Discussion

1

PSYCHOLOGICAL REACTANCE THEORY 17

Psychological Reactance Theory

Psychological Reactance Theory

Method

Participants

In 2024, a total of 155 individuals from diverse backgrounds at Florida International University voluntarily participated in this study. The gender distribution included 43.2% male ( n = 67), 53.5% female ( n = 83), and 3.2% unspecified ( n = 5). Participants' ages ranged from 17 to 67 years, with an average age of M = 24.56 ( SD = 8.605). Ethnicity frequencies within the sample were as follows: 45.2% Latino/a ( n = 70), 15.5% black ( n = 24), 26.5% White ( n = 41), 5.2% Asian ( n = 8), 1.3% Indigenous ( n = 2), and 8.6% classified as Middle East/North Africa (MENA) ( n = 10). See tables below:

Statistics
Part D: Gender (1 = M, 2 = F, 3 = NB, 4 = O)
N	Valid	150
	Missing	5
Mean	1.55
Std. Deviation	.499
Minimum	1
Maximum	2

Part D: Gender (1 = M, 2 = F, 3 = NB, 4 = O)
	N	%
Male	67	43.2%
Female	83	53.5%
Missing	System	5	3.2%

Descriptive Statistics
	N	Minimum	Maximum	Mean	Std. Deviation
Part D: Age	155	17	67	24.56	8.605
Valid N (listwise)	155

Race
	N	%
White	41	26.5%
Latino/a	70	45.2%
Indigenous	2	1.3%
Black	24	15.5%
Asian	8	5.2%
MENA	10	6.5%

Materials and Procedure

The study commenced with an explicit and structured procedure to ensure consistency across participants. First and foremost, researchers approached participants individually to obtain oral informed consent to participate in the study. The researcher asked each potential participant for their willingness to partake in a study for the research methods class, explaining that the study would take approximately five to ten minutes with no inherent risks. A verbal "Yes" or "No" response sufficed, and in case of refusal, the researcher sought another participant.

If a participant agreed to take part in the study, he/she wa presented with three distinct versions of the questionnaire. This manipulation aimed to investigate the impact of language tone on reactance. In the "High Controlling Language" condition, participants encountered explicit and forceful language, using phrases like "prohibited" and "must never use." In contrast, the "Low Controlling Language" condition utilized more implicit and polite expressions such as "encouraged" and "should." The "Neutral Language" condition employed neutral and informative language, devoid of controlling elements.

Random assignment of participants to the three conditions added robustness to the study's internal validity. This allocation was crucial, ensuring that any observed effects could be confidently attributed to the language manipulation rather than pre-existing participant differences.

The dependent variables (DVs) consisted of survey questions designed to gauge participants' reactions, feelings, and intentions regarding the proposed AI policy. Participants provided demographic information including gender, age, and race, each recorded as categorical variables. The survey responses were collected on a Likert scale ranging from 1 (Strongly Disagree) to 7 (Strongly Agree) for various statements. For instance, statements related to reactance included items such as "The AI policy gives me too little freedom to decide how I can use AI."

The key DVs analyzed in the study focused on participants' reactance, feelings, and intentions towards the AI policy. Specifically, statements assessing reactance included participants' perceived threat to freedom, feelings of anger, and the extent to which they found the policy controlling. Intentions were probed through statements gauging participants' likelihood to follow the AI policy.

A critical element in ensuring participants' attention to the study was the manipulation check question embedded in the demographic section. Participants were asked to recall the tone of the AI policy they read, effectively gauging whether the language manipulation was successful.

Upon completion of the survey, participants were debriefed about the study's focus on Psychological Reactance Theory, emphasizing their valuable contributions and the study's broader implications. The debriefing aimed to maintain ethical standards and transparency in participant involvement.

Results

The first analysis was conducted on the manipulation check. Researchers ran a chi square test of examining the relationship between the experimental conditions (High Controlling, Low Controlling, and Neutral) and participants' responses to the attention check. The chi-square test yielded a significant effect, χ²(4) = 68.49 , p < .001, indicating that there were notable differences in the distribution of responses among the three conditions. The cross tabulation of experimental conditions and participants' responses to the attention check revealed distinct patterns. Specifically, participants in the High Controlling condition predominantly associated the AI usage with being "prohibited" (68.9%), while those in the Low Controlling condition leaned towards the response "discouraged" (69.6%). Participants in the Neutral condition exhibited a more varied distribution, with responses spread across "prohibited," "discouraged," and "inappropriate."

