The Ten Minutes That Shocked the World—Teaching Generative AI to Analyze the Trump–Zelensky Multimodal Debate

Abstract

Today, foundation models simulate humans’ skills in translation, literature review, fact checking, fake-news detection, novel and poetry production. However, generative AI can also be applied to discourse analysis. This study instructed the Gemini 2.5 model to analyze multimodal political discourse. We selected some fragments from the Trump–Zelensky debate held at the White House on 28 February 2025 and annotated each sentence, gesture, intonation, gaze, and facial expression in terms of LEP (Logos, Ethos, Pathos) analysis to assess when speakers, in words or body communication, rely on rational argumentation, stress their own merits or the opponents’ demerits, or express and try to induce emotions in the audience. Through detailed prompts, we asked the Gemini 2.5 model to run the LEP analysis on the same fragments. Then, considering the human’s and model’s annotations in parallel, we proposed a metric to compare their respective analyses and measure discrepancies, finally tuning an optimized prompt for the model’s best performance, which in some cases outperformed the human’s analysis: an interesting application, since the LEP analysis highlights deep aspects of multimodal discourse but is highly time-consuming, while its automatic version allows us to interpret large chunks of speech in a fast but reliable way.

1. Introduction

The course of human events is highly determined by words and body behaviors, and the capacity to catch their meanings may provide us with insights into history.

On 28 February 2025, the Presidents of USA and Ukraine, Donald J. Trump and Volodymir Zelensky, met in the Oval Office of the White House “https://www.youtube.com/watch?v=UhFqdAZ9Xnw (accessed on 26 January 2026)”, but the conclusions of the meeting were very different from what was hoped by Zelensky—obtaining help to end the war and the Russian aggression that started on 24 February 2022. In contrast, in this “Oval Office bullying” (so termed by the political analyst Ruth Deyermond), Zelensky was humiliated by Trump and his collaborators, who aimed to force him into accepting a peace plan favorable to Moscow. The meeting had heavy and long-lasting geopolitical consequences: it revealed the human, communicative, and political style of Trump, and in some way, determined the future pace of world politics.

The words uttered and the gestures, postures, and facial expressions displayed in that meeting impacted on subsequent events and the interactants’ relationships, and generally changed the ways, rules, and styles of communication in political affairs, as humans all over the world soon understood. However, could have a foundation model have caught such a relevant change?

In the last few years, generative AI has rapidly evolved toward AI agents—modular LLM-based systems integrated with external tools via advanced prompt engineering—then transitioning toward agentic AI, characterized by multi-agent collaboration, dynamic task decomposition, persistent memory, and orchestrated autonomy [1].

The remarkable capabilities of current AI models rely on a fundamental shift in architecture and computational scale. The Transformer architecture [2], with its self-attention mechanism, unlike previous sequential models (recurrent neural networks), can process an entire input sequence simultaneously, enabling parallel training on massive datasets. This scalability was further fueled by specialized hardware, with NVIDIA GPUs (A100/H100) becoming the industry standard, while Google developed proprietary tensor processing units (TPUs v4/v5p) specifically optimized for training its Gemini models [3].

The growth of these models follows the so-called Scaling Laws [4,5], stating that performance improves predictably with increases in computing, data, and parameters. The most recent frontier—and the focus of this study—is native multimodality in multimodal large language models (MLLMs), in other words, foundation models trained from the outset on mixed modalities (text, images, audio, video) rather than combining separate unimodal components post hoc. In this sense, Gemini is an example of a native MLLM, pretrained on multimodal data, enabling deeper cross-modal semantic alignment between video, audio, and text. This capability can be further enhanced by Mixture-of-Experts architectures [6,7], which balance massive knowledge with inference efficiency.

The emergence of standards such as the Model Context Protocol enabling seamless interoperability and tool integration, and new capabilities and uses for foundation models [1,8,9,10], enhances hope that AI tools may also acquire sophisticated skills of discourse analysis.

This paper, after overviewing recent works on the application of MLLM to discourse production, comprehension and analysis (Section 2), takes the Trump–Zelensky meeting of 28 February 2025 as a case study for the analysis of multimodal speech. After recalling the literature on the subtleties of body communication in that meeting in Section 3, Section 4 analyses it in terms of a theoretical model of multimodal political communication, and Section 5 illustrates a study instructing the Gemini Pro 2.5 model in this analysis. Gemini’s analyses resulting from subsequent prompts are compared to a human’s, proposing a benchmarking model with criteria and measures to assess their discrepancies. Section 6 summarizes the successful results attained and highlights the effectiveness of instructing an MLLM in the proposed analysis, whereby large chunks and sophisticated aspects of multimodal discourse (so highly time-consuming when run by humans) are correctly interpreted in a fast but reliable way.

2. Related Work—Generative AI in Text and Discourse Analysis

Generative AI may be usefully applied to the qualitative analysis of discourse [11,12]. This is generally a highly time-consuming and error-prone task, and it would be very useful to have AI carry out the transcription, categorization, and quantification of speeches and debates.

After a first automatization of discourse analysis in the first years of the century, through tools for automatic qualitative analysis and the detection of semantic contents, from political discourse to diaries of hospital patients (e.g., LIWC [13]), the rise of foundation models now allows for a deeper analysis of discourse, including, besides the bare count of words, the skill to make inferences and capture implicit contents.

The first attempts at the automatic coding of discourse did not achieve standard thresholds [11]. In some text annotation tasks (e.g., identifying types of social media posts), these models outperform crowd workers [14], but in other cases, they make trivial mistakes unlike humans: as goes the “generative AI paradox”, foundation models are better at producing than at evaluating texts.

However, presently, works on text analysis exploiting large language models are flourishing. Text analyses are conducted on textbox questionnaire responses, interview data, and transcribed think-aloud data: ChatGPT is able to generate sentiment scores [15], GPT 3.5-Turbo can perform topic segmentation and inductive thematic analysis (TA) and single out explicit and latent meanings in qualitative data, showing itself capable of inferring most of the main themes with a good degree of validity [16]. Hamilton et al. [17] also found that human coders sometimes recognize some themes that ChatGPT does not, and vice versa. Chew et al. [18], exploiting LLM-assisted content analysis (LACA) to perform deductive coding, a qualitative research method for determining the prevalence of themes across documents, found that GPT-3.5 can often perform deductive coding at levels of agreement comparable to human coders, and demonstrated that LACA can help refine prompts for deductive coding, identify codes for which an LLM is randomly guessing, and help assess when to use LLMs vs. human coders.

