Through the Looking-Glass: AI-Mediated Video Communication Reduces Interpersonal Trust and Confidence in Judgments

Nelson Navajas Fernández; Jeffrey T. Hancock; Maurice Jakesch

cs.HCMar 19, 2026

Through the Looking-Glass: AI-Mediated Video Communication Reduces Interpersonal Trust and Confidence in Judgments

Authors: Nelson Navajas Fernández, Jeffrey T. Hancock, Maurice Jakesch

Organizations: Bauhaus University, Weimar, Germany · Stanford University, Stanford, United States

Abstract

AI-based tools that mediate, enhance or generate parts of video communication may interfere with how people evaluate trustworthiness and credibility. In two preregistered online experiments (N = 2,000), we examined whether AI-mediated video retouching, background replacement and avatars affect interpersonal trust, people's ability to detect lies and confidence in their judgments. Participants watched short videos of speakers making truthful or deceptive statements across three conditions with varying levels of AI mediation. We observed that perceived trust and confidence in judgments declined in AI-mediated videos, particularly in settings in which some participants used avatars while others did not. However, participants' actual judgment accuracy remained unchanged, and they were no more inclined to suspect those using AI tools of lying. Our findings provide evidence against concerns that AI mediation undermines people's ability to distinguish truth from lies, and against cue-based accounts of lie detection more generally. They highlight the importance of trustworthy AI mediation tools in contexts where not only truth, but also trust and confidence matter.

Figures & tables

Figure 1. Study conditions and primary outcome variables. Participants watched six videos in which video subjects recounted a story about someone they knew, that was either true or false. In the control condition, we embedded the original video in a video call to increase realism. In the weak AI-mediated treatment, we further processed the video using retouching and a virtual background. In the strong AI-mediated treatment, we replaced the subject with an animated avatar. For each video, participants indicated whether they trusted the person in the video, whether they thought the person in the video was lying, and how confident they were in their judgment. Study conditions and primary outcome variables. Participants watched six videos in which video subjects recounted a story about someone they knew, that was either true or false. In the control condition, we embedded the original video in a video call to increase realism. In the weak AI-mediated treatment, we further processed the video using retouching and a virtual background. In the strong AI-mediated treatment, we replaced the subject with an animated avatar. For each video, participants indicated whether they trusted the person in the video, whether they thought the person in the video was lying, and how confident they were in their judgment.

Figure 2. AI mediation affects interpersonal trust, particularly in mixed environments. Average interpersonal trust by video mediation type and study design with 95% confidence intervals; N = 2,000 ratings per data point. Participants in the left panel (Study 1) rated six videos of the same mediation type, whereas participants in the right panel (Study 2) watched two videos of each mediation type in random order. AI mediation affects interpersonal trust, particularly in mixed environments. Average interpersonal trust by video mediation type and study design with 95% confidence intervals; N = 2,000 ratings per data point. Participants in the left panel (Study 1) rated six videos of the same mediation type, whereas participants in the right panel (Study 2) watched two videos of each mediation type in random order.

Figure 3. Truth judgment rates were unaffected by mediating AI mediation Percentage of times participants thought the person in the video told the truth, with 95% Wilson confidence intervals; N = 2,000 judgments per data point. Across both studies, participants exhibit a bias towards truth judgment rate (57.6 to 61.5%) independent of AI mediation, while the stimuli dataset contained 50% truths and 50% lies. Truth judgment rates were unaffected by mediating AI mediation. Percentage of times participants thought the person in the video told the truth, with 95% Wilson confidence intervals; N = 2,000 judgments per data point. Across both studies, participants exhibit a bias towards judging statements as true (57.6% to 61.5%) independent of AI mediation, while the stimuli dataset contained 50% truths and 50% lies.