	Cases
	Valid	Missing	Total
	N	Percent	N	Percent	N	Percent
Condition (1 = High C, 2 = Low C, 3 = Neutral) * Part D: Attention Check	155	100.0%	0	0.0%	155	100.0%

Condition (1 = High C, 2 = Low C, 3 = Neutral) * Part D: Attention Check Crosstabulation
	Part D: Attention Check
	Using AI is "prohibited"	Using AI is "discouraged"
Condition (1 = High C, 2 = Low C, 3 = Neutral)	High Controlling	Count	42	7
		Expected Count	20.5	15.4
	Low Controlling	Count	16	32
		Expected Count	19.7	14.8
	Neutral	Count	3	7
		Expected Count	20.9	15.7
Total	Count	61	46
	Expected Count	61.0	46.0

Condition (1 = High C, 2 = Low C, 3 = Neutral) * Part D: Attention Check Crosstabulation
	Part D: Attention Check	Total
	Using AI is "inappropriate"
Condition (1 = High C, 2 = Low C, 3 = Neutral)	High Controlling	Count	3	52
		Expected Count	16.1	52.0
	Low Controlling	Count	2	50
		Expected Count	15.5	50.0
	Neutral	Count	43	53
		Expected Count	16.4	53.0
Total	Count	48	155
	Expected Count	48.0	155.0

Chi-Square Tests
	Value	df	Asymptotic Significance (2-sided)
Pearson Chi-Square	133.411a	4	<.001
Likelihood Ratio	133.481	4	<.001
Linear-by-Linear Association	85.011	1	<.001
N of Valid Cases	155
a. 0 cells (0.0%) have expected count less than 5. The minimum expected count is 14.84.

Symmetric Measures
	Value	Approximate Significance
Nominal by Nominal	Phi	.928	<.001
	Cramer's V	.656	<.001
N of Valid Cases	155

The second manipulation utilized the one-way ANOVA; researchers investigated the impact of experimental conditions (High Controlling, Low Controlling, and Neutral) on participants' responses to the attention check in Part D. The analysis revealed a significant condition effect, F(2, 152) = 94.75, p < .001. Subsequent Tukey post hoc tests were conducted to compare mean differences between conditions. The post hoc analysis indicated significant differences in participants' responses among the three conditions. Participants in the High Controlling condition ( M = 1.36, SD = 0.58) had significantly lower mean ratings compared to those in the Low Controlling ( M = 2.00, SD = 0.56) and Neutral ( M = 2.83, SD = 0.52) conditions (p < .001). Furthermore, the Low Controlling and Neutral conditions did not differ significantly from each other. The effect size analysis revealed a substantial effect of the experimental conditions on participants' responses, with an estimated Eta-squared of 0.555, indicating that 55.5% of the variability in the attention check responses can be attributed to the manipulation of language tone.

Descriptives
Condition (1 = High C, 2 = Low C, 3 = Neutral)
	N	Mean	Std. Deviation	Std. Error	95% Confidence Interval for Mean
					Lower Bound
Using AI is "prohibited"	61	1.36	.578	.074	1.21
Using AI is "discouraged"	46	2.00	.558	.082	1.83
Using AI is "inappropriate"	48	2.83	.519	.075	2.68
Total	155	2.01	.826	.066	1.88

Descriptives
Condition (1 = High C, 2 = Low C, 3 = Neutral)
	95% Confidence Interval for Mean	Minimum	Maximum
	Upper Bound
Using AI is "prohibited"	1.51	1	3
Using AI is "discouraged"	2.17	1	3
Using AI is "inappropriate"	2.98	1	3
Total	2.14	1	3

Tests of Homogeneity of Variances
	Levene Statistic	df1	df2
Condition (1 = High C, 2 = Low C, 3 = Neutral)	Based on Mean	4.632	2	152
	Based on Median	1.858	2	152
	Based on Median and with adjusted df	1.858	2	147.468
	Based on trimmed mean	4.795	2	152

Tests of Homogeneity of Variances
	Sig.
Condition (1 = High C, 2 = Low C, 3 = Neutral)	Based on Mean	.011
	Based on Median	.160
	Based on Median and with adjusted df	.160
	Based on trimmed mean	.010

ANOVA
Condition (1 = High C, 2 = Low C, 3 = Neutral)
	Sum of Squares	df	Mean Square	F	Sig.
Between Groups	58.261	2	29.131	94.750	<.001
Within Groups	46.732	152	.307
Total	104.994	154

ANOVA Effect Sizesa
	Point Estimate	95% Confidence Interval
		Lower	Upper
Condition (1 = High C, 2 = Low C, 3 = Neutral)	Eta-squared	.555	.448	.629
	Epsilon-squared	.549	.441	.624
	Omega-squared Fixed-effect	.547	.439	.623
	Omega-squared Random-effect	.377	.281	.452
a. Eta-squared and Epsilon-squared are estimated based on the fixed-effect model.

