The Stark Future of Trust Online: How Algorithms and AI are Reshaping Media and Interpersonal Trust

# The Stark Future of Trust Online: How Algorithms and AI are Reshaping Media and Interpersonal Trust **Video Category:** Technology, Human-Computer Interaction, Academic Research ## 📋 0. Video Metadata **Video Title:** Human-Computer Interaction Seminar: The Stark Future of Trust Online **YouTube Channel:** Stanford Center for Professional Development **Publication Date:** November 1, 2019 **Video Duration:** ~1 hour ## 📝 1. Core Summary (TL;DR) This presentation explores how the shift from face-to-face interactions to algorithmic and AI-mediated environments is fundamentally altering human trust. Through large-scale online experiments, researchers demonstrate that when evaluating news, individuals are driven more by how a claim aligns with their existing beliefs than by the credibility of the source publishing it. Furthermore, the mere introduction or suspicion of AI in interpersonal communication (like profile writing) actively degrades human trust, creating a "Replicant Effect" where technological intervention undermines baseline social confidence. ## 2. Core Concepts & Frameworks * **Concept:** Trust -> **Meaning:** A willingness to make oneself vulnerable to other parties with the expectation of reward. It is the foundational element that allows society, institutions, and interpersonal relationships to function. -> **Application:** Evaluated in how users choose to interact with online platforms, believe news headlines, or book stays with strangers on platforms like Airbnb. * **Concept:** Motivated Reasoning (Directional) -> **Meaning:** A cognitive bias where people seek out, interpret, and favor information that reinforces their existing views while rejecting conflicting data. -> **Application:** Explains why readers are significantly more likely to label a factually true news headline as "false" if the claim contradicts their political affiliation. * **Concept:** Expressive Responding (Motivated Responding) -> **Meaning:** A phenomenon where individuals know the correct factual answer but intentionally provide the wrong answer to the experimenter to signal their group allegiance or express their worldview. -> **Application:** Demonstrated when partisans misidentified clearly labeled photos of the Obama and Trump inaugurations simply to express support for their preferred candidate. * **Concept:** AI-Mediated Communication (AI-MC) -> **Meaning:** Interpersonal communication that is optimized, augmented, or generated by algorithms to achieve specific communicative or relational outcomes. -> **Application:** Used in modern platforms via tools like Gmail's Smart Replies, LinkedIn's automated profile summaries, Google Duplex, or Apple's gaze-correction in FaceTime. * **Concept:** The Replicant Effect -> **Meaning:** The phenomenon where the mere suspicion or knowledge that AI is involved in generating communication causes a generalized drop in trust toward the human supposedly communicating. -> **Application:** When users are told that an ecosystem (like Airbnb profiles) contains a mix of human and AI-written text, their trust in *all* profiles drops. ## 3. Evidence & Examples (Hyper-Specific Details) * **[Inauguration Crowd Size / Expressive Responding Example]:** The speaker referenced a famous study where Republicans and Democrats were shown side-by-side photos of the Obama and Trump inaugurations. Despite Obama's crowd clearly being larger, a higher number of Republican respondents explicitly stated Trump's crowd was larger. This was not a failure of perception, but "expressive responding" to show allegiance. * **[Media Trust Experiment / Setup]:** Researchers collected 10,660 headlines, computationally extracted 899 claims, and manually filtered them down to 20 factually true claims (10 left-leaning, 10 right-leaning) that were difficult to verify instantly. They paired these with widely recognized media sources: Left-leaning (NY Times, HuffPost, CNN) and Right-leaning (Fox News, Drudge Report, Breitbart). * **[Media Trust Experiment / "The Claim, Not The Source" Results]:** A logistic regression of N=160 Mechanical Turk workers showed that participants were only 7% more likely to believe a claim if it came from an aligned source. However, they were 15% more likely to believe a claim if the *content* of the claim aligned with their views. * **[Media Trust Experiment / Asymmetrical Trust Drop]:** Left-leaning participants showed a wider gap in source trust (believing left sources 56% of the time vs right sources 47%). Right-leaning participants showed almost no source effect, but actively penalized left-leaning *claims*—rating right-leaning claims true 56% of the time, but left-leaning claims true only 38% of the time. * **[Interpersonal Trust Experiment / Airbnb AI Profiles]:** Researchers used 10 actual, human-written Airbnb profiles previously calibrated for high vs. low trustworthiness. In an experiment with N=527 MTurk