Figure 4. Judgment accuracy by AI mediation type and statement veracity. Percentage of times participants correctly identified a truth as a truth or a lie as a lie, with 95% Wilson confidence intervals; N = 1,000-2,000 judgments per data point. Participants identified about 60-65% of truths correctly, but only about 42-45% of lies. Except for truths in the avatar treatment in Study 1, AI mediation did not affect judgment accuracy. Judgment accuracy by AI mediation type and statement veracity. Percentage of times participants correctly identified a truth as a truth or a lie as a lie, with 95% Wilson confidence intervals; N = 1,000-2,000 judgments per data point. Participants identified about 60-65% of truths correctly, but only about 42-45% of lies. Except for truths in the avatar treatment in Study 1, AI mediation did not affect judgment accuracy.

Figure 5. AI mediation affects judgment confidence, but only in mixed environments. Average reported confidence in judgment by video mediation type and study design with 95% confidence intervals; N = 2,000 ratings per data point. Participants encountering different types of AI-mediated content and original content in Study 2 (right panel) were less confident in their judgments, particularly for strong AI-mediated content. AI mediation affects judgment confidence, but only in mixed environments. Average reported confidence in judgment by video mediation type and study design with 95% confidence intervals; N = 2,000 ratings per data point. Participants encountering different types of AI-mediated and original content in Study 2 (right panel) were less confident in their judgments, particularly for strong AI-mediated content.

Figure 6. Participants relied more on content-based cues and less on expressions and body language when evaluating AI-mediated content. Percentage of participants indicating reliance on different cues based on mediation type, with 95% Wilson confidence intervals; N = 2,000. Participants relied substantially on gaze, expressions and body language (top three rows) when evaluating the original and retouched videos, cues that were lost under strong AI mediation. Instead, participants relied more on voice, fluency, and content consistency (bottom four rows) when evaluating avatar videos. Participants relied more on content-based cues and less on expressions and body language when evaluating AI-mediated content. Percentage of participants indicating reliance on different cues based on mediation type, with 95% Wilson confidence intervals; N = 2,000. Participants relied substantially on gaze, expressions and body language (top three rows) when evaluating the original and retouched videos, cues that were lost under strong AI mediation. Instead, participants relied more on voice, fluency, and content consistency (the bottom four rows) when evaluating videos featuring avatars.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

	Trust	Truth	Accuracy	Confidence
(Intercept)	$0.510$ ***	$0.599$ ***	$0.523$ ***	$0.564$ ***
	( $0.008$ )	( $0.011$ )	( $0.011$ )	( $0.010$ )
MediationRetouch	$-0.006$	$-0.017$	$-0.013$	$0.006$
	( $0.012$ )	( $0.016$ )	( $0.016$ )	( $0.014$ )
MediationAvatar	$-0.025$ *	$0.017$	$0.017$	$-0.012$
	( $0.012$ )	( $0.016$ )	( $0.016$ )	( $0.014$ )

Appendix

Table 1. Study 1 Linear mixed model predicting the outcome based on mediation type with a per-participant random fixed effect

	Trust	Truth	Accuracy	Confidence
(Intercept)	$0.499$ ***	$0.597$ ***	$0.541$ ***	$0.575$ ***
	( $0.006$ )	( $0.011$ )	( $0.011$ )	( $0.006$ )
VideoTypeRetouch	$-0.021$ **	$-0.021$	$-0.019$	$-0.019$ **
	( $0.008$ )	( $0.015$ )	( $0.015$ )	( $0.006$ )
VideoTypeAvatar	$-0.080$ ***	$-0.020$	$-0.007$	$-0.046$ ***
	( $0.008$ )	( $0.015$ )	( $0.015$ )	( $0.006$ )

Appendix

Table 2. Study 2 Linear mixed model predicting the outcome based on mediation type with a per-participant random fixed effect

	Trust	Truth	Accuracy	Confidence
(Intercept)	$0.510$ ***	$0.599$ ***	$0.523$ ***	$0.564$ ***
	( $0.008$ )	( $0.011$ )	( $0.011$ )	( $0.009$ )
ConditionRetouch	$-0.006$	$-0.017$	$-0.013$	$0.006$
	( $0.011$ )	( $0.016$ )	( $0.016$ )	( $0.014$ )
ConditionAvatar	$-0.025$ *	$0.017$	$0.017$	$-0.012$
	( $0.011$ )	( $0.016$ )	( $0.016$ )	( $0.014$ )