Post Hoc Tests

Multiple Comparisons
Dependent Variable: Condition (1 = High C, 2 = Low C, 3 = Neutral)
Tukey HSD
(I) Part D: Attention Check	(J) Part D: Attention Check	Mean Difference (I-J)	Std. Error	Sig.
Using AI is "prohibited"	Using AI is "discouraged"	-.639*	.108	<.001
	Using AI is "inappropriate"	-1.473*	.107	<.001
Using AI is "discouraged"	Using AI is "prohibited"	.639*	.108	<.001
	Using AI is "inappropriate"	-.833*	.114	<.001
Using AI is "inappropriate"	Using AI is "prohibited"	1.473*	.107	<.001
	Using AI is "discouraged"	.833*	.114	<.001

Multiple Comparisons
Dependent Variable: Condition (1 = High C, 2 = Low C, 3 = Neutral)
Tukey HSD
(I) Part D: Attention Check	(J) Part D: Attention Check	95% Confidence Interval
		Lower Bound	Upper Bound
Using AI is "prohibited"	Using AI is "discouraged"	-.90	-.38
	Using AI is "inappropriate"	-1.73	-1.22
Using AI is "discouraged"	Using AI is "prohibited"	.38	.90
	Using AI is "inappropriate"	-1.10	-.56
Using AI is "inappropriate"	Using AI is "prohibited"	1.22	1.73
	Using AI is "discouraged"	.56	1.10

*. The mean difference is significant at the 0.05 level.

Homogeneous Subsets

Condition (1 = High C, 2 = Low C, 3 = Neutral)
Tukey HSDa,b
Part D: Attention Check	N	Subset for alpha = 0.05
		1	2	3
Using AI is "prohibited"	61	1.36
Using AI is "discouraged"	46		2.00
Using AI is "inappropriate"	48			2.83
Sig.		1.000	1.000	1.000
Means for groups in homogeneous subsets are displayed.
a. Uses Harmonic Mean Sample Size = 50.877.
b. The group sizes are unequal. The harmonic mean of the group sizes is used. Type I error levels are not guaranteed.

Means Plots

Discussion

In this study, the researchers aimed to investigate the influence of language tone in AI policies on individuals' psychological reactance, feelings, and intentions. They hypothesized that participants exposed to the AI policy in high-controlling language would exhibit higher levels of reactance compared to those in low-controlling and neutral language conditions. The specific reactance-related statements included beliefs that the policy restricts freedom, threatens autonomy, and lacks respect for decision-making autonomy. The researchers' hypothesis was confirmed, revealing that participants exposed to the AI policy in high-controlling language exhibited heightened reactance compared to low-controlling and neutral conditions. This aligns with established psychological reactance theories, emphasizing the influence of language on user attitudes. Practical implications include the need for policymakers and designers to adopt language that respects autonomy, potentially reducing resistance. 

The prediction that participants exposed to the AI policy in high-controlling language would report more negative feelings about the policy compared to those in low-controlling and neutral language conditions was confirmed. Specific feelings-related statements included beliefs that the policy induces anger, is too extreme, and uses demanding language. The researchers validated their prediction, as participants exposed to the AI policy in high-controlling language reported more negative feelings than those in low-controlling and neutral conditions. The identified feelings, such as anger, perception of extremity, and the use of demanding language, underscore the emotional impact of language choice in policy communication.

The third hypothesis posited that participants exposed to the AI policy in high-controlling language would express lower intentions to follow the policy compared to those in low-controlling and neutral language conditions. Specific intention-related statements included beliefs that participants intend to use AI even in prohibited situations, intend to ignore the AI policy, and intend to use AI according to their judgment. The results aligned with the third hypothesis, revealing that participants exposed to the AI policy in high-controlling language expressed lower intentions to adhere to the policy compared to their counterparts in low-controlling and neutral language conditions. Noteworthy intention-related statements included participants expressing the intent to use AI in prohibited situations, disregarding the AI policy, and exercising personal judgment in AI usage.

Generally, the language tone in AI policies significantly influences individuals' psychological reactance, feelings, and intentions. By adopting less controlling language, policymakers and designers can potentially mitigate reactance and foster positive attitudes and intentions among users. This study provides practical insights for crafting effective AI policies that respect user autonomy and reduce resistance. Future research should explore the generalizability of these findings in diverse populations and settings, further advancing our understanding of the nuanced relationship between language, reactance, and user compliance.