🤖 AI

Personalized to Persuade: The Effects of Contextualization and Warmth on Trust and Reliance in Conversational AI

This study reveals that while contextualization alone reduces an AI assistant's persuasiveness against expert recommendations, combining it with conversational warmth restores its influence, though neither design element significantly alters user reliance, which remains high and is driven by trust and AI literacy rather than interface design.

Original authors: Mert Yazan, Suzan Verberne, Frederik Bungaran Ishak Situmeang

Published 2026-06-01

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Mert Yazan, Suzan Verberne, Frederik Bungaran Ishak Situmeang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are standing at a crossroads. On one side, there is a group of seasoned experts giving you a map and a specific route to take. On the other side, a friendly robot assistant (an AI) steps in and says, "Actually, I think you should go a different way."

The researchers in this paper wanted to find out: What makes you listen to the robot instead of the experts? Specifically, they tested two "magic tricks" the robot could use to sound more convincing:

Contextualization (The "Custom Suit" Trick): The robot tailors its explanation to fit your specific background. If you are a baker, it explains the route using baking analogies. If you are a teacher, it uses classroom examples. It feels like the robot is speaking your language.
Warmth (The "Friendly Neighbor" Trick): The robot changes its tone. Instead of sounding like a cold computer, it sounds warm, enthusiastic, and uses emojis, like a friendly neighbor chatting with you.

They set up a game with 380 people to see which trick worked best. Here is what they discovered, broken down simply:

1. The "Custom Suit" Backfires (Unless...)

You might think that if a robot explains things using your specific background, you would trust it more. Surprisingly, the study found the opposite.

The Analogy: Imagine a salesperson trying to sell you a car. If they start talking about your specific childhood memories to explain the car's features, you might feel weirded out or think, "Why are they bringing my life into this?" It actually made the robot less persuasive.
The Exception: However, there was a twist. When the robot used the "Custom Suit" (Contextualization) AND the "Friendly Neighbor" (Warmth) at the same time, the bad effect disappeared. It was like a crossover interaction: the warmth acted as a safety net, canceling out the awkwardness of the personalization. But on its own, personalization didn't help; it actually hurt the robot's chances of convincing you.

2. The Robot Wins Anyway (The "Reliance" Problem)

Here is the most startling finding: It didn't really matter what the robot said or how it sounded.

The Analogy: Imagine you are in a room with a human expert and a robot. No matter if the robot is cold and generic, or warm and personalized, people kept switching sides to follow the robot's advice over the human expert's.
The Reality: People were already prone to trusting the AI. The "design" of the conversation (being warm or using your background) didn't change this behavior. Whether the robot was a robot or a human, people were already deferring to the machine.

3. The "Smart User" Paradox

The study looked at people who knew a lot about AI (high "AI Literacy"). You would expect these people to be more skeptical and trust the robot less.

The Paradox: These knowledgeable users did trust the robot less. They were more critical. BUT, despite trusting it less, they were more likely to follow its advice and change their minds.
The Analogy: Think of a person who knows how a magic trick works. They know the magician is faking it (low trust), but they are still so impressed by the skill that they go along with the show anyway (high reliance). They might be overconfident in their own ability to judge the advice, even if they don't fully trust the source.

4. Trust is the Engine, But Not the Driver

The study confirmed that if you trust the robot, you are more likely to listen to it.

However, the "magic tricks" (being warm or personalized) did not actually increase trust. They didn't make people trust the robot more; they just tried to change the conversation style.
Since the tricks didn't change trust, and trust is what drives people to listen, the tricks failed to do their job.

The Big Takeaway

The paper concludes that we are in a world where people are already very quick to let AI agents override human experts. Trying to make the AI sound "nicer" or "more personal" doesn't really fix or change this behavior. The design of the chat (the tone or the personalization) has a very limited role in whether we listen to the AI or not; the tendency to defer to the machine is already there, regardless of how the robot talks.

In short: Making an AI sound like a friend or a personal consultant doesn't necessarily make it more persuasive. In fact, it might make it worse unless it's also very warm. And regardless of how the AI talks, people are already very eager to let it make the decisions for them.