Nguyen-Trung [19] introduced guided AI thematic analysis (GAITA), positioning the researcher as an intellectual leader guiding GPT-4 in four stages: data familiarization; preliminary coding; template formation and finalization; and theme development. Additionally, the ACTOR framework combines different effective prompting techniques when working with GenAI for qualitative research purposes.

Socially interesting topics are confronted in the field of LLM text analysis: Törnberg [20] showed how LLMs can identify populism in political texts, while Chiu et al. [21] found that with zero- and one-shot learning, GPT-3 could identify sexist or racist text with an average accuracy between 55% and 67%, depending on the category of text and type of learning. With few-shot learning, the model’s accuracy could be as high as 85%. On a similar topic, Huang et al. [22], testing ChatGPT’s capacity to provide natural language explanations (NLEs) for implicit hateful speech detection, found that it correctly identified 80% of the implicit hateful tweets, and that for the remaining 20%, it aligned with lay people’s perceptions. Moreover, ChatGPT-generated NLEs can reinforce human perception and tend to be perceived as clearer than human-written NLEs. Hence, the authors raised an important ethical issue: if ChatGPT can be convincing, this might lead to the risk of misleading laypeople when its decision is wrong.

Other works have focused on conversation and dialogue analysis: according to Fan et al. [23], when instructed in two discourse analysis tasks, topic segmentation and discourse parsing, ChatGPT demonstrates proficiency in identifying topic structures in general-domain, sometimes better than humans, but is not as skilled in specific-domain conversations, and only linearly parses the hierarchical rhetorical structures that are in fact more complex than topic structures.

Labruna et al. [24] used ChatGPT to generate and annotate goal-oriented dialogues (task-oriented, collaborative, and explanatory) in English and Italian, and in two generation modes (interactive and one-shot). Evaluating them through an in-depth analysis, they concluded that their quality was comparable to dialogues generated by humans, but that the task of dialogue annotation schema, due to its complexity, still requires human supervision.

These works generally compare the coding by LLMs with previous human analyses [25], and some of them have more specifically focused on the issue of prompting. De Paoli [26] proposed open-ended prompts, Fan et al. [23] used ablation studies on various prompt components, while Huang et al. [22] proposed investigating the effects of different prompt designs.

Other works [27] have highlighted the ethical issues, a typical theme of critical discourse studies, raised by the use of LLMs for text production and analysis: the problems of power and inequality, authorship and the ownership of texts, linguistic homogenization, and the privilege of the mainstream induced by LLM usage.

Another scholar who recently explored the use of generative AI for qualitative discourse analysis is Susan Herring [12,28], a pioneer linguist who proposed CMDA (computer-mediated discourse analysis) as an approach to researching online behavior and also produced foundational work on online (im)politeness and “flaming” [29]. Wondering whether an MLLM could be taught to perform a complex discourse-analytic task, she asked GPT-4o and GPT-o3 to code a Reddit thread using CMDA conventions and compared their output to human coding. Through a detailed metric system to evaluate their discrepancies, distinguishing types of divergence, from “better than human” to “simply wrong”, and systematic errors in politeness coding (omission, over-application, incorrect assignment of type and tone), Herring [12] showed that differences between human and AI coding were frequent; that some AI errors resembled those of inexperienced students but that AI reasoning was often sound even when the coding was incorrect; and that prompt refinements can improve performance. Herring claims that for students, LLMs can support understanding and reduce workload but risk reinforcing shallow or mistaken interpretations, and for researchers, they offer scalability and potential contributions to coding scheme development, though their output remains imperfect. She stresses that LLM-assisted CMDA must be accompanied by careful human oversight, recommending systematic verification of AI outputs and continued direct engagement with the data [30].

A peculiar aspect of discourse analysis concerns multimodal discourse. The term multimodality may define, either (1) written text (e.g., in social media) that contains images, videos, graphics, gifs, where the issue is how to capture the semantics not only of words, but also of images, and their relationships [12]; (2) the intertwining of speech with body communication, whose analysis entails capturing, besides the meanings of words and sentences, those of intonation, gestures, facial expressions, gaze, body movements, and their reciprocal relationships.

Taking multimodality in this latter sense, to our knowledge, no systematic study has yet applied generative AI to the analysis of multimodal interaction. Our work aims at filling this gap by instructing an MLLM to analyze multimodal political discourse, extracting its rhetorical structure and the features of its persuasive import. The type of analysis we instructed the AI model in was the LEP (Logos, Ethos, Pathos) analysis, based on a socio-cognitive model of persuasion and multimodal communication, and its specific object was the Trump–Zelensky debate.

3. The Ten Minutes That Shocked the World—A Case Study

The Trump–Zelensky meeting of 28 February 2025, concerning the possibilities of stopping the Russia—Ukraine war, went down in history for unhinging all the rules of diplomacy and causing a heavy impact on future geopolitics. It was held publicly in front of Trump’s important collaborators, ministers, and Vice President Vance, open to journalists’ questions, and followed a set of previous meetings where Trump tried to tune a deal with Zelensky regarding the exploitation of Ukrainian raw earth, clearly implying this deal was a precondition to any further help from the U.S. to Ukraine, and somehow a payment for the weapons that the U.S. had previously provided to Ukraine.

3.1. The Meeting

The meeting “https://www.youtube.com/watch?v=UhFqdAZ9Xnw (accessed on 26 January 2026)” at first follows the diplomatic protocol: greetings, ritual formulas, and reminders of friendship between the U.S. and Ukraine. Then, Trump mentions that he has spoken with President Putin, and that he also wants to bring the war to an end. Trump shows sadness and worry about so many soldiers killed on both sides, hoping that all the money will be put to different uses, and the rebuilding. While acknowledging how brave Ukrainian soldiers were, he places the blame of letting this war start and continue on Biden, while it would have finished immediately had he, Trump, been there at the time.

Zelensky thanks him for the invitation and support, but also asks that the deal include security guarantees for Ukraine, and counts on the U.S.’s strong position to stop Putin. After proposing to share the license of Ukrainian drones in exchange for that of U.S. air defense, he goes into more detail about the unfortunate Ukrainian situation—20,000 stolen children and mistreated prisoners—implying a very bad picture of the Russians.

