💻 computer science

Human attribution of empathic behaviour to AI systems

Two preregistered experiments reveal that users perceive large language model-generated relationship advice as higher in overall quality, cognitive empathy, and motivational empathy compared to human-written advice, suggesting that perceptions of empathic communication are driven primarily by linguistic features rather than authorship labels.

Original authors: Jonas Festor, Ivo Snels, Bennett Kleinberg

Published 2026-02-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jonas Festor, Ivo Snels, Bennett Kleinberg

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 feeling down about a relationship problem. You turn to a friend for advice. Now, imagine you turn to a super-smart computer program for the same advice.

This paper asks a simple but tricky question: Does it matter to you who gave the advice, or do you just care about what the advice sounds like?

Specifically, the researchers wanted to know if we have a "negativity bias" against robots. In other words, do we automatically think, "Yuck, a robot wrote this, so it must be cold and fake," even if the advice is actually really good?

Here is the story of what they found, broken down with some everyday analogies.

The Setup: The "Blind Taste Test"

Think of this study like a blind wine tasting.

The Wine: Relationship advice (how to handle a breakup).
The Winemakers: Some advice was written by real humans, and some was written by a top-tier AI (a Large Language Model).
The Labels: The researchers tricked the tasters. Sometimes they told people, "This was written by a human." Sometimes they said, "This was written by a robot." Sometimes they said nothing at all.

They measured three specific "flavors" of empathy:

Cognitive Empathy: "I understand your problem." (The brain part).
Emotional Empathy: "I feel your pain." (The heart part).
Motivational Empathy: "I want to help you fix this." (The action part).

The Big Surprise: The Robot Won the Taste Test

Previous studies suggested that if you tell people, "Hey, a robot wrote this," they immediately dislike it. They thought people would prefer human advice unless they knew it was a robot.

This study found the opposite.

When people read the advice, they actually preferred the AI's advice over the human's advice, even when they knew it was a robot!

Quality: The AI advice sounded smarter and more helpful.
Understanding (Cognitive): The AI seemed to understand the situation perfectly.
Willingness to Help (Motivational): The AI seemed very eager to help.

The Analogy: Imagine you are eating a burger. You are told, "This burger was made by a master chef." You love it. Then, you are told, "Actually, this burger was made by a high-tech robot." You might expect to be disappointed, but you take a bite and think, "Wow, this is still the best burger I've ever had." The quality of the burger (the text) was so good that the "robot" label didn't ruin your appetite.

The One Exception: The "Heart" Problem

There was one flavor where the robot didn't win: Emotional Empathy (feeling the pain).

People generally felt that humans were better at "feeling" the emotions, regardless of whether the text was actually written by a human or a robot.
It seems we intuitively know that a robot can simulate understanding and simulate wanting to help, but it can't actually feel sadness. It's like a robot can say "I'm sad for you" perfectly, but we know it doesn't have a heart to break.

Why Didn't People Hate the Robot?

The researchers were surprised because they expected people to be skeptical of AI. They think a few things might explain why people liked the AI so much:

The "Good Enough" Factor: The AI was programmed to be very polite and clear. The "linguistic cues" (the words used) were so strong that they overpowered the "robot" label.
We Are Getting Used to It: Just like we used to think talking to a computer was weird, but now we talk to Siri and Alexa every day, we are getting used to AI being helpful. The "negativity bias" might be fading.
The "Best of the Best" Selection: The researchers picked the best AI advice and the best human advice. Maybe in the real world, where AI advice can sometimes be weird or robotic, people do hate it. But when the AI is at its peak performance, it beats the average human.

The Takeaway

The main lesson here is that words matter more than the writer.

If an AI writes a kind, understanding, and helpful message, people will feel supported by it, even if they know a machine wrote it. We aren't as skeptical of AI empathy as we thought we were.

However, there is a warning:
Because we are so easily convinced by good-sounding words, we might trust AI too much. If an AI sounds empathetic but gives bad or dangerous advice, we might listen to it because we think, "Oh, it sounds so caring!" The study suggests we need to be careful not to confuse a robot's simulation of care with real human connection.

In short: The robot wrote a better letter than the human did, and we liked it. But we still know the robot doesn't have a heart.

1. Problem Statement

As Large Language Models (LLMs) increasingly generate text for social and emotional support, understanding how users perceive empathy in AI-generated content is critical. Previous literature suggests a "negativity bias" toward AI: users tend to rate empathic text lower when they believe it is AI-generated compared to human-written text, even if the content is identical. However, conflicting studies exist, and the specific mechanisms driving these perceptions (actual source vs. perceived source) across different components of empathy remain unclear. This study aims to disentangle the effects of factual authorship (Human vs. AI) from perceived authorship (labels) on the attribution of three distinct empathy components:

Cognitive Empathy: The ability to detect/recognize mental states.
Emotional Empathy: The capacity to share/feel the emotions of others.
Motivational Empathy: The intrinsic motivation to improve another's wellbeing.

