[01]Summary
I designed the protocol, ran the data collection and analysed the results: 78 students from three programmes in Belfort (Info and EDIM at UTBM, HSE for Hygiene, Safety, Environment, outside UTBM), randomly assigned, stratified by programme, to four conditions of 18 to 21 people.
In our sample, the submission rate goes from 5.6% to 26.3% with the Dark Pattern, a clear trend that still needs to be confirmed on a larger sample.
Hypocrisy induction, meant to protect, did not reduce justifications: if anything, we observe the opposite. All results are presented as exploratory trends.
[02]The question
Dark Patterns (deceptive design) are interfaces built to steer users' choices against their own interest: artificial urgency, the safe option made discreet, costs left unsaid.
Do these interfaces push students to give up their data for less than the price they set themselves, and can an intervention based on cognitive dissonance protect them?
[03]Theoretical framework
Four notions are enough to read the experiment. References point to the bibliography (section 08) and to pages of the thesis.
ELM and the peripheral route
According to the Elaboration Likelihood Model (Petty and Cacioppo), persuasion follows either a central route, which requires analysing arguments, or a peripheral route, where people rely on mental shortcuts. Under time pressure or overload, the peripheral route dominates: a badge, a colour or social proof weigh more than arguments. This is the entry point of Dark Patterns.
Sourcesthesis p. 7
Cognitive dissonance
Theorised by Leon Festinger, dissonance is the discomfort felt when an action contradicts one's values or beliefs. To reduce it, people tend to rationalise: they adjust their attitude to fit what they just did. The experiment tries to provoke dissonance in order to block this self-justification, then measures what remains of it.
SourcesSimplyPsychology (2023) • Courbet et al. (2013) • thesis p. 8 and 9
Induced hypocrisy
This paradigm, developed by Elliot Aronson, first asks people to publicly defend a norm (for example the importance of strong passwords), then confronts them with their own recent shortcomings. The gap between words and deeds is expected to create strong dissonance, which pushes people to align their future behaviour with the norm.
SourcesStone and Fernandez (2008) • Ryoo and Kim (2024) • thesis p. 10 and 14 to 16
Neutralisation techniques
These are the excuses people give themselves to justify risky behaviour: denial of injury, condemnation, defence of necessity ("I had no time"), bypass ("I use a junk password"). Siponen, Puhakainen and Vance (2020) report, in a field study of 80 employees, a drop in 5 of the 8 techniques tested after a dissonance-based intervention. Our post-test measures them on 1 to 5 scales.
SourcesSiponen, Puhakainen and Vance (2020) • thesis p. 13 and 14
[04]The protocol
The protocol merges two methods from the literature: the Transaction Test of Bogliacino et al. (2023), which traps the user mathematically instead of asking whether they feel manipulated, and the dissonance intervention of Siponen, Puhakainen and Vance (2020). The scenario is a student sign-up on "Bons Plans UTBM", a fictional student-discounts app.
The scenario in five steps
1. Price elicitation
Participants state the minimum value, in gift cards per month, they would require to create an account in Express mode (weak security, data sharing): 2, 5, 10, 15 or 20 EUR, or a refusal. This value is recorded without their knowledge as a reference.
2. Treatment
All participants watch an official password-prevention video. Depending on the group, it is followed or not by a hypocrisy questionnaire, for example: "Have you ever used a very simple password because you were too lazy to remember a better one?".
3. Filler task
A short socio-demographic form (age, gender, semester, programme) occupies working memory, so that participants do not keep their price in mind.
4. The trap
Participants choose between the Secure Profile (no gain) and the Express Profile (+ 5 EUR immediately), shown in a neutral interface or in a Dark Pattern interface. Choosing the Express Profile while requiring more counts as an inconsistency.
5. Likert post-test
They rate on 1 to 5 scales how much they agree with justifications (neutralisation techniques): denial of injury, condemnation, defence of necessity, bypass.
The two interfaces of the trap
Same screen, two versions. The neutral interface shows two cards of equal weight, with costs and data sharing stated plainly. The Dark Pattern interface combines wording tricks (trick wording) and visual interference. The captures are those of the experiment (my own work); the texts stay in French, as they were for participants.
Marker legend
The numbered markers on the Dark Pattern capture point to these six elements.
