Empathy Fog: Attenuation in Empathic Processes

• Empathy fog is a heterogeneous concept defined as the degradation or distortion of empathic signals due to distance, mediated channels, or AI calibration errors.
• It spans multiple disciplines, from social neuroscience and HCI to VR and computational modeling, highlighting challenges in cue transmission and measurement.
• Empirical studies using EEG, VR experiments, and formal models demonstrate that small shifts in physical or communicative parameters can significantly impact empathic responses.

Empathy fog is a heterogeneous research term for the attenuation, distortion, or latent reconfiguration of empathic processes when interpersonal understanding is obscured by physical distance, mediated channels, communicative mismatch, AI calibration errors, or imprecise operationalization. Across social neuroscience, HCI, VR, human–AI interaction, NLP, and mathematical modeling, the cited literature does not treat empathy fog as a single standardized construct. Taken together, it suggests a family of related phenomena in which empathic registration, transmission, or measurement becomes degraded, ambiguous, or difficult to detect (Lomoriello et al., 2018, Lee, 2019, Meng et al., 23 Feb 2026, Roshanaei et al., 2024, Siyan et al., 25 Jun 2026, Lahnala et al., 24 Jan 2025, Calderoli et al., 30 Jan 2026).

1. Conceptual scope and definitional variants

The literature uses empathy fog in several distinct but overlapping ways. In some work, it names a reduction in neural empathic response caused by perceived physical distance; in others, it denotes the communicative haze introduced by mediated channels, the co-constructed breakdown of mutual understanding, the distorted empathy profile of AI systems, the invisible behavioral effect of subtle empathic cues, or the epistemic blur created by coarse task definitions (Lomoriello et al., 2018, Lee, 2019, Meng et al., 23 Feb 2026, Roshanaei et al., 2024, Siyan et al., 25 Jun 2026, Lahnala et al., 24 Jan 2025, Calderoli et al., 30 Jan 2026).

Domain	Paper	Formulation of empathy fog
Social neuroscience	"Out of sight out of mind: Perceived physical distance between the observer and someone in pain shapes observer's neural empathic reactions" (Lomoriello et al., 2018)	Perceived physical distance attenuates the P3/LPP-type neural signature of empathy for pain
Computer-mediated communication	"Computer-mediated Empathy" (Lee, 2019)	Mediated channels create an invisible barrier, with loss of nonverbal signals and flattened interaction
Embodied VR	"52-Hz Whale Song: An Embodied VR Experience for Exploring Misunderstanding and Empathy" (Meng et al., 23 Feb 2026)	Communicative mismatch creates a haze between people that can be addressed through role-shifting and mediation
Human–AI empathy evaluation	"Talk, Listen, Connect: Navigating Empathy in Human-AI Interactions" (Roshanaei et al., 2024)	AI produces a foggy or distorted empathy profile through over-expression and blurred nuance
Behavior-change chatbots	"Invisible Impact of Empathy on Behavioral Change: Isolating the Effect of Empathy in Long-term Physical Activity Coaching Chatbot Interactions" (Siyan et al., 25 Jun 2026)	Users fail to perceive empathy differences explicitly, yet behavior and motivation still shift
NLP operationalization	"The Muddy Waters of Modeling Empathy in Language: The Practical Impacts of Theoretical Constructs" (Lahnala et al., 24 Jan 2025)	Coarse or indirect task definitions obscure transferable empathic behaviors
Double empathy modeling	"A Distinct Communication Strategies Model of the Double Empathy Problem" (Calderoli et al., 30 Jan 2026)	Feedback-driven empathy collapse emerges from asymmetrical channel weighting and defensivity

A recurring misconception is that empathy fog refers only to reduced feeling. The cited work is broader. It includes failures of cue transmission, failures of mutual calibration, failures of conscious access to empathy effects, and failures of construct specification. Another misconception is that the term necessarily implies an intrinsic empathy deficit in one party. The double-empathy modeling literature explicitly rejects that interpretation, instead locating degradation in dyadic feedback and communication preferences (Calderoli et al., 30 Jan 2026).