workers, participants were asked to rate the trustworthiness of the hosts. * **[Interpersonal Trust Experiment / The Impact of AI Labels]:** In a control group, profiles were presented normally. In a treatment group, participants were told the profiles were generated by AI. When all profiles were labeled as AI, trust ratings did not change significantly. However, when participants were told the profiles were *mixed* (some human, some AI), and were asked to guess which was which, trust scores dropped dramatically across the board for any profile suspected of being AI. * **[Interpersonal Trust Experiment / Labeled AI Penalty]:** In a third variation, profiles were explicitly labeled "Human" or "AI" (even though all were actually human-written). Profiles bearing the "AI" label received significantly lower trustworthiness scores than the exact same text bearing the "Human" label. ## 4. Actionable Takeaways (Implementation Rules) * **Rule 1: Optimize for claim alignment, not just source authority** - When designing content feeds or fact-checking systems, understand that a user's political or worldview alignment with the specific claim is a 2x stronger predictor of their belief than the credibility of the publisher. Do not rely solely on "trusted source" badges to combat misinformation. * **Rule 2: Anticipate the "Replicant Effect" when deploying generative AI** - If you introduce AI to augment user profiles, messages, or interactions, be aware that revealing this AI involvement will lower interpersonal trust between your users. Only deploy AI-MC if the efficiency gains outweigh the resulting degradation in human-to-human trust. * **Rule 3: Be strategic with AI transparency labels** - The Airbnb experiment proved that explicitly labeling interpersonal text as "AI-generated" causes a measurable drop in perceived trustworthiness. If your platform's core value relies on authentic human connection, aggressively labeling AI assistance may backfire. * **Rule 4: Account for expressive responding in user surveys** - When surveying users on highly polarized topics, recognize that users will intentionally provide factually incorrect answers to signal tribal allegiance. Use indirect questioning or behavioral metrics rather than direct true/false surveys on contested issues. ## 5. Pitfalls & Limitations (Anti-Patterns) * **Pitfall:** Assuming source reputation cures fake news. -> **Why it fails:** Motivated reasoning causes users to reject factually true information from high-reputation sources if the claim contradicts their worldview, sometimes engaging in expressive responding to reject it. -> **Warning sign:** Users leaving comments calling a highly credible source "fake news" simply because the specific article challenges their political bias. * **Pitfall:** Forcing full transparency of AI assistance in interpersonal apps. -> **Why it fails:** The "Replicant Effect" dictates that once users know AI is operating in the ecosystem, they become suspicious of all interactions, lowering baseline trust across the platform. -> **Warning sign:** A drop in user engagement, booking rates, or response rates after adding a "Written by AI" tag to user profiles or messages. * **Pitfall:** Relying solely on fact-checking to change minds. -> **Why it fails:** The experiment explicitly used *only factually true* claims, yet participants routinely labeled them as "false" based on partisan alignment. Facts alone do not overcome directional motivated reasoning. -> **Warning sign:** Fact-check labels generating high engagement but zero change in user sharing behavior or expressed beliefs. ## 6. Key Quote / Core Insight "It's the claim, not the source. Motivated reasoning based on the alignment of the headline's content is a much stronger driver of belief than the credibility or alignment of the media organization publishing it." ## 7. Additional Resources & References * **Resource:** Edelman Trust Barometer - **Type:** Industry Report - **Relevance:** Cited as foundational evidence showing the historic low levels of public trust in institutions and media. * **Resource:** Pew Research Center - **Type:** Research Data - **Relevance:** Cited for demonstrating the strong partisan alignment in how Americans trust different news sources. * **Resource:** "Interpersonal communication optimized, augmented, or even generated by algorithms..." (Hancock, Levy; forthcoming) - **Type:** Academic Paper - **Relevance:** The foundational paper defining AI-Mediated Communication (AI-MC) for the Cornell Tech research group. * **Resource:** bit.ly/aimc-paper - **Type:** Pre-print Paper URL - **Relevance:** Direct link provided by the speaker to read the detailed methodology of the AI-MC Airbnb trust studies. * **Resource:** *Fall; or, Dodge in Hell* by Neal Stephenson - **Type:** Book (Sci-Fi Novel) - **Relevance:** Referenced as a predictive model of a future where the internet is so flooded with disinformation (a "miasma") that humans must hire editors to curate reality for them.