust in conversational UIs is especially fragile in those situations.
Deceptive design example:
GUI: Amazon Prime “Iliad” flow (FTC v. Amazon, 2025) → 4-page, 6-click, 15-option cancellation.
CUI: SiriusXM chat transcripts → 28-minute cancellation with repeated retention offers (NY AG, 2023; user complaints).
The research gap
Venn diagram: GUI research on the left, CUI research on the right, and the unstudied comparison of deceptive design in the overlap, where this study sits.
Main research question

How do Conversational UIs shape users’ detection, trust, and behavior when facing deceptive design, compared to traditional GUIs?

Deep Dive

User Perception in CUI

vs GUI.
  • CASA Paradigm [5]
  • Gricean Pragmatics [6]
  • Social Presence & Reciprocity [7]
  • Anthropomorphism & Warmth [8]
  • Sequential Framing [9]
  • Trust & Attribution [10]
  • Politeness Norms [11]
  • “People treat chatbots like humans.”
  • “Users assume conversation is complete & truthful.”
  • “Chat feels social, triggers obligation.”
  • “Human‑like cues enable subtle manipulation.”
  • “Step‑by‑step hides info, raises cognitive load.”
  • “Users blame the AI, not just the company.”
  • “Harder to ‘exit’ a chatbot without feeling rude.”
Study design

User Study Process

User study process Interviews Survey Scenarios Research 40+ Articles 16 Scenarios 22 Participants 3 Interviews

Why these 4 deceptive patterns?

Filters applied
5 strategy categories from Gray [12]
Single-scenario operationalisable
Distinct psychological mechanism
Documented in both GUI and CUI
Deceptive patterns

Obstruction: making a task easy to get into, but hard to get out of.

Confirmshaming: using shame language to steer users away from opting out.

Forced Action: requiring information or actions not necessary for the task.

Sneaking: hiding, disguising, or delaying disclosure of relevant information.

After coming up with the 4 deceptive patterns I would study, I collected specific examples of them happening to real users, and mapped identical scenarios after them, but under a fake brand name. You can see all of the scenarios below. I also created the neutral versions of them: the same eCommerce scenario, without the deceptive pattern.
Study design: 4 patterns times 2 modalities times 2 conditions equals 16 scenarios, each participant views all 16 within-subjects, feeding a Phase 1 survey (N=22) and Phase 2 interviews (N=3), integrated at interpretation.
Example

Real Scenarios

GUI dark: StreamPlus obstruction flow, 4 steps from account page to a hard-to-leave cancellation flow.
GUI neutral: FreshBox account and cancellation flow, the same task without the deceptive pattern.
CUI neutral: CloudVault support chat, a straightforward cancellation conversation.
CUI dark: MusicRadio Pro support chat, a cancellation conversation with repeated retention attempts.
Scenario screens coming soon.
Scenario screens coming soon.
Scenario screens coming soon.
Self-Hosted HTML Survey File

Survey Flow

Consent
Demographics
16 Scenarios
Debrief & Re‑Rate
Finish
Optional Interview
Methodology

Data Analysis

Quantitative
  • UEQ scoring
  • Likert‑scale ratings
  • Descriptive statistics
  • Inferential statistics
Qualitative
  • Open coding & taxonomy building
  • Thematic analysis of participant comments
  • Interpretive comparison of GUI vs CUI mockups
  • Interpretation of user perceptions (e.g., “rudeness,” “betrayal”)
Cross‑Referencing
  • Literature synthesis (merging prior dark pattern taxonomies)
  • Survey vs. user‑study comparison
  • Pre‑ vs. post‑debrief recognition check
  • GUI vs. CUI modality comparison
Findings

The modality effect is real.
It just isn’t one effect.

Amplified

Obstruction + Sneaking

CUI makes it feel worse. The trust violation feels personal.

100% detected in CUI

Invisible

Forced Action

CUI hides the pattern. The steps feel like guidance.

74% GUI vs 25% CUI

Independent

Confirmshaming

Modality makes no difference. Shame is already linguistic.

~27% in both

Design Implications

Transparency in CUIs

Disclosures need to be explicit, not buried across conversational turns the way a chatbot can bury them.

Limit anthropomorphic cues

Warmth and friendliness in a chatbot's voice can mask manipulation that would be obvious on a screen.

Sequential awareness

Give users the “big picture” of a choice, not just the one turn they're currently inside.

Exit clarity

Leaving a conversation should feel as neutral as closing a tab, without the social pressure of politeness norms.

Policy Implications

CUI‑specific standards

GUI‑based dark pattern definitions miss vulnerabilities that only exist in a conversational flow.

Disclosure requirements

Costs, data use, and opt‑outs need to be stated upfront in a CUI, not surfaced only if you ask.

Accountability

When a chatbot manipulates someone, is the company responsible, or the agent acting on its behalf?

Financial contexts

Trust in a conversational agent is most fragile in money‑related interactions, so protections should be strongest there.

Component