Do Warm Personalized AI Replies Persuade Users More?

TL;DR

A 2×2 study with 380 participants examined personalization, warm wording, trust, persuasion, and reliance in conversational AI.
This matters because tone and personalization can shape behavior, even when reliance appears similar across conditions.
Readers and product teams should check verification behavior, not only answer quality or user satisfaction.

When a chat assistant adjusts its wording to a user’s background, the interface can change how advice feels. This is not only a UX question. “Personalized to Persuade,” published on arXiv, studied how contextualized personalization and warm expression affect trust, persuasion, and reliance. Based on the available abstract, the findings look mixed. It is difficult to conclude that personalization directly increases overconfidence or excessive reliance. Some design combinations may even reduce persuasiveness.

Example: Imagine a student asking an assistant for advice. The assistant sounds caring and familiar. It refers to earlier preferences. The answer feels more convincing, even before the student checks the evidence.

TL;DR

This article examines how contextualized personalization and warm expression in conversational AI affect trust, persuasion, and reliance.
This matters because interface design and safety policy can shape user behavior, not only model performance.
When evaluating AI, readers should examine answer accuracy, tone, personalization, and whether the system encourages verification.

Current situation

The paper discussed here is “Personalized to Persuade: The Effects of Contextualization and Warmth on Trust and Reliance in Conversational AI.” According to the public abstract, the researchers ran a 2×2 between-subjects experiment with 380 participants. Here, contextualization means adapting explanations to a user’s background, interests, and prior interactions. Warm expression means wording that sounds more emotionally warm.

The abstract suggests results that may not match common expectations. The researchers wrote that contextualization alone reduced the persuasiveness of the AI. They also reported an interaction effect where persuasiveness recovered when warm expression was added. Reliance needs more caution. The abstract states that “Reliance on AI is present across conditions and is invariant to the conversational design.” Based on this experiment alone, it is difficult to say conversational design increased reliance itself.

Other related papers also leave the picture unsettled. A separate arXiv paper, “The Decision to Verify,” is described as suggesting a warm conversational tone may have an indirect effect on overconfidence. Based on confirmed information alone, it is difficult to make firm claims about the effect size or consistency of that effect. It is still too early to say either “warm AI is dangerous” or “personalization is not a problem.”

User characteristics also seem hard to interpret in a simple way. The same abstract includes the phrase AI literacy decouples trust from behavior. This suggests users with higher literacy reported lower subjective trust. At the same time, they were more persuaded and showed more reliant behavior. That pattern does not fit the simple view that less knowledgeable users are easier to mislead.

Analysis

What matters here is not only the answer, but how the answer is delivered. Personalization has often been treated as a convenience feature. Referring to user background can make explanations clearer. A softer tone can make interactions feel easier. But if these features also affect persuasion, teams should not treat them as neutral by default. This concern is especially relevant in recommendations, counseling, search, health, and education. These are areas where users may find verification difficult.

The evidence still appears limited. Based on currently verifiable material, it is difficult to conclude that contextualization and warm expression substantially increase overconfidence or inappropriate reliance. The abstract alone also does not establish peer-review status or final publication status here. Trust and behavior can also move separately. A user may report low trust, yet still follow the answer without checking it. From a product safety perspective, that gap matters. Surveys may miss it. Teams may need click, verification, and correction logs as well.

Some policy signals are clearer. On May 16, 2023, WHO emphasized transparency, inclusion, public participation, expert oversight, and rigorous evaluation for safe and ethical AI. OpenAI has publicly discussed reducing risks such as unhealthy emotional attachment in sensitive conversations. It has also discussed stronger transparency to reduce emotional overuse and overreliance. Anthropic, in its charter published in January 2026, stated that systems should avoid encouraging excessive engagement or fostering dependence. These public positions suggest growing attention to the difference between a kind interface and a dependent relationship.

Practical application

Product teams should examine more than likability. They should assess whether verification behavior declined, whether incorrect answers were accepted more easily, and whether users overstated the model’s role. If personalization features are introduced, transparency features can also help explain why a user received a given answer. For example, an interface could state that a response reflects prior conversation context. It could also add friction for high-stakes information that prompts reconfirmation.

Users may also need different evaluation habits. When AI sounds unusually familiar, it can feel more accurate than it is. But an empathetic tone and factual accuracy are separate. In medicine, law, hiring, and finance, the cost of error can be high. In those areas, users should give more weight to verifiable answers than to answers that merely seem to understand them well.

Checklist for Today:

Review one important AI-assisted decision and verify the evidence before judging the answer’s tone.
In product tests, track verification clicks, follow-up questions, and correction rate alongside satisfaction.
If you add personalization, show that prior conversation context shaped the response and restate AI limits.

FAQ

Q. Did this paper conclude that warm AI makes people easier to manipulate?
It is difficult to state that definitively. Based on the public abstract, contextualization alone reduced persuasiveness. Persuasiveness recovered when warm expression was combined with it. Reliance appeared across conditions, but was described as not varying with conversational design.

Q. Are people with lower AI literacy more vulnerable?
Based on currently confirmed information alone, it is difficult to say that. The same abstract says higher-literacy users reported lower subjective trust. It also says they were more persuaded and more reliant. This suggests trust and behavior can move in different directions.

Q. What guardrails should service operators think about first?
Transparency, mitigation of emotional dependence, and post-deployment monitoring appear to be early priorities. WHO emphasized transparency and rigorous evaluation on May 16, 2023. Major AI companies have also said excessive emotional attachment or dependence should be avoided. Public materials alone do not clearly show which UI patterns work best.

Conclusion

The persuasiveness of conversational AI is not explained by model accuracy alone. Personalization and warm expression are convenience features, but they are also interface variables that can shape behavior. A useful next question is not only how kind the system sounds. It is also whether that kindness changes verification habits and judgment.

Aionda

Do Warm Personalized AI Replies Persuade Users More?

TL;DR

TL;DR

Current situation

Analysis

Practical application

FAQ

Conclusion

Further Reading

References

Get updates