2. Physical distance and neural attenuation

Schiano Lomoriello et al. tested whether manipulating retinal image size, and therefore perceived physical distance, modulates neural empathic reactions to others’ pain (Lomoriello et al., 2018). Experiment 1 used a between-subjects pain-decision task with forty healthy undergraduates, with 17 usable EEG recordings per group after artifact rejection. Participants saw Caucasian face stimuli for 400 ms, either upright or inverted, and either painfully stimulated by a syringe or neutrally stimulated by a cotton-tip applicator. Stimuli were drawn from the Eberhardt Lab Face Database and scaled to two sizes: a close condition of $2.5^\circ \times 3.3^\circ$, implying approximately $2$ m social distance, and a far condition of $1.6^\circ \times 2.5^\circ$, implying approximately $3$ m. Each participant completed 576 trials in four 144-trial runs. Accuracy and reaction time did not differ by distance, orientation, or stimulation, with all $F$s $< 1$ and $p$s $> .30$.

Continuous EEG was recorded from 64 scalp sites in a 10–20 montage, referenced online to the left earlobe and re-referenced offline to linked earlobes. Signals were band-pass filtered $0.01$–$80$ Hz, digitized at $2$0 Hz, and epoched from $2$1 to $2$2 ms around face onset. Trials with blinks, eye movements, or amplitude excursions $2$3 were discarded, approximately $2$4 of trials. Following prior work by Fan and Han and by Sessa and Meconi, the analysis targeted the P3/LPP component as mean amplitude $2$5–$2$6 ms post-stimulus at fronto-central and centro-parietal electrode pools. The critical Stimulation $2$7 Distance interaction was significant at both pools: fronto-central $2$8, and centro-parietal $2$9. Planned tests showed that only the close group exhibited larger P3 amplitudes for painful than neutral faces. At fronto-central sites in the close group, painful faces yielded $1.6^\circ \times 2.5^\circ$0 $1.6^\circ \times 2.5^\circ$1 versus neutral $1.6^\circ \times 2.5^\circ$2 $1.6^\circ \times 2.5^\circ$3, $1.6^\circ \times 2.5^\circ$4; at centro-parietal sites, painful faces yielded $1.6^\circ \times 2.5^\circ$5 $1.6^\circ \times 2.5^\circ$6 versus neutral $1.6^\circ \times 2.5^\circ$7 $1.6^\circ \times 2.5^\circ$8, $1.6^\circ \times 2.5^\circ$9.

Experiment 2 addressed a straightforward confound: smaller faces might simply be harder to identify. Twenty-two new students performed a match-to-sample task with the same two upright face sizes across 528 trials. Accuracy and reaction time again did not differ for close versus far sizes, with accuracy $3$0 and reaction time $3$1. The reported interpretation therefore attributes the ERP effect to perceived distance rather than discriminability.

The theoretical framing combines Hall’s proxemics, Construal Level Theory, and Embodied Simulation. In the CLT reconstruction provided in the source material, construal level increases with distance, $3$2, with $3$3. In the embodied-simulation interpretation, resonance strength is modeled as decreasing with distance, $3$4, $3$5. This suggests that even a modest shift from approximately $3$6 m to $3$7 m can be sufficient to damp the neural empathic marker without changing task performance.

3. Mediated channels and the loss of empathic cue richness

In "Computer-mediated Empathy," empathy is framed as a dyadic interaction between the empathizee, who must engage in self-expression and self-reflection, and the empathizer, whose role is specified through Wiseman’s four attributes: perspective taking, non-judgmental stance, recognizing emotions, and communicating understanding (Lee, 2019). Within this framework, empathy fog is characterized as the invisible barrier introduced by mediated channels. Its reported consequences are loss of nonverbal signals, reduced emotional resonance, and flattened interactions.

The paper is a position paper rather than an original empirical study, and it explicitly does not present original user studies. Its significance lies in decomposition. Self-expression is defined as making internal states available through rich multimodal channels such as audio, video, text, and creative artifacts. Self-reflection is treated as guided introspection that increases self-awareness and clarity about what to share. Perspective taking is linked to mixed-reality systems that place the empathizer into the empathizee’s situation; non-judgmental stance is linked to asynchronous text channels and algorithmic filters; recognizing emotions is linked to affective-computing pipelines; and communicating understanding is linked to explicit feedback mechanisms such as awareness widgets and acknowledgment prompts.