Trump reiterates his pragmatic stance, insisting on rare earths but leaving out the guarantee issue, while reminding us of the dead soldiers. Whilst seemingly kidding, he compares himself to Washington and Lincoln, highlighting his activism aimed at sparing so many deaths in Russia and Ukraine. As Zelensky reminds us that it was the Russians who invaded their territory, and that Putin broke the deals and ceasefire 25 times, Trump magnifies his own abilities as a mediator, criticizing Biden for sending a lot of weapons and Europe for not doing enough, and blaming the hatred shining through Zelensky’s words, which does not facilitate negotiation.

Conflict escalates in the last ten minutes of the meeting (39.54–49.20). Vice President J.D. Vance reminds Zelensky that the right path to peace is diplomacy, as pointed out by Trump, and when Zelensky shows distrust in diplomacy, given the many agreements signed but broken by Putin, Vance complains of his lack of respect and gratitude to President Trump. As Zelensky remarks that the U.S. do not have the same war problems as Ukrainians just thanks to an ocean in the middle, and asks them to imagine how the U.S. would feel if they were in the same situation, Trump, showing anger and offence, tells Zelensky that he is in no position to dictate what Americans are going to feel, that he does not have the cards, and that he is gambling with World War III; he again blames his ingratitude, remarking that he is running low on soldiers who, although brave, would be nothing without U.S. weapons. Finally, Trump stops the meeting, sarcastically calling it a piece of great television.

3.2. Body Language and Stances in the Meeting

The Trump–Zelensky meeting at the White House became a focal point of international attention not only for its diplomatic implications, but for the intense and revealing non-verbal exchange between the two leaders. According to analysts Caroline Goyder and Darren Stanton “https://www.washingtonpost.com/world/2025/03/01/zelensky-trump-video-body-language/ (accessed on 26 January 2026)”, the encounter deviated sharply from conventional protocol, unfolding more like a media spectacle than a diplomatic negotiation. Goyder emphasized that the usual “courtly behavior pattern” was visibly stripped away. First, no interpreter was used, increasing the potential for tension and miscommunication; second, the evident power imbalance between Trump and Zelensky, as to both military status and geopolitical leverage, was played out nonverbally, and they embodied contrasting performance frames: the former, calm and courtroom-like, signaled control from a position of domestic stability; the latter, displaying “warrior energy”, brought urgency, military posture, and emotional intensity shaped by Ukraine’s wartime reality. Stanton observed that Zelensky’s posture shifted from engaged to defensive, particularly at Vance’s intervention, with gestures displaying frustration and a solely ego-driven communication. Trump, in contrast, deployed his classic dominance repertoire—bone-crusher handshakes, power pats, and steepled fingers. Paul Boross “https://news.sky.com/video/what-does-the-body-language-between-trump-and-zelenskyy-tell-us-13319187 (accessed on 26 January 2026)” framed the event as a clash of alpha energy versus moral resilience: Trump’s wide stance, pointed gestures, and aggressive gaze were met by Zelensky’s composed presence, sincere hand movements, and steady eye contact: a non-verbal duel where Trump sought to overpower, while Zelensky aimed to project unshakable integrity and national dignity. Khyati Bhatt “https://simplybodytalk.com/blog/trump-zelensky-body-language-analysis/ (accessed on 26 January 2026)” interpreted the exchange as an alpha–submissive dynamic, with Trump asserting territorial dominance through expansive postures and Zelensky exhibiting physical constriction, anxious facial expressions, and appeasing gestures—a cue to a leader under pressure. Traci Brown “https://youtu.be/yxI-8yaVsZE?feature=shared (accessed on 26 January 2026)” traced the meeting’s progression from polite gestures to open confrontation, highlighting Trump’s use of the accusatory finger-point as a symbolically loaded act of aggression and Zelensky’s upward chin tilt and heart-touch as signals of moral defiance. Judi James “https://www.dailymail.co.uk/news/article-14449261/donald-trump-white-house-smackdown-zelensky-jd-vance.html (accessed on 26 January 2026)” described the encounter as a high-stakes chess match, where Trump’s physiological escalation—from flushed cheeks to commanding gestures—marked a transition from calculated authority to combative dominance, and emphasized Vance’s role as Trump’s “provocateur”, igniting tension and enabling Trump to deliver the climactic verbal blows.

These expert readings agree that beneath the surface of formal diplomacy, the meeting exposed raw, performative power struggles, where every gesture, glance, and shift in posture contributed to a psychological battle in full view of the world.

This jeopardizing event, beyond analysis by political commentators, may go through a more systematic and detailed view, exploiting one of the numerous discourse analysis approaches, from classical rhetoric or pragma-linguistics to lexicographic automatic analysis. However, these models, as applied by humans, are highly time-consuming, especially when annotating not only the verbal content of the interaction, but also its multimodal communication; therefore, it would be useful to instruct an MLLM on how to apply one of these models.

In the following, we illustrate a socio-cognitive model of multimodal political discourse, analyzing, as an example, a fragment from the Trump–Zelensky meeting.

4. A Socio-Cognitive Model of Persuasion and the LEP Analysis of Political Discourse

According to a socio-cognitive model that interprets mind, social interaction, emotion, and communication in terms of goals and beliefs [31,32,33,34,35], communication occurs when a Sender performs a verbal or bodily communicative act, that is, a sentence (speech act) or a gesture, gaze, facial expression, posture (nonverbal act) that conveys a performative of information, question, request, and a propositional content—the information given or asked, the requested action. Therefore, communicating is always an act of influence, since it asks an Addressee to do something, assume some information, or provide one.

Persuasion is a communicative action aimed at having someone do something: a form of social influence (i.e., raising or lowering the likeliness of someone to pursue some goal) [32]. A influences B when A causes B to pursue some goal GA or not to pursue some goal GB. In the peculiar social influence of persuasion [36], persuader A proposes persuadee B to pursue a goal GA by showing that bringing about GA is a means for B to achieve one’s previous goal GB. Unlike manipulation, in which B is persuaded without realizing that someone wants to persuade him, in persuasion, A communicates to B one’s intent to influence, and finally B accepts pursuing the proposed goal not through seduction or induced by a promise, threat, or the use of force, but because B is convinced that the goal proposed by A is indeed a means to achieve one’s goal GB. To bring about this process, the persuader exploits three different strategies proposed by Aristotle’s “Rhetoric” (325 BCE): logos (rational argumentation), ethos (the Orator’s character), and pathos (the Audience’s emotions).