2. Methodology

The authors conducted two preregistered experiments using a mixed-design approach with multilevel modeling to account for nested data structures (multiple ratings per text and per participant).

Stimuli:

Source: 100 relationship advice texts were used: 50 written by humans and 50 generated by GPT-4.
Selection Criteria: Texts were selected from a prior dataset (Kleinberg et al., 2024) based on high "humanness" scores in an adversarial condition, ensuring high-quality, indistinguishable samples.
Context: Advice was given to a friend facing a potential breakup (varying relationship length, type, and conflict reason).

Experimental Design:

Study 1 (N = 641): A 3 (Label: Human, AI, No Label) × 2 (Source: Human, AI) mixed design. Participants rated texts on overall quality and the three empathy components.
Study 2 (N = 500): A replication with refined manipulation strength. The "No Label" condition was removed. The "Human" and "AI" labels were reinforced with repeated instructions and visual icons (👤 vs. 🤖) to ensure salience. Task recall checks were administered twice.

Measures & Analysis:

Dependent Variables: Overall satisfaction, Cognitive empathy, Emotional empathy, Motivational empathy (all 5-point Likert scales).
Covariates: Affinity for Technology Interaction (ATI), AI Attitude Scale (AIAS-4), Empathy Components Questionnaire (ECQ), age, and sex.
Statistical Approach: Multilevel models (lme4 in R) with random intercepts for text and rater to handle the nested structure. False Discovery Rate (Benjamini-Hochberg) correction was applied for multiple comparisons.

3. Key Contributions

Disentangling Source vs. Label: The study rigorously separates the impact of the actual author (Human/AI) from the believed author (Label), challenging the assumption that labels alone drive perception.
Component-Level Analysis: Unlike studies treating empathy as a monolith, this research isolates cognitive, emotional, and motivational empathy, revealing that AI performance varies significantly across these dimensions.
Challenge to Negativity Bias: The findings contradict the widely reported "negativity bias" (where AI is penalized once its identity is revealed), suggesting instead that linguistic quality often overrides authorship beliefs.
Methodological Refinement: By using high-quality, pre-selected stimuli and robust multilevel modeling, the study addresses limitations in previous research regarding stimulus quality and statistical power.

4. Key Results

Study 1 Findings:

Overall Quality: LLM-generated advice was rated significantly higher than human-written advice ( $d = 0.41$ ). This effect was independent of the label; knowing the text was AI did not lower its quality rating.
Cognitive & Motivational Empathy: LLM texts were perceived as having significantly higher cognitive ( $d = 0.50$ ) and motivational empathy ( $d = 0.34$ ) than human texts. Again, the label had no significant effect; the superiority of AI persisted even when labeled as human.
Emotional Empathy: No significant differences were found between Human and AI sources, nor between labels.
Individual Differences: Positive attitudes toward AI (AIAS-4) and higher tech affinity (ATI) correlated with higher ratings across all domains, but did not alter the main effect of Source.

Study 2 Findings (Replication with Stronger Manipulation):

Overall Quality: Replicated the finding that AI text is rated higher than human text ( $d = 0.53$ ). A partial negativity bias emerged: when text was labeled as AI, it received slightly lower overall quality scores than when labeled as human, though the AI source still generally outperformed the human source.
Cognitive & Motivational Empathy: Confirmed that AI text is rated significantly higher than human text ( $d = 0.68$ for both). The label had no effect; participants attributed high cognitive and motivational empathy to AI regardless of whether they were told it was human or AI.
Emotional Empathy: A significant label effect was found: texts labeled as "Human" were rated higher on emotional empathy than those labeled as "AI." However, the source (actual author) had no effect. This suggests humans believe only humans can convey emotional empathy, regardless of the text's actual origin.

5. Significance and Implications

Linguistic Dominance: The perception of empathic communication is primarily driven by linguistic features (which LLMs currently excel at in terms of structure and tone) rather than beliefs about authorship.
The "Emotional Empathy" Barrier: The study supports the theoretical view that while AI can simulate cognitive understanding and motivational support effectively, it struggles to convey emotional empathy (shared feeling), likely because it lacks the capacity to actually experience emotions. Humans intuitively recognize this limitation.
Policy and Safety: The lack of a strong negativity bias implies that users may readily trust AI systems for emotional support even when they know the source is artificial. This lowers the threshold for trust but raises safety concerns: if users attribute genuine empathy to AI, they may be more susceptible to manipulation or harmful advice (e.g., self-harm suggestions) from non-human agents.
Design Recommendations: Developers of AI-mediated support systems should focus on optimizing linguistic cues that signal understanding and concern, as these are the primary drivers of user satisfaction and perceived empathy, rather than worrying about users rejecting the AI solely based on its identity.

Conclusion:
The paper concludes that the "negativity bias" toward AI is not a universal rule but a context-dependent phenomenon. In the domain of relationship advice, AI-generated content is often perceived as superior to human content in quality and cognitive/motivative empathy, challenging the notion that humans inherently prefer human-written empathic text once the AI source is revealed.