The technical contribution is therefore architectural and design-oriented. The paper outlines possible support layers, including VR system pipelines, moderation pipelines, speech-to-prosody and text-to-transformer emotion-recognition pipelines, empathy-aware chat interfaces, reflective dashboards, and “empathy token” style UI mechanisms. It also identifies risks: over-reliance on algorithmic emotion recognition, privacy and consent issues in capturing intimate self-expression, and empathy fatigue from continual emotional labor. A plausible implication is that empathy fog, in this strand of work, is less a failure of empathic capacity than a channel-capacity problem in which low-bandwidth or poorly scaffolded interfaces suppress the cues on which empathy ordinarily depends.

4. Communicative mismatch, embodiment, and formal collapse dynamics

Meng et al. operationalize empathy fog as the haze that settles between individuals when communicative mismatches prevent genuine understanding (Meng et al., 23 Feb 2026). Their embodied VR system, "52-Hz Whale Song," uses the real-world “52-Hz whale” as a metaphor for interactional mismatch rather than as a representation of any specific social group. The experience has a three-act arc: Act I, failed communication, in which players embody a whale whose calls generate golden waves and haptic pulses that rebound from an invisible “sonic net”; Act II, agency, in which a survival crisis emphasizes that self-help alone cannot restore contact; and Act III, mediation, in which players re-embody as a pink dolphin that translates whale fragments into intelligible signals and physically positions itself between pods.

The evaluation used a preliminary mixed-methods design with $3$8, randomly assigned to experimental $3$9 and control $F$0 conditions. At baseline, all participants watched a 5-minute documentary on immigration, completed the Social Distance Scale and Interpersonal Reactivity Index, and participated in a brief interview. The experimental group then completed a 15-minute VR experience; the control group rested. Post-study, all participants repeated the scales and interviews. Quantitatively, the experimental group showed larger reductions on the Social Distance Scale, with $F$1 $F$2 versus control $F$3 $F$4, $F$5, including $F$6, $F$7, and $F$8 at $F$9–$< 1$0. Perspective-Taking gains were $< 1$1 $< 1$2 versus $< 1$3 $< 1$4, $< 1$5. Fantasy gains were $< 1$6 $< 1$7 versus $< 1$8 $< 1$9, $p$0. Qualitatively, the reported themes were bodily frustration, emotional resilience, role-shift epiphany, and contextual transfer. The design sequence “Subject $p$1 Resilient Agent $p$2 Mediator” is presented as a transferable interaction pattern.

Calderoli et al. provide a complementary formalization in the context of the double empathy problem (Calderoli et al., 30 Jan 2026). Their model treats empathy degradation as a bidirectional feedback process driven by communication preferences rather than by a one-sided deficit. Registered empathy is defined as

$p$3

with empathy gap

$p$4

Defensivity is the dynamical state:

$p$5

and verbal and nonverbal empathy outputs decline linearly with defensivity:

$p$6

The mechanism is recursive: higher defensivity lowers output, lower output enlarges the partner’s empathy gap, and the loop can spiral toward “empathy collapse.”

The stability analysis introduces a loop gain,

$p$7

with Jury conditions for fixed-point stability. In 120 simulation runs varying $p$8 and $p$9, collapse occurred in 20 runs $> .30$0, all with $> .30$1, and in none with $> .30$2. The phase plane contains a low-defensivity fixed point near $> .30$3, asymptotically stable with eigenvalues $> .30$4, and a high-defensivity fixed point near $> .30$5, also stable once reached. This makes empathy fog a quantitatively testable interactional attractor rather than a metaphor alone.

5. Human–AI empathy fog: distortion, over-expression, and invisible effects

In "Talk, Listen, Connect," empathy is analyzed as affective empathy, cognitive empathy, and a relational, bidirectional process, with measurement anchored in the Interpersonal Reactivity Index and, for HCI, the Perceived Empathy of Technology Scale (Roshanaei et al., 2024). The empirical study compares human–human and human–AI empathy judgments using 756 narratives from 126 Stanford undergraduates and 2,586 MTurk annotators, and then evaluates base GPT-4o, persona-prompted GPT-4o, and two fine-tuned variants. The central finding is systematic over-expression by base GPT-4o. Overall empathy ratings were Human $> .30$6 $> .30$7 versus GPT-4o $> .30$8 $> .30$9, with $0.01$0, $0.01$1, Wasserstein $0.01$2, and $0.01$3. Empathy-Affective showed $0.01$4, $0.01$5, $0.01$6, $0.01$7, and Empathy-Cognitive showed $0.01$8, $0.01$9, $80$0, $80$1. Persona prompting produced little change, although a perspective-taking persona slightly improved cognitive alignment. Instruction fine-tuning was more consequential: GPT-4o FT Story-only improved overall empathy to $80$2, $80$3, and GPT-4o FT All reduced the overall difference to $80$4, $80$5, $80$6, while improving Empathy-Cognitive to $80$7, $80$8. The paper therefore describes empathy fog here as a distorted empathy profile: blurred nuance, failure in pleasant contexts, and mis-weighting of similarity signals.

Siyan et al. isolate a different manifestation, which they explicitly term “Invisible Empathy” or empathy fog, in long-term physical-activity coaching chatbots (Siyan et al., 25 Jun 2026). Their six-week within-subject study $80$9 compares three WhatsApp chatbots differing only in empathy level: Non-Empathetic, Standard, and Empathetic. The Empathetic version uses a Clinical Empathy Module that classifies Empathy Opportunities with a tuned GPT-4o-mini classifier at top-3 accuracy $2$00, samples response strategies from real healthcare distributions, and injects corresponding system instructions into GPT-4. Participants often could not consciously distinguish the empathy conditions, and the non-empathetic version was often rated as more engaging and useful. Yet linear mixed-effects modeling showed reliable latent effects: Days predicted higher step counts, $2$01; the interaction $2$02 predicted faster improvement in intention, $2$03, and self-efficacy, $2$04. LLM-based EPITOME annotation also confirmed higher empathy scores for the more empathic variants, with Emotion $2$05, $2$06, $2$07, Exploration $2$08, $2$09, $2$10, and Interpretation $2$11, $2$12, $2$13. This directly challenges the assumption that empathy must be consciously recognized in order to alter motivation or behavior.

6. Operationalization, measurement, and unresolved problems

"The Muddy Waters of Modeling Empathy in Language" relocates empathy fog to the level of construct definition itself (Lahnala et al., 24 Jan 2025). The paper distinguishes direct, abstract, and adjacent empathy tasks, and represents each task $2$14 by a three-dimensional construct vector

$2$15

where $2$16 is definition granularity, $2$17 is link correspondence between labels and construct, and $2$18 is conduciveness of the communication context. Across 18 tasks, transfer learning used RoBERTa single-bottleneck adapters and compared intermediate-task transfer to target-only baselines. Significant improvements were rare overall, occurring in fewer than $2$19 of task pairs. Direct empathy sources produced the most improvements and the fewest harms; abstract sources never significantly helped abstract targets; adjacent sources sometimes helped direct targets but often harmed or failed elsewhere. Inter-annotator agreement on task characterization was reported as Krippendorff’s $2$20 for Definition, $2$21 for Link, and $2$22 for Conduciveness, with Spearman’s $2$23, all $2$24. In permutation-importance analysis, $2$25 was the top predictor of transfer success, with a $2$26 decrease in $2$27 when permuted, and $2$28 was second at $2$29. The authors’ core claim is that precise, multidimensional operationalization is not ancillary; it is the main determinant of whether models learn transferable empathic behavior.

Taken together, the literature suggests that empathy fog is simultaneously phenomenological, infrastructural, and epistemic. It can arise because a face appears $2$30 m rather than $2$31 m away, because a medium suppresses nonverbal and contextual cues, because communicative preferences are misaligned, because AI systems overshoot human norms while missing nuance, because empathic effects operate through peripheral cues beneath conscious report, or because datasets collapse distinct empathic dimensions into abstract scores. Future directions explicitly proposed in the cited work include standardized empathy-aware UI components and longitudinal field studies for mediated settings (Lee, 2019), indirect behavioral evaluation rather than self-report alone in chatbot studies (Siyan et al., 25 Jun 2026), parameter-estimation experiments for communication weights and output asymmetries in double-empathy modeling (Calderoli et al., 30 Jan 2026), and the refinement of context-sensitive empathy metrics such as PETS and multilevel models in human–AI evaluation (Roshanaei et al., 2024). A plausible implication is that resolving empathy fog requires not one intervention but coordinated work on distance cues, interaction design, role structure, model calibration, and construct validity.