TRUST-Pathos: Trust via Emotional Communication

• TRUST-Pathos is a multidimensional framework that defines trust as both a latent attitude and a behaviorally expressed emotional response through speech and rhetoric.
• It integrates Aristotelian concepts with modern computational methods to assess trust and pathos in contexts like human-robot interaction, political speech, and formal argumentation.
• The framework offers actionable insights for calibrating trust in dynamic interactions while addressing ethical and methodological challenges.

Taken together, the relevant literature suggests that TRUST-Pathos is a pathos-centered treatment of trust: trust is approached not only as a latent attitude, but also as something expressed through speech, rhetorical framing, affective behavior, commitment, and audience interpretation. In human-robot interaction, speech has been proposed as an objective, passive, real-time proxy for trust, organized around what users say, when they say it, and how they say it, with the operational goal of calibrating trust toward a level appropriate to the robot’s actual abilities (Velner et al., 2021). In computational rhetoric and political-speech analysis, pathos is treated not as generic sentiment, but as hearer-oriented emotional persuasion or as the divisive or unifying rhetorical impact of emotional language (Gajewska et al., 2024, Dietrich, 21 May 2026). The result is a cross-domain research agenda in which trust and pathos are jointly modeled as dynamic, socially situated, and measurable.

1. Conceptual foundations

The conceptual background of TRUST-Pathos is Aristotelian. Ethos concerns credibility and moral character; pathos concerns appeals to affective states; logos concerns argument and reason. Recent computational work preserves this distinction and explicitly warns against collapsing pathos into sentiment analysis, generic emotion labeling, or low-level acoustic affect (Gajewska et al., 1 Jul 2026). In social-media annotation, pathos is defined as rhetoric intended “to induce specific affective states in the hearer rather than simply express emotions,” and in data-visualization theory it is treated as emotional appeal that overlaps with, but is not identical to, rhetoric and aesthetics (Prantl et al., 2023).

A second foundational distinction is between pathos as emotional persuasion and pathos as commitment. In the Trichotomic Argument Interchange Format, pathos is represented via weighted edges from actors to illocutions, capturing their level of commitment to propositions, whereas ethos is represented via weighted edges between actors that encode trust (Göttlinger et al., 2018). This formalization suggests that TRUST-Pathos spans both the affective impact of communication and the speaker’s stance toward what is being said.

Domain	TRUST-Pathos operationalization	Source
Human-robot interaction	Speech as an objective, passive, real-time proxy for trust	(Velner et al., 2021)
Online rhetoric	Hearer-oriented appeals to affective states	(Gajewska et al., 2024)
Reader interpretation	Ethos and pathos preserved or altered in silent-audience interpretations	(Gajewska et al., 1 Jul 2026)
Political speech	Ordinal score from $-2$ to $+2$ for divisive or unifying rhetorical impact	(Dietrich, 21 May 2026)
Formal argumentation	Weighted commitment edges for pathos; weighted trust edges for ethos	(Göttlinger et al., 2018)
Data visualization	Emotional appeal distinct from aesthetics and logos-centered design	(Prantl et al., 2023)

2. Speech as trust signal in human-robot interaction

In human-robot interaction, the most explicit TRUST-Pathos proposal is “Speaking of Trust — Speech as a Measure of Trust,” which argues that speech can serve as an objective, passive, real-time proxy for trust (Velner et al., 2021). The motivating problem is practical: trust affects relationship formation, self-disclosure, reliance on the robot, and whether the resulting trust is appropriate to the robot’s actual capabilities. Existing measures are described as inadequate because questionnaires are subjective and vulnerable to bias, while active objective measures such as trust games are difficult to use in natural interaction and are generally not real-time.

The proposed speech cues are organized into three dimensions. The first is what the user says, including words used by the user, overlap or matching with robot language, and possibly speech acts. The second is when the user speaks, including response duration, pauses, turn-taking behavior, and conversational timing dynamics. The third is how the user speaks, including pitch, intensity, speech rate, emphasis, and possibly voice quality. The proposal therefore treats trust as behaviorally and affectively expressed in language use, timing, and prosody rather than only in explicit self-report.

The intended use is calibration, not mere detection. The robot is envisioned as operating in a sense–think–act cycle: it listens to what, when, and how the user speaks; infers trust level from these cues; and adjusts behavior to move trust toward an appropriate level. If overtrust is sensed, the robot should adapt to reduce it; if undertrust is sensed, it should adapt to increase it. The emphasis on appropriate trust is central, because the target is neither maximal reliance nor systematic skepticism.

A common misconception is to treat this proposal as an evaluated speech-trust system. It is not. The paper is a conceptual workshop short paper and explicitly does not present a participant study, labeled trust corpus, feature extraction pipeline, machine-learning model, evaluation metric, accuracy result, or formal calibration equation (Velner et al., 2021). Its contribution is a research agenda: speech may support continuous trust inference, but the empirical, statistical, and control-theoretic instantiation remains open.