- Marker 1:
- Interface element
- Badge "Offre Limitée - Recommandé" (limited offer, recommended)
- Mechanism
- Artificial urgency and scarcity, implicit recommendation
- In the neutral interface
- No badge
- Marker 2:
- Interface element
- "Profil Sécurisé" greyed out, low contrast
- Mechanism
- Visual hierarchy: the safe option looks disabled
- In the neutral interface
- Two cards of equal visual weight
- Marker 3:
- Interface element
- "Contrainte manuelle", "reconnexion forcée", "imposé" (manual constraint, forced reconnection, imposed)
- Mechanism
- Negative framing of security (trick wording)
- In the neutral interface
- "Création manuelle", "Reconnexion à chaque visite" (manual creation, reconnection at each visit)
- Marker 4:
- Interface element
- "+ 5 € CRÉDITÉS DIRECT" in green, every line ticked
- Mechanism
- Immediate gain highlighted, everything looks positive
- In the neutral interface
- "+ 5 € OFFERTS", costs marked with a cross
- Marker 5:
- Interface element
- "Expérience personnalisée (Synchronisation du réseau partenaire)", almost invisible and ticked
- Mechanism
- Data sharing hidden and euphemised
- In the neutral interface
- "Partage avec nos partenaires" (sharing with our partners) marked with a cross, explicit
- Marker 6:
- Interface element
- "Transmission de l'historique" (history transmission) has disappeared
- Mechanism
- Omission of a cost
- In the neutral interface
- "Exploitation (Transmission de l'historique)" marked with a cross
2×2 factorial design
Two crossed factors, the interface (neutral or Dark Pattern) and the treatment (neutral video or video followed by a hypocrisy questionnaire), give four groups assigned at random, stratified by programme. Group 1 is the reference; group 4 is the core of the study: does dissonance block the Dark Pattern? The rates shown are the results, detailed in section 05.
Panel breakdown
| Group | Info | EDIM | HSE | Total |
|---|---|---|---|---|
| 1 • Control | 4 | 8 | 6 | 18 |
| 2 • Dark Pattern | 4 | 8 | 7 | 19 |
| 3 • Dissonance | 4 | 8 | 9 | 21 |
| 4 • Dissonance + Dark Pattern | 4 | 8 | 8 | 20 |
| Total | 16 | 32 | 30 | 78 |
Variables
Independent variables
- IV 1 • student profile (programme)
- IV 2 • treatment: neutral video or induced dissonance
- IV 3 • final interface: neutral or Dark Pattern
Dependent variables
- DV 1 • inconsistency (binary, 0 or 1): does the student give up their data for 5 EUR despite their initial requirement?
- DV 2 • rationalisation (score from 1 to 5): agreement with "false excuses" measured with Likert scales
Hypotheses
They were written before data collection, in part IV of the thesis (p. 22 and 23). The statements of H2 to H4 follow its wording, abridged.
H1 • Programme specificity
At equal failure rates, Info students will justify their choice by technical bypass, EDIM students by ergonomic convenience.
H2 • The attack
The Dark Pattern interface (which combines emotional Trick Wording and visual interference) will significantly increase the rate of choosing the Express Profile (inconsistent choice) compared with the neutral interface.
H3 • The defence
The groups that underwent the induced hypocrisy treatment (groups 3 and 4) will show significantly lower scores for the use of neutralisation techniques (DV 2, measured in the post-test) than the control groups.
H4 • The mental shield
Group 4 (which faces the Dark Pattern attack after being ethically awakened) will show a significantly lower inconsistency rate than group 2 (which faces the Dark Pattern alone).
[05]Results
In our sample, the submission rate goes from 5.6% to 26.3% with the Dark Pattern, a clear trend that still needs to be confirmed on a larger sample. With about twenty people per group, every percentage point weighs heavily: everything below reads as an exploratory trend.
Submission rate to the trap
- Interface: we observe 5.6% in G1 and 26.3% in G2. Pearson's chi-squared test, without correction, on this comparison gives 2.93 and p = 0.087: not significant at the 0.05 threshold.
- Mental shield: with the Dark Pattern, the rate is 15% (3 out of 20) after hypocrisy induction (G4) against 26.3% (5 out of 19) without it (G2), about 11 points lower, a gap of 2 people. This goes in the direction of H4, without allowing us to assert it.