The first strategy, logos, is logical argumentation: to provide pertinent information and good reasons (i.e., to highlight the other goals of B that motivate or enhance B’s choice to pursue the goal GA proposed by A as a means for B’s goals).

Ethos is the strategy of enhancing trust in the persuader: we are convinced not only by what someone tells us, but also by who tells us this. Therefore, the persuader must project a positive image of oneself to elicit the persuadee’s trust and make one’s arguments more credible. B trusts A [37] when B believes s/he cannot or does not want to pursue one’s goal GB, and consequently delegates A to bring it about. However, to trust A, B must believe (1) that A has the competence—intelligence, knowledge, comprehension, inference and planning skills—that are required to bring about B’s goals, and (2) that to realize B’s goals, A will act in the interest of B, and not in one’s own: that is, B also relies on A’s benevolence. This always holds in everyday interaction: a car seller or a dentist must be both expert and honest. However, in politics, one more characteristic is required to obtain the electors’ trust and their votes: dominance, the capacity of being and showing more strength than the opponent, to exert one’s power over others, and not to submit to others’ argumentations and decisions.

Political communication is generally a triangular persuasive job: in rallies and debates, both politicians A and B aim to persuade the Voter V; and to have V vote for A, A must persuade them that A is better than B. Therefore, politicians must display a positive image of themselves, conveying “I am more (good, smart, strong) than my opponent”. However, if one (even unconsciously) intimately feels that they are not really endowed with such skills, one can take another route: to discredit their opponents (i.e., to diminish their image).

This is why the political orator, in speeches, debates, and social media, while working on one’s own ethos, to project a positive image of oneself, also often works on the opponents’ ethos by trying to diminish their image, delegitimate them by criticism, calumny, accusation or gossip, insult or ridicule, aiming to highlight one’s own merits and remark on the others’ demerits, alternating communicative acts of self-praise and discredit.

The third Aristotelian persuasive strategy, pathos, consists of eliciting and managing the persuadee’s emotions to induce them to pursue the proposed goal. Although the appeal to emotions is considered a fallacy by argumentation theory (e.g., [38]), emotions are often and heavily exploited in political communication [39,40] thanks to their high motivating power due to their function of monitoring the primary goals of people, and to their triggering urgent and pressing goals, for example, “I feel fear if my goal of survival is at risk; anger if someone thwarted my goal of justice; compassion if my goal of others’ well-being is thwarted”, and anger triggers aggression, fear, flight, compassion, and the impulse to help.

The three strategies of logos, ethos, and pathos can be all exploited in persuasive discourse, including political speech [35,36], but different speakers may employ them in different ways, relying more on rational argumentation, on the presentation of oneself or the opponent, or on the emotions they can trigger. Moreover, in their self-presentation and discredit strategy, the evaluations they stress more—competence, benevolence, or dominance—can tell us something about what values are most important for them, and in the emotions they most often express or try to trigger, we see which goals are most important for them. Therefore, by carrying out an LEP analysis (Logos, Ethos, Pathos analysis) [41] of political speeches and debates, we can characterize the respective “political discourse style” of different politicians by comparing the relative frequency with which they employ the three strategies: how much they go into the details of political programs (logos), whether they highlight positive aspects of self or negative aspects of their opponents (ethos), or leverage on emotions (pathos), and how their combination of preferences characterizes their discourses.

4.1. The LEP Analysis of Multimodal Political Discourse

Poggi et al. [41] presented the LEP model (LEP analysis of political discourse) and applied it to the qualitative analysis of four inauguration speeches delivered by three Italian Prime Ministers. The LEP analysis showed that it was particularly fit to detect the main aspects of their personal history, their communicative style, and their political ideology. In that work, the analysis was carried out only on the verbal side of each speech, but a detailed quantitative analysis inspired by the LEP model can also be carried out on a whole multimodal arrangement of speeches.

In speech, the sentences produced are only a small part of what a speaker intends to communicate: communication is multimodal in that intonation and prosody, face, gaze, gesture, and posture all convey meanings, and even not considering the information unconsciously leaking from the speaker’s body behavior, such as manipulation gestures (licking one’s lips, grasping a pen), a great part of the contents borne by the body are deliberately conveyed by the speaker as functional to the transmission of the global intended meaning. Also in political communication, the behavior of the body is sometimes as relevant as the verbal one because it can not only confirm but also add to and complete the meanings conveyed verbally, or sometimes even possibly contradict them. A staring gaze may convey a person’s dominance, a contemptuous face can discredit the interlocutor, and a fast moving gesture can incite and transmit enthusiasm; all these pieces of information make up the whole mosaic of a politician’s communication, with its peculiar combination of logos, ethos, and pathos. Moreover, each communicative item, whether word or bodily signal, besides its literal meaning, may also convey one or more indirect meanings: presupposition, implications, or other semantic contents that the sender wants to be inferred by the addressee.

Therefore, to provide a more complete grasp of the differences in communication style among politicians’ speeches, an LEP analysis can be conducted on multimodal communication according to an annotation scheme where each verbal or bodily communicative item (sentence, prosody, gesture…) is attributed both its literal and its possible indirect meanings, and these meanings are classified in terms of logos, ethos, and pathos. This work is the first in which a multimodal LEP analysis was exploited. Let us see (

Table 1. Multimodal LEP analysis. An annotation scheme of multimodal political communication.

1 Time and Sender	2 Modality	3 Signal	4 Literal Meaning	5 Indirect Meaning	CLASSIFICATION
					6 Ritual	LOGOS	ETHOS		PATHOS
						7 Logos	8 Our Merits	9 Their Demerits	10 Expressed Emotion	11 Induced Emotion
13.38 Trump	Verb	There has never been a first month like uh like we’ve had	I did many things in my first month as a President	I am better than all other Presidents			C		Pride
	Gaze	Looks forward	I address you all	I am confident					Self-confidence
13.52 Trump	Verb	Think of the parents whether they are Russian or Ukraine	You should feel compassion for both Russian and Ukrainian parents	I am neutral, I do not take sides for Ukraine → this is how to make a deal + I am skilled in making deals		X	B C			Compassion for Russian victims’ parents
	Head	Shakes head	No, no	We must stop this tremendous death		X
	Gaze	Squints eyes	I feel sorrow for them	You should feel compassion					Sorrow	Compassion
13.52 Zelensky	Verb
	Head	Turns face rightward (the opposite direction of Trump)	I do not want to hear him while saying so	I do not agree with Trump				PS -B
	Mouth	Left corner lowered	I feel disgust for Russians	Russians are evil → They do not deserve compassion → I do not agree with Trump				PS TB -B	Disgust	No compassion

Legenda: PS = Present Speaker, B = Benevolence, C = Competence, → = indirect meaning

) how fragments of the Trump–Zelensky meeting can be analyzed in terms of this model.