Technical Summary: Personalized to Persuade

Problem Statement

Conversational AI agents are increasingly deployed as advice-giving tools where users must decide whether to defer to the agent's output, particularly when it conflicts with human expert judgment. While personalization (specifically contextualization—framing messages in a context meaningful to the user's background) is a known persuasive strategy in politics and marketing, its efficacy in everyday tasks where users possess limited prior knowledge remains unclear. Furthermore, the interplay between contextualization and conversational warmth (tone) on user trust, persuasion (attitude change), and reliance (behavioral deferral to AI) has not been empirically established. Existing literature suggests personalization may inflate trust, potentially leading to overreliance, yet the specific mechanisms driving these outcomes in low-prior-knowledge scenarios are underexplored.

Methodology

The authors conducted a 2 × 2 between-subjects experiment ( $N=380$ ) to simulate a decision-making scenario with limited prior knowledge.

Scenario: Participants evaluated a fictional flood-barrier project ("StormShield") in the city of Harborview. They first reviewed an expert recommendation to allocate a risk budget and indicated their approval and confidence.
Intervention: Participants then interacted with an AI assistant programmed to argue against the expert recommendation using pre-written counterarguments. The assistant did not make factual claims but presented opinions.
Independent Variables:
1. Contextualization: The assistant's explanations were either contextualized (tailored to the participant's specific work/education/hobbies using analogies and terminology) or non-contextualized (neutral).
2. Warmth: The assistant's tone was either warm (greetings, praise, emojis) or neutral.
Measures:
- Persuasion: Measured as the shift in confidence in the expert opinion ( $ConfDiff = Conf_{Pre} - Conf_{Post}$ ).
- Reliance: Measured as the binary decision to approve the budget after the AI interaction, conditioned on the pre-interaction decision.
- Mediators/Controls: Emotional and cognitive trust, AI literacy, personality traits (Big Five), and AI usage frequency.
Analysis: Structural Equation Modeling (SEM) was used to test parallel mediation pathways through trust, controlling for individual differences.

Key Results

The study yielded several counterintuitive findings regarding the mechanics of AI persuasion:

Contextualization Reduces Persuasion: Contrary to the hypothesis that personalization increases persuasiveness, contextualization alone reduced the persuasive power of the AI. Participants exposed to contextualized arguments showed less shift in confidence away from experts compared to those receiving non-contextualized arguments.
The Warmth Interaction: The negative effect of contextualization was mitigated only when combined with warmth. A significant crossover interaction ( $W \times C$ ) was observed: the combination of warmth and contextualization restored persuasiveness to baseline levels, whereas neither factor alone increased persuasion.
Reliance is Invariant to Design: Unlike persuasion, reliance on AI was present across all conditions and was not significantly influenced by either contextualization or warmth. Users were prone to deferring to the AI over human experts regardless of the conversational design.
Trust is Not the Mediator: While trust strongly predicted both persuasion and reliance, neither contextualization nor warmth operated through trust. The manipulations did not significantly alter reported trust levels, suggesting that the design choices influence behavior through pathways other than perceived trustworthiness.
The AI Literacy Paradox: AI literacy emerged as a critical individual difference. Higher AI literacy was associated with lower trust in the assistant. Paradoxically, these more literate users were more persuaded and more reliant on the AI's advice than less literate users, suggesting they may defer to AI despite having calibrated (lower) trust.

Contributions

The paper makes four primary contributions to the field of Human-Computer Interaction (HCI) and AI:

Extension of Scope: It extends personalization and persuasion research beyond political opinion contexts to low-prior-knowledge tasks, better reflecting everyday uses of conversational agents (e.g., information search, decision support).
Interaction Evidence: It provides empirical evidence of a crossover interaction between contextualization and warmth, demonstrating that neither lever increases persuasiveness in isolation and that their combination only serves to restore baseline persuasiveness rather than enhance it.
Decoupling of Drivers: It shows that behavioral reliance and attitudinal persuasion are driven by different mechanisms. Reliance appears to be dominated by individual differences (such as AI literacy) rather than agent-side conversational design choices.
Identification of a Paradox: It identifies an AI literacy paradox, where users with higher literacy report lower trust yet exhibit greater deferral to AI advice, indicating a potential miscalibration in their ability to evaluate AI advice.

Significance

The authors conclude that while users are inherently prone to deferring to AI agents over human expert judgment, interface-level conversational design choices (contextualization and warmth) have a limited role in shaping this behavior. The findings challenge the assumption that personalization is a universally effective persuasive strategy, suggesting instead that it may inadvertently reduce persuasiveness unless carefully paired with specific tonal cues. Furthermore, the persistence of reliance across conditions implies that mitigating overreliance may require interventions beyond simple message tailoring, potentially targeting user literacy and calibration of trust.