- G1 • neutral interface, neutral video (control)1 of 185.6%
- G2 • Dark Pattern, neutral video5 of 1926.3%
- G3 • neutral interface, video + hypocrisy questionnaire4 of 2119%
- G4 • Dark Pattern, video + hypocrisy questionnaire3 of 2015%
Justification profile
Justifications are measured on Likert scales from 1 to 5. The defence of necessity (lack of time) is the excuse used most by all students; it reaches 3.39/5 among groups 3 and 4 (value for groups 1 and 2 not published). The chart compares the Info and EDIM programmes.
- Info programme3.19/5
- EDIM programme2.81/5
The other techniques (denial of injury, condemnation) appear on the radar charts of the defence without numeric values: I only reuse the published values here.
Cross analysis: Info versus EDIM
In our sample (16 Info, 32 EDIM), technical profiles (Info) rationalise more overall than design profiles (EDIM) on every criterion. They are more ready to admit using a "junk password" (bypass: 2.56 against 2.28), while the defence of necessity remains the number one excuse for both profiles (3.19 and 2.81). The defence slides only compare Info and EDIM: the 30 HSE students are not part of this comparison.
On the value given to data, 6 Info students out of 16 (37.5%) refuse any price, against 3 out of 32 in EDIM, where the most frequent price is 15 EUR (12 out of 32). The two profiles do not treat data as a currency in the same way (defence slides, slide 14).
- Info programme2.56/5
- EDIM programme2.28/5
- Info programme4 of 1625%
- EDIM programme26 of 3281.25%
Hypotheses review
The defence rated H2 and H4 as "validated". With these sample sizes, I now read them as trends. H2: non-significant Pearson chi-squared test (p = 0.087); H4: a gap of 2 people (5/19 against 3/20), not tested.
- H1 • programmeCautious reading Partial trendDistinct rationalisation profiles, but a shared reliance on the defence of necessity.
- H2 • attackCautious reading Trend in the expected direction5.6% against 26.3%, but a non-significant Pearson chi-squared test (p = 0.087).
- H3 • defenceCautious reading Not supportedOpposite trend observed: rationalisation went up rather than down.
- H4 • mental shieldCautious reading Weak trend in the expected direction15% (3 out of 20) against 26.3% (5 out of 19) with the Dark Pattern, a gap of 2 people, to be confirmed.
[06]Limits and validity
Showing why this work is not conclusive is part of the result. Here is what bounds what can be said.
Very small samples
Each condition has 18 to 21 people. 26.3% is 5 people out of 19, and 5.6% is 1 person out of 18: a single participant moves the rate by about 5 points.
Tests were run, but their application conditions are not met
I ran the planned statistical tests. On the main comparison (G1 against G2), Pearson's chi-squared value, without correction, is 2.93 with p = 0.087, above the 0.05 threshold set in the thesis. Above all, sample sizes are too small for the application conditions of this test to be met (expected counts below 5 in several cells). Fisher's exact test is the appropriate test for counts this small. I computed it a posteriori, for this page, after the defence (it is not part of the original work): G1 against G2, p ≈ 0.18; G2 against G4, p ≈ 0.45 (two-sided). All results should therefore be read as exploratory trends, never as established effects.
A narrow sample
Students from three programmes in Belfort (Info and EDIM at UTBM, HSE outside UTBM), panels of unequal size (16, 32 and 30 people).
A simulated, self-reported setting
The app is fictional: the decision does not involve real personal data. Justifications are measured by questionnaire, hence exposed to social desirability bias (the tendency to rationalise one's choices after the fact).
An incomplete hypocrisy induction
The Aronson paradigm described in the theory section starts with a public commitment to the norm, before the confrontation with one's own shortcomings (thesis p. 10). In the experiment, the induction was limited to the questionnaire on transgressions, without this public commitment step (thesis p. 18). This may partly explain the opposite trend observed.
What it would take to conclude
Many more participants per condition, hypotheses and an analysis plan pre-registered before data collection, and a second field less close to the first. This is the natural next step of this work.
[07]What I take away as an engineer
Four design leads, phrased as working hypotheses and not as established rules.
Friction can push toward workarounds
Among EDIM students, 26 out of 32 (81.25%) say they would reuse an existing password for a student-app account, which suggests that a constraint that adds friction risks producing junk passwords; the challenge is security without friction.