Table 1 shows the annotation scheme of the multimodal LEP analysis. In the first three columns, we list who the Sender of the analyzed communicative item is (Time and Sender, col. 1), the modality of that item (Modality, 2), and the transcription of the verbal signal or a description of the bodily signal (Signal, 3). In the following two columns, we list the semantic analysis of the signal: its literal meaning in col. 4 (e.g., a verbal translation of the gesture or gaze under analysis) and its indirect meaning (if any) in col. 5. Columns 6 through 11 contain the classification of the communicative moves performed by the meanings in cols. 4 and 5: whether barely ritual (col. 6), or bearing on logos (7), on ethos—the Sender’s and their side’s merits (8), the Opponents’ demerits (9)—or finally on pathos: the emotions expressed by the Sender (10) and ones s/he wants to elicit in the Audience or the Opponents (11).

Here, for example, Trump, at minute 13.38 (Col. 1) says (Verbal modality, col. 2): “There has never been a first month like uh like we’ve had” (col. 3), an informative sentence conveying (Literal meaning, col. 4) “I did many things in my first month as a President”, which, however, implies (indirect meaning, col. 5) an evaluative information: “I am better than all other Presidents”, a self-praise about competence (C, col. 8, Ethos our merits), and displays pride (expressed emotion, Pathos col. 10). In the gaze modality (subsequent line, col. 2), Trump, at the same time of his sentence, looks forward at the camera and the Audience, conveying (literal meaning, col. 3) “I address you all” (col. 4), which implies “I am confident” (5), thus expressing the emotion of self-confidence (col. 10).

At 13.52, Trump’s requestive act, think of the parents whether they are Russian or Ukraine (col. 3), means “you should feel compassion for both Russian and Ukrainian parents” (4), but indirectly implies informative acts (5) “I am neutral, I do not take sides for Ukraine”, further indirectly conveying “this is how to make a deal”, which again highlights the evaluative information “I am skilled in making deals”. The first implication is a communicative act of Logos (col. 7) but also conveys Trump’s own Benevolence (8), while by the other two implications, he again self-praises his own Competence; and the general meaning of the verbal act tries to induce compassion in the Audience (col. 11). At the same time, his head-shake (col. 3) means “no no” (4), implying “we must stop this tremendous death” (i.e., so many people dying) (5), a requestive act of Logos (7), while by gaze (2), his squinting eyes (3) perform an informative act of expressing sorrow (10) in order to induce compassion (11).

The very words “whether they are Russian” elicit the (only bodily) reaction of Zelensky who, at the same time (13.52), in the head modality (col. 2) “turns face rightward (the opposite direction of Trump)” (3), meaning “I do not want to hear him while saying so” (4), conveying the implication “I do not agree with Trump” (5): an Ethos move discrediting PS (the Present Speaker Trump) (col. 9) for lack of Benevolence (-B). At the same time, Zelensky’s mouth (2) “lowers left corner” (3), meaning “I feel disgust for Russians” (4), implying “Russians are evil, they do not deserve compassion”, hence “I do not agree with Trump” (5), which casts discredit (col. 9) for a lack of Benevolence on both Russians (Third Person, TP) and Trump (PS), and by expressing disgust (10), tries to induce the Audience not to feel compassion for Russians (11).

These classifications of communicative acts can be quantified by computing how many of them fall in each class, and how many are performed in each modality: this outlines a “profile” of each interactant in the fragment, and somehow their communicative style. For example, from Table 1, the results show that Trump performs one verbal and one head act of logos, two verbal acts of self-praise for Competence and one for Benevolence, one verbal act expressing pride, and by gaze, he expresses self-confidence once, sorrow once, and induces compassion by one verbal and one gaze act. Zelensky conveys the others’ lack of benevolence via his head and mouth, and through his mouth, expresses disgust and tries to induce a lack of compassion.

From these counts, Trump appears to be a person quite satisfied with himself, inclined to strive for the others’ good, and Zelensky as a bitter person, seeing others as enemies. However, this is the profile resulting from these few lines only: in many other cases, Trump displays all of his narcissism, his primary interest in money and power, his contempt and criticism for Biden’s stupidity, and his attempt to induce humiliation in Zelensky; whereas Zelensky shows his firmness in confirming Russian responsibility for the war (-B) but also fairness to Europe who always backed his actions (B), and tries to elicit worry even in Trump and Vance for Putin’s possible action toward Europe and the USA.

4.2. Can Generative AI Run a Multimodal LEP Analysis?

As may be clear from this example, the LEP analysis of multimodal speeches and debates is very effortful and time-consuming: brief fragments of communication require hours of annotation and classification of all communicative signals. A challenging task, then, right for MLLMs.

5. Instructing Gemini 2.5 Pro in the LEP Analysis of a Political Debate

To test this hypothesis, we designed a study aimed at assessing if and how an MLLM can be instructed to categorize a multimodal political discourse in terms of the LEP analysis. As a case study, we selected the Trump–Zelensky debate, and as the MLLM to instruct, we used Gemini (Pro 1.5/2.5) via Google AI Studio, given its specific technical advantages crucial for political discourse analysis. First, Gemini stands out as a native multimodal model with an extensive context window (up to 2 million tokens), allowing it to process the entirety of the Trump–Zelensky meeting without fragmentation. Crucially, Google AI Studio offers a unique feature: direct YouTube integration. Unlike other MLLMs that require complex pre-processing (downloading videos, extracting frames, or transcribing audio separately), Gemini can ingest and analyze YouTube videos directly. This capability allows the model to “watch” and “listen” to the debate in its original format, processing non-verbal cues (prosody, facial expressions) alongside the verbal transcript in a unified inference stream. This technological affordance is essential for applying the LEP framework to the interplay between words and body language.