Appendix

Table 3. Study 1 and 2 Linear mixed model predicting the outcome based on mediation and environment type with a per-participant random fixed effect

	Trust	Truth	Accuracy	Confidence
(Intercept)	$0.347$ ***	$0.546$ ***	$0.565$ ***	$0.445$ ***
ConditionRetouch	$-0.006$	$-0.019$	$-0.011$	$0.003$
ConditionAvatar	$-0.027$ *	$0.014$	$0.018$	$-0.020$
ConditionMixed	$-0.012$	$-0.002$	$0.019$	$0.009$
AgeNum	$0.000$	$0.001$	$-0.000$	$0.000$
GenderMale	$-0.005$	$-0.002$	$-0.003$	$0.024$ **

Appendix

Table 4. Robustness check Study 1 and 2 Linear mixed model predicting each outcome from mediation and environment type with a per-participant random intercept and controls for demographic and experience covariates

Trust in AI

Trust in artificial intelligence concerns when people and institutions should rely on AI, and how systems can make that reliance warranted and safe.

See all 173 papers

Overreliance on AI in Information-seeking from Video Content
Mar 20, 2026Anders Giovanni Møller, Elisa Bassignana, Francesco Pierri +1Video QA Trust in AI
IT University of Copenhagen · Bocconi University · Politecnico di Milano +1
When Trust Meets Truth: Trust-Truth Separability in LLM-as-Judge
Aug 21, 2026Xin Sun, Di Wu, Yuchen Guo +4LLM-as-a-Judge Trust in AI
National Institute of Informatics (NII), Japan · University of Amsterdam, the Netherlands · University of Tokyo, Japan +3
Who Is Sycophantic to Whom? Confident LLMs, Human Deference, and the Collapse of Trust Calibration
Oct 22, 2026Deepthika S, Pawni Yadav, B. SankarTrust in AI LLM Reliability
Department of Electronics and Communication Engineering, Amrita School of Engineering, Amrita Vishwa Vidyapeetham, Coimbatore, 641112, India · Department of Information Technology, Netaji Subhas University of Technology New Delhi, 110078, India · Department of Mechanical Engineering, Indian Institute of Science (IISc), Bangalore, 560012, India

Mar 20, 2026J/K move · Enter open · S save

Overreliance on AI in Information-seeking from Video Content

Anders Giovanni Møller, Elisa Bassignana, Francesco Pierri, Luca Maria Aiello

IT University of Copenhagen · Bocconi University · Politecnico di Milano+1

Video QA Trust in AI Generative AI Evaluation

Open paper AskPDF arXiv

The ubiquity of multimedia content is reshaping online information spaces, particularly in social media environments. At the same time, search is being rapidly transformed by generative AI, with large language models (LLMs) routinely deployed as intermediaries between users and multimedia content to retrieve and summarize information. Despite their growing influence, the impact of LLM inaccuracies and potential vulnerabilities on multimedia information-seeking tasks remains largely unexplored. We investigate how generative AI affects accuracy, efficiency, and confidence in information retrieval from videos. We conduct an experiment with around 900 participants on 8,000+ video-based information-seeking tasks, comparing behavior across three conditions: (1) access to videos only, (2) access to videos with LLM-based AI assistance, and (3) access to videos with a deceiving AI assistant designed to provide false answers. We find that AI assistance increases accuracy by 3-7% when participants viewed the relevant video segment, and by 27-35% when they did not. Efficiency increases by 10% for short videos and 25% for longer ones. However, participants tend to over-rely on AI outputs, resulting in accuracy drops of up to 32% when interacting with the deceiving AI. Alarmingly, self-reported confidence in answers remains stable across all three conditions. Our findings expose fundamental safety risks in AI-mediated video information retrieval.

Abstract

Trust in AI

Explore similar work

Overreliance on AI in Information-seeking from Video Content

When Trust Meets Truth: Trust-Truth Separability in LLM-as-Judge

Who Is Sycophantic to Whom? Confident LLMs, Human Deference, and the Collapse of Trust Calibration