Technical profiles are not immune
Info students rationalise more, and admit to the bypass more readily (2.56 against 2.28). A quarter of them (4 out of 16) would still reuse an existing password for a student-app account.
Blaming the user can backfire
In our sample, hypocrisy induction came with more justifications, not fewer. Ryoo and Kim (2024) describe a related mechanism: guilt pushes people to repair, shame pushes them to flee.
Visual balance is a security decision
In the Dark Pattern interface, the safe option is greyed out and costs are omitted or euphemised. As a developer, I see these as design parameters that can be audited: equal visual weight, explicit costs, neutral wording.
[08]Bibliography
References of the thesis (pages 26 and 27), with the links it provides. External links open in a new tab.
Online books
- DAFFIN, Lee W. Jr., LANE, Carrie. Principles of Social Psychology. 2nd ed. Washington State University, 2020, 238 p..Open the source
- JHANGIANI, Rajiv S., TARRY, Hammond. Principles of Social Psychology. 1st International H5P Edition. BCcampus, 2022, 736 p..Open the source
- STANGOR, Charles. Principles of Social Psychology. 1st International Edition. eCampusOntario, 2017, 620 p..Open the source
Online book chapters
- DALHOUSIE UNIVERSITY. What is Social Psychology? In: Introduction to Psychology & Neuroscience. Dalhousie University Libraries, 2020.Open the source
- COURBET, Didier, FOURQUET-COURBET, Marie-Pierre, BERNARD, Françoise et al. Communication persuasive et communication engageante pour la santé : Favoriser des comportements sains avec les médias, Internet et les serious games. In: BLANC, Nathalie (dir.). Publicité et Santé : des liaisons dangereuses ? Le point de vue de la psychologie. In Press, 2013, p. 15.Open the source
Online articles
- MATZ, Sandra C., KOSINSKI, Michal, NAVE, Gideon et al. Psychological targeting as an effective approach to digital mass persuasion. Proceedings of the National Academy of Sciences, 2017, vol. 114, n° 48, p. 12714-12719.Open the source
- WEINMANN, Markus, SCHNEIDER, Christoph, BROCKE, Jan vom. Digital Nudging. Business & Information Systems Engineering, 2016, vol. 58, p. 433-436.Open the source
- GUÉGUEN, Nicolas, JACOB, Céline. Fund-Raising on the Web: The Effect of an Electronic Foot-in-the-Door on Donation. CyberPsychology & Behavior, 2001, vol. 4, n° 6, p. 705-709.Open the source
- STONE, Jeff, FERNANDEZ, Nicholas C. To Practice What We Preach: The Use of Hypocrisy and Cognitive Dissonance to Motivate Behavior Change. Social and Personality Psychology Compass, 2008, vol. 2, n° 2, p. 1024-1051.Open the source
- BOGLIACINO, Francesco, PEJSACHOWICZ, Leonardo, LIVA, Giovanni et al. Testing for Manipulation: Experimental Evidence on Dark Patterns. SocArXiv, 2023.Open the source
- SIPONEN, Mikko, PUHAKAINEN, Petri, VANCE, Anthony. Can individuals' neutralization techniques be overcome? A field experiment on password policy. Computers & Security, 2020, vol. 88, p. 101617.Open the source
- RYOO, Yuhosua, KIM, WooJin. Approach versus avoidance: A self-regulatory perspective on hypocrisy induction in anti-cyberbullying CSR campaigns. Journal of Business Ethics, 2024, vol. 189, n° 2, p. 345-364.Open the source
Web pages
- SIMPLYPSYCHOLOGY. Social Identity Theory in Psychology. SimplyPsychology, 2023.Open the source
Conferences and proceedings
- FOGG, B. J. A Behavior Model for Persuasive Design. Proceedings of the 4th International Conference on Persuasive Technology, Association for Computing Machinery, New York, 2009.Open the source
[09]Downloads
Thesis
Theoretical framework, critical analysis of three protocols, protocol design and analysis plan. Written before data collection: it does not contain the results.
PDF • FR • 0.9 MB • 28 p.
[ PDF ]Defence slides
Protocol, descriptive results (n = 78), cross analysis and statistical discussion.
PDF • FR • 0.9 MB • 19 p.
[ PDF ]
Documents in French only.