Given the length of the meeting (49 min), before feeding Gemini 2.5 Pro pieces of the dialogue and the instructions on how to analyze it, it was necessary to select only some fragments of the Trump–Zelensky meeting. Such selections required a process in more phases that combined human and AI contributions.

The first phase required a thorough human analysis of the meeting, and the second included: (a) the construction of prompts to make the model analyze fragments of the meeting, (b) comparison of its analyses with the human’s, and (c) optimization of a final prompt, aimed at making the model as much or even more of an expert than the human in the LEP analysis.

5.1. Method

The human analysis of the debate included three steps.

Debate transcription: the integral video of the meeting was transcribed through the Turboscribe “https://turboscribe.ai (accessed on 26 January 2026)” tool, resulting in a faithful transcription of the dialogue as well as encompassing pauses, overlaps and interruptions, timing, and distinction of different speakers, thanks to the tool’s advanced skills of vocal recognition.

Fragment selection: after thorough reading, the most significant fragments of the meeting were selected for a whole detailed LEP analysis: namely,

• the fragment from minute 13.38 through to 13.57 was chosen to provide a detailed model of analysis within the prompt to feed to Gemini 2.5. Pro. Here, Trump urges for a deal to be made given the “tremendous death” taking place, and invites the Audience to think of the parents of the dead soldiers, both Russian and Ukrainian. Zelensky first reacts with physical signs of tension to Trump’s (for him) excessive equidistance from the two contenders, and finally blurts out “They came to our territory!”.
• This fragment is interesting, mainly as a test of whether the GAI model also decides to represent and classify the bodily communication delivered by Zelensky, who in this case is but the Interlocutor, and not a Present Speaker.
• three fragments from the last 10 min of the meeting, 38.18 through 49.22. This is the most crucial part of the debate: the level of aggression of Trump and Vance toward Zelensky raises, jeopardizing his possible previous expectations about the conclusion of the debate, and his hope to obtain some help for Ukraine from the USA. It is also the most widely discussed, commented, and viral part of the meeting spread on social media. These are the three fragments analyzed:

  ▪

  (39:14) Trump:

  I’m aligned with the world and I want to get this thing over with. You see the hatred he’s got for Putin. It’s very tough for me to make a deal with that kind of hate. He’s got tremendous hatred and I understand that, but I can tell you the other side isn’t exactly in love with, you know, him either. (39:31)

  Here, Trump tries to shift the blame for the failed agreement onto Zelensky, and does so by also exploiting an ironic euphemistic understatement; the fragment then provides a chance to test Gemini’s skill to catch indirect meanings and rhetorical figures.

  ▪

  (40:33) Zelensky:

  Sure. Yeah. Yeah. Okay. So he occupied it, uh, our parts, big parts of Ukraine, parts of East and Crimea. Uh, so he occupied it on 2014. So during a lot of years, I’m not speaking about just Biden, but those time was Obama, then President Obama, then President Trump, then President Biden. (40:58)

  This excerpt shows Zelensky while rattling off facts, years, and names to explain the context without controversy, thus presenting him as consistent, rational, and respectful, but also precise and strategic. We chose this to test Gemini’s knowledge and skill to take geopolitical issues into account.

  ▪

  (44:45) Trump:

  You’re not winning. You’re not winning this. I you have a damn good chance of coming out.

  (44:50) Zelensky:

  Okay. Because the President we are staying in our country, staying strong from the very beginning of the war. We’ve been alone and we are thankful. I said thanks.

  (44:58) Trump:

  You haven’t been in this cabinet. We gave you the President three hundred and fifty billion dollars. We gave your military equipment. Your men are brave, but they had to use our military. (45:10)

  This last exchange is relevant for the level of aggression reaching white heat, with Trump’s attempt to humiliate Zelensky and pretend his gratitude and submission, and Zelensky defending the dignity and legitimacy of his struggle: a test for Gemini’s skills in detecting, among others, expressed and induced emotions.

Human LEP Analysis of the three fragments: Three of the authors provided a qualitative analysis of the three selected fragments transcribing, for each interactional move, the verbal and nonverbal signals, and classifying them in terms of logos, ethos, and pathos. The annotations were carried on independently and then discussed together by finally finding an agreement on possibly divergent descriptions or classifications. This first human analysis was a basis for reference and comparison with Gemini’s analysis.

The second phase, MLLM analysis, included:

Prompt construction and MLLM analyses

Two distinct prompts—a “synthetic prompt” and a “detailed prompt”—were fed to Gemini 2.5. Pro asking it to analyze the last 10 min of the meeting.

The “synthetic prompt” simply requested the MLLM to annotate, for each verbal expression, gesture, posture, or modulation of voice (1) the minute of occurrence, its (2) literal, and (3) indirect meaning (including implications, subtexts, possible strategic or diplomatic interpretations), its (4) ritual meaning (i.e., the role of the phrase or gesture within the formal and informal communication dynamics of a political debate), and (5) classification in terms of logos, ethos (our merits and their demerits), and pathos (emotions expressed—e.g., anger, confidence, irony, frustration, hope—and to be elicited—e.g., fear, trust, indignation, compassion). The prompt also solicited it to highlight any discrepancies between verbal and non-verbal language. It finally asked to present the results in the form of a table, with a column for each of the listed criteria (see Supplementary Materials, 1. Synthetic Prompt).

The “detailed prompt” guided the MLLM step by step, describing the bulk of the LEP analysis, and incorporating a “Gold Standard” human annotation (fragment 13.38–13.57) as a reference example [42] (see Supplementary Materials, 2. Detailed Prompt).

The two prompts were fed to the MLLM and the two respective analyses were recorded and compared.