3. Rhetorical annotation, audience interpretation, and formal representation

In computational rhetoric, TRUST-Pathos is grounded by corpora and annotation schemes that distinguish pathos from adjacent constructs. The PolarIs corpora were created for studying polarization in social media through ethos and pathos across COVID-19 vaccines, climate change, and US 2016 elections, on Reddit and Twitter (Gajewska et al., 2024). Across five datasets, PolarIs contains 162,278 words and 15,588 sentences. Ethos is present in 20.1% of sentences overall, with 4.6% supports and 15.5% attacks; pathos is present in 31.0% of sentences, with 5.3% positive and 25.7% negative. The corpus design is thread-aware rather than comment-isolated, reflecting the fact that targets of trust attack or support are often interpretable only in context.

The annotation logic is explicitly hearer-oriented. A sentence is labeled as pathos if it is intended to evoke emotion in the hearer, and the polarity decision concerns whether that induced emotional effect is positive or negative. This excludes simple speaker-emotion reading. The authors also report that pathos annotation is more difficult than ethos detection, with lower agreement, and identify irony, sarcasm, short utterances, and ambiguity between expressed and evoked emotion as recurrent error-prone cases. Within this framework, TRUST-Pathos is tightly coupled to polarization because ethos models credibility attack and support, while pathos models emotional mobilization.

A further step is provided by the study of “silent audiences.” Using OrigamIM, a dataset of 2,018 English sentences paired with about five human-written interpretations each, for a total of 9,851 interpretations, ethos and pathos are labeled on both source sentences and interpretations (Gajewska et al., 1 Jul 2026). The principal result is that interpretations diverge from the original sentence in 30% of cases. More specifically, 74.2% of interpretations match the source’s ethos label and 70.4% match the source’s pathos label, while full alignment of both dimensions across all interpretations of a sentence occurs in only 15.9% of cases. Neutral sentences are significantly more stable than rhetorically charged ones, and positive ethos and positive pathos show especially high interpretive variability. Ethos and pathos in the original sentence also predict attitudes toward the author, indicating that rhetorical appeals shape perception even when readers do not comment.

Formal argumentation work provides an additional layer of precision. T-AIF extends structured argumentation by representing logos in the usual argument graph, ethos as weighted trust edges between entities, and pathos as weighted actor-to-illocution edges representing commitment (Göttlinger et al., 2018). On this view, pathos is not merely emotional tone; it is the degree to which a speaker stands behind a proposition. This makes it possible to define properties such as similarity, agreement, rationality, justified trust, and trust compliance on a graph that combines argument structure, social trust, and speaker commitment.

4. Multimodal measurement and real-time analytics

A distinct operationalization appears in political-speech analysis. “Beyond Acoustic Emotion Recognition” defines TRUST-Pathos not as a generic emotion score, but as the societal impact of emotional language in a speech, ranging from $+2$ for language that unifies across party lines, through $0$ for neutral or in-group-oriented language, down to $-2$ for language that is actively divisive (Dietrich, 21 May 2026). This score is produced by the TRUST Multimodal Pipeline v1.0, in which a three-advocate LLM supervisor ensemble evaluates segments and aggregates judgments by median consensus. In the reported case study, 51 utterances from a Bundestag plenary speech were segmented, and 41 remained after filtering.

The central empirical question is whether acoustic speech emotion recognition can stand in for this political-rhetorical construct. The answer is largely negative. Gemini Valence correlates strongly with TRUST-Pathos at $\rho = +0.664$, $p < 0.001$, whereas emotion2vec Valence does not, at $\rho = +0.097$, $p = 0.499$ (Dietrich, 21 May 2026). Gemini Arousal also correlates with TRUST-Pathos, but negatively, while emotion2vec Arousal does not. The authors interpret this as a construct mismatch: acoustic models capture voice-level affect, whereas TRUST-Pathos captures political-semantic valence, including sarcasm, accusation, irony, and appeal. Acoustic features remain informative for low-level arousal estimation, but they do not reliably recover the rhetorical social meaning of emotionally charged political speech.

Real-time trust analytics in dialogue move the same general agenda into human-AI communication. VizTrust is a Streamlit-based web application with a front-end chatbot using Llama-3.1-8B and a back-end analysis system that combines a six-agent AutoGen trust-evaluation pipeline with additional NLP and ML routines (Wang et al., 10 Mar 2025). Trust is modeled along four dimensions—competence, integrity, benevolence, and predictability—and the system produces turn-by-turn 7-point Likert scores with supporting evidence. It also tracks engagement, politeness strategies, and emotional tones, including anger, fear, joy, love, sadness, and surprise.

VizTrust’s contribution is architectural rather than benchmark-driven. After each user utterance, the transcript is analyzed, outputs are stored, and a dashboard visualizes trust and behavioral signals over time. The trust turning point is therefore a visual and interpretive notion rather than the output of a dedicated change-point detector. The paper presents a storyboard case study rather than a formal controlled evaluation, but it concretely demonstrates the TRUST-Pathos premise that trust and affect co-evolve turn by turn and can be made observable in real time (Wang et al., 10 Mar 2025).

5. Trust calibration, affective trust, and sociotechnical embedding

Adjacent trust research clarifies where pathos-like mechanisms sit within broader trust theory. In a financial decision-support setting, anthropomorphic design is hypothesized to influence risk perception through two routes: cognitive trust and affective trust (Reani et al., 14 Feb 2026). The experiment uses an 8-level anthropomorphic design varying visual cue, identity cue, and communicative cue, and the manipulation significantly affects perceived anthropomorphism. Human-like communication is the strongest lever, with a reported contrast of $\Delta = 0.290$, $+2$0, $+2$1, while the human name manipulation is essentially inert.

The trust results are dual rather than unitary. Perceived anthropomorphism increases cognitive trust and affective trust. Cognitive trust decreases risk perception, whereas affective trust does not reliably lower risk and in OLS has a small positive association with risk; in SEM, the direct affective-trust path to risk perception is not significant (Reani et al., 14 Feb 2026). Domain knowledge further moderates these pathways. This is important for TRUST-Pathos because it shows that affective trust is not simply a benign amplifier of adoption: pathos-like design cues can alter trust, but the downstream effect depends on expertise, task, and the distinction between competence-based and warmth-based trust.

A more explicitly sociotechnical account is offered by the bowtie model of trust in LLMs. This model ties trustor contextual factors—such as demographics, ideologies, background, perceptions, moral values, past experiences, and familiarity—to trustee systemic elements such as AI/LLMs, scientists, products, and stakeholders (Paraschou et al., 11 Jun 2025). In the mixed-methods study built around a ChatGPT-powered political discourse tool, past experiences and familiarity significantly shape trust-related actions. Human-in-the-loop features enhance trust, while lack of transparency decreases it. The relevance to TRUST-Pathos is direct: the affective, attitudinal, and value-laden side of trust is treated not as noise, but as a structured part of the trustor side of the system.

A complementary shift appears in the Leap of Faith framework for ML trust. Rather than relying on trust self-reports, the framework operationalizes declared, demonstrated, and deserved trust through intervention choice, follow-through, and outcomes, comparing an ML agent with an expert-validated rules-based agent (Frame et al., 2024). This suggests a methodological extension of TRUST-Pathos: pathos-rich interaction may matter, but trust claims become analytically stronger when linked to observable reliance and outcome quality rather than to attitudinal declaration alone.

6. Ethical issues, misconceptions, and open problems

Ethics is not peripheral to TRUST-Pathos; in several papers it is part of the core contribution. In speech-based trust sensing, speech is explicitly described as privacy-sensitive because it can reveal identity, personality, emotions, and other personal information (Velner et al., 2021). The proposed safeguards are concrete: explicit consent before recording or analyzing speech, a clear statement that the data are used only for the interaction’s benefit, storage only when necessary, and protection through encryption and restricted access. These requirements follow from the fact that TRUST-Pathos systems often operate on data streams that are intimate, continuous, and difficult for users to audit.

Manipulation is the most persistent risk. If systems can calibrate trust, they can also be used to create overtrust for persuasion or commercial gain. The speech-based HRI paper gives the example of a robot used to sell dubious products by increasing the user’s trust; the social-media interpretation paper notes that its findings could be exploited for targeted disinformation or manipulative persuasion (Velner et al., 2021, Gajewska et al., 1 Jul 2026). Related work also warns that design choices may reinforce stereotypes, as in the example of using a female voice because it is perceived as more trustworthy. These concerns are magnified for vulnerable populations such as children, who are explicitly discussed as a motivating but high-risk group in child-robot interaction.

Methodological limitations are equally recurrent. The speech proposal has no experiment or model; VizTrust is validated only through system design and a storyboard case study; the silent-audience study relies on silver-standard rhetoric labels and notes that these should be interpreted as model-predicted rhetorical signals rather than absolute truths (Velner et al., 2021, Wang et al., 10 Mar 2025, Gajewska et al., 1 Jul 2026). In multimodal political-speech analysis, the critique of EMO-DB—acted speech, cultural bias, and category incompatibility—underscores a broader problem: benchmark corpora for basic emotion are often poor proxies for context-sensitive rhetorical constructs (Dietrich, 21 May 2026). Generalizability therefore remains open at multiple levels: across populations, languages, modalities, domains, and definitions of trust itself.

The overall trajectory of the literature is nevertheless clear. TRUST-Pathos consistently reorients trust research away from single post-hoc scores and toward dynamic, context-sensitive, and behaviorally grounded signals. Speech timing, lexical choice, prosody, commitment, audience reinterpretation, emotional framing, and multimodal rhetorical context all become candidate observables. The shared implication is not that trust is reducible to pathos, but that pathos is one of the principal ways trust becomes measurable, calibratable, and contestable in contemporary human-machine and human-human communication.