Comparison of the analyses stemming from the two prompts

The analyses resulting from the two prompts were quite different. The “synthetic” prompt showed some interesting strengths. First, the MLLM clearly identified the function of portions of discourse such as opening, closing, contrast, or stakes raising, and the global scheme of the meeting was quite clearly portrayed, providing a good basis for a general interpretation of the event. However, more interestingly, Gemini showed a fair grasp of the context and the political and rhetorical implications of the signals, for example, it captured:

• Indirect meanings
  Gemini catches Trump’s indirect meaning when it identifies that his reference to Zelensky’s “tremendous hatred” (39:22) for Putin is an indirect blame. It notes: “Shifts potential blame for negotiation failure onto Zelensky’s emotional state. Suggests Trump is the rational actor hindered by emotion”, thus recognizing Trump’s use of emotional framing to obscure responsibility, a sophisticated rhetorical strategy where the literal meaning (Zelensky hates Putin) masks the pragmatic implication (Zelensky’s emotions are an obstacle to Trump’s deal-making).
• Trump’s Self-Presentation as a Mediator
  As Trump claims, “I’m aligned with the world… I want to see if we can get this thing done” (39:14), the analysis captures: “Reasserts his primary focus is deal-making and resolution above taking sides. Emphasizes his role as a facilitator for broader resolution”, recognizing that Trump is employing strategic ambiguity—appearing neutral while simultaneously establishing his authority as the problem-solver.
• Hidden presuppositions and manipulative moves
  A particularly striking exchange demonstrates Gemini’s ability to identify manipulative framing. Trump’s statement “You’re playing cards… you’re gambling with the lives of millions of people” (43:53) employs a metaphorical escalation that the MLLM correctly parses: “Trump escalates the metaphor to accuse Zelensky of recklessly endangering lives. Turns Zelensky’s situation into a high-stakes gamble he is losing without U.S. help.” This reveals the MLLM’s recognition that Trump is exploiting the tragic situation (death of Ukrainian citizens) to reframe Zelensky’s negotiating position as irresponsible gambling. The hidden presupposition is: “Zelensky’s refusal to make concessions will cause Ukrainian deaths”.
• Manipulative exploitation of “gratitude”
  The MLLM correctly identified the exploitation of gratitude as a power dynamic. When Vance blames Zelensky for not showing gratitude and explicitly pretends it, the analysis notes: “Vance attempts to reinforce the narrative of Zelensky’s ingratitude by demanding a specific, performative act of thanks… Zelensky pushes back, claiming prior thanks”. The MLLM identified how thanking Trump serves an instrumental political function: it would constitute public acknowledgment of dependency and validate Trump’s narrative of American generosity. By demanding this ritual performance, Vance (and implicitly Trump) seeks to establish a hierarchical relationship where gratitude becomes a form of subordination.

Despite these brilliant performances, the output of this first prompt was lacking on the microanalytic side, which was the bulk of the LEP analysis. Here are some of its flaws:

• The MLLM’s table lacked some relevant columns: it did not distinguish the signals by modality, not specifying verbal, vocal or other; no marking of the signals opening and closing the interaction; the category ethos does not set apart “our merits” and “their demerits”, nor does pathos distinguish expressed from induced emotions;
• The analysis mainly takes verbal and vocal behavior into account; only some gestures and postures are quoted, while face, eyes and body are almost never described;
• The Interlocutor’s behavior concomitant with the Sender’s is not taken into account, thus lacking any analysis of the interaction and bodily or vocal response of the Interlocutor (e.g., its signals of acceptance or disagreement).
• The analysis failed to classify the analyzed signals in terms of the LEP categories, such as “Pathos–Disgust”, “Ethos–Competence”, or the like, and was confined to a set of discursive descriptions not systematically coded. This prevented any kind of computation and quantitative comparison of the assessed phenomena.

More generally, the output generated based on the “synthetic” prompt looked more like a discursive interpretation than a detailed semantic/interactional analysis of the fragment.

Coming to the “detailed prompt”, the output was more complete and correct than the former, resulting in a table strictly complying with the proposed criteria and containing a fairly thorough LEP analysis. However, since the prompt was long and complex, the 10 minutes were split into six segments, each time providing the MLLM with the same prompt and asking it to resume from the preceding segment. This iterative approach [43] allowed the model to maintain context fidelity across the timeline, though it occasionally led to fragmentation: Gemini omitted pieces of dialogue or stopped just at key points, wasting the logical continuity and coherence of the analysis. Furthermore, like with the synthetic prompt, it only analyzed the Speaker’s communication, overlooking the Listener’s reactions. Moreover, sometimes it filled the column concerning Pathos with some mental states that were not strictly emotions, such as “pragmatism”, “determination”, and “autonomy”. Finally, since a quantitative table was not displayed in the prompt, Gemini issued different, sometimes ineffective, forms of schematization. Only the analysis of Segment 5 was close to the expected model, with better clarity of classifications and higher adherence to the LEP format.

Despite these problems, the MLLM analysis showed good levels of interpretative inferences, sometimes capturing the rhetorical and strategic implications of the analyzed discourse, for instance, highlighting attempts to shift responsibility, mark superiority or deference, or build an ideological frame of legitimization for one’s discourse.

Meta-Prompting and Automatic Optimization

Finally, an advanced “Meta-Prompting” strategy [44] was used. A secondary LLM (ChatGPT-4o) served as an optimizer [45], refining the instructions based on Gemini’s previous errors (e.g., overlooking the interlocutor). This recursive process generated an “Optimized Prompt” that achieved the highest performance.

ChatGPT-4o was provided with (1) the theoretical principles of LEP analysis, with various examples of application, (2) the “detailed” prompt fed to Gemini 2.5 Pro, and (3) the resulting analysis of the analyzed fragments, with a description of the errors discussed above.

For a more accurate analysis, GPT-4o was advised to add explicit instructions to prevent Gemini 2.5. Pro from overlooking portions of the dialogue, to always include both the Sender and Interlocutor, to add a list of emotions to include in columns 10 (expressed) and 11 (induced), and to propose an example of a quantitative summary table. The new “optimized” prompt was then fed to Gemini 2.5. Pro, asking it to provide an LEP analysis of the same three fragments analyzed by human annotators (see Supplementary Materials, 3. Optimized Prompt).

The results were promising, because the errors found in the previous analyses were overcome: (1) Aside from the Sender, the Interlocutor’s communication was also accurately analyzed; (2) columns 10 and 11 were only filled with pertinent emotions, coherent with the persuasive goals of the discourse; and (3) the tables were correctly organized.

However, although obtaining a general idea that the new analysis by the MLLM was quite successful, we opted for a more strict assessment of its accuracy.

5.2. Benchmarking Multimodal Reasoning—A Model to Evaluate an MLLM’s LEP Analysis

While traditional benchmarks like the MMLU (Massive Multitask Language Understanding) test general knowledge [46], evaluating a model’s ability to interpret complex visual dynamics requires newer standards. Benchmarks such as MMMU (Massive Multi-discipline Multimodal Understanding) [47] and MathVista [48] have set the bar for visual reasoning. However, analyzing political rhetoric involves subtle socio-emotional intelligence that standardized tests often miss [49,50]. In this study, we made a first attempt to propose one.

We built a benchmarking model aimed to assess the similarities and discrepancies between the human and MLLM analyses by comparing them according to specific criteria of evaluation and finally rating the AI’s performance.

In our comparison model (

Table 2. Comparison, evaluation, and rating of the MLLM’s performance.

1 Signal	2 Human	3 AI	4 Comparison	5 Evaluation	6 Rating
Section S (Signal)
Signal (Voice)	Slightly raised pitch, faster pace	High pitch, stressed tone on “hatred”	D2	BC	1
Signal (Gaze)	Fixed Gaze	Maintains direct eye contact with Trump	D1	W	−2
Signal (Gesture)	Right hand, index finger slightly extended, moves slightly rightward to Z.	Right hand gestures toward Zelensky (“him”), then left hand gestures away (“the other side”)	D2	BC	1
Section M (Meaning)
Literal Meaning (Verbal)	Zelensky hates Putin	Zelensky has a lot of hatred for Putin	D2	NR	0
Literal Meaning (Body)		I am engaged but tense	R	BC	2
Indirect Meaning (Verbal)		I am benevolent enough to understand, but competent enough to see the bigger picture	R	BC	2
Indirect Meaning (Gaze)	This makes a deal impossible	I am revealing something important about him to you	D1	W	−2
Section C (Classification)
Ethos—Our Merits (Gestures)		C	R	BC	2
Ethos—Their Demerits (Gaze)	I. -B		A	W	−2
Pathos—Expressed Emotion (Verbal)		Frustration	R	BC	2
Pathos—Expressed Emotion (Gaze)	Alarm, Worry	Confidence	D1	W	−2
TOTAL SCORE					2

Legenda: D2 = Minor discrepancy; D1 = High Discrepancy; R = Redundancy; A = Absence; BC = Better Choice; W = Wrong; NR = Not Relevant.

), three distinct sections on subsequent lines respectively compare three aspects of the multimodal LEP analysis: signal, meaning, and classification.

Table 2 shows the comparison between the two annotations of the same fragment by the human and AI, distinguishing them into three sections. Section S (Signal, first 3 lines of the scheme) compares the MLLM’s annotations of the first three columns of the LEP scheme (in Table 1 above, col. 1 Time and Sender, 2. Modality, 3. Signal) with ones agreed upon by the three human annotators (Table 2, col. 2 Human) to assess whether the AI (col. 3) correctly identified the type of signal (verbal, vocal, gestural…), the Sender, and the exact minute. Section M (Meaning) compares columns 4 and 5 of Table 1 (Literal and Indirect Meaning) to assess whether the AI caught the signals’ meaning, but also the implications that can be inferred from them. Section C (Classification) compares the human and AI as to columns 6 through 11 of Table 1 to assess whether the meanings extracted received the same classification by the human and AI in terms of the LEP categories: ritual, logos, ethos (our merits and their demerits), and pathos (expressed and induced emotions).

In all three sections of Table 2, the comparison went across six columns: in col. 1, we listed which item was compared, whether time, sender, modality, or signal; in col. 2 (Human), the human’s annotation, and in col. 3 (AI), the MLLM’s annotation as instructed by the optimized prompt for the same line and column of the human’s analysis; Col. 4 (Comparison) contains the results of the comparison between the values in columns 2 and 3, coded according to the coding system below:

• I (Identical): AI detects exactly the same value as Human (H);
• A (Absence): H detects the value, AI does not;
• PA (Partial Absence): AI detects part of the correct value, but omits other aspects;
• R (Redundancy): AI detects a value H did not detect;
• PR (Partial Redundancy): AI detects something more than H, and partially coherent;
• D1 (High Discrepancy): Both AI and H detect a value, but very different between them;
• D2 (Minor Discrepancy): Both detect a value, with small or only formal differences.

Column 5 (Evaluation), except for cases classified as “Identical” in Col. 4, contains a qualitative evaluation of the difference assessed in col. 4, trying to determine whether AI’s annotation is a more or less relevant error, or even a better choice than the human’s:

• W (Wrong): AI makes a mistake (e.g., it misses an important value or adds an inappropriate one);
• LR (Limited Relevance): AI makes a light or marginal error: the value can be considered either not so relevant or opinionable;
• BC (Better Choice): AI gives a better or more correct value than did H;
• NR (Not Relevant): The discrepancy between AI’s and H’s value is not relevant for the analysis.

As shown in

Table 3. Comparison, evaluation, and rating.

4 Comparison	5 Evaluation	6 Rating
A (Absence)	Wrong Choice	−2
A (Absence)	Better Choice	+2
PA (Partial Absence)	Wrong Choice	−1
R (Redundancy)	Better Choice	+2
PR (Partial Redundancy)	Wrong Choice	−1
D1 (High Discrepancy)	Wrong Choice	−2
D1 (High Discrepancy)	Better Choice	+2
D2 (Minor Discrepancy)	Not Relevant	0
D2 (Minor Discrepancy)	Better Choice	1
I (Identical)	Good Choice	1

, Column 6 (Rating) lists a numeric score attributed to the evaluation of col. 5, allowing one to quantify AI’s mistaken and correct analyses; each combination “Comparison + Evaluation” was converted into a positive or negative numeric score that measured the accuracy and pertinence of the AI’s annotations.

For instance, if the H analysis was (Table 2, col. 2) slightly raised pitch, faster pace and the AI’s was High pitch, stressed tone on “hatred”, it counted as a minor discrepancy (D2, col. 4), but the AI’s was a better choice (BC) than the H’s, so the rating was 1 (col. 6).

5.3. Results and Discussion

The sum of the final scores in all three tables (Signal, Meaning, Classification) allowed us to assess whether AI’s output was superior, inferior, or equivalent to human analysis, further providing a detailed analysis of the nature and quality of the discrepancies detected.

In general, in this benchmarking, the annotations concerning the Signal (Section S) showed a good level of accuracy of AI in associating the signals to the right time frame and to the Sender, with a fair amount of correspondences classified as “Identical”. However, some cases of “Absence”, mainly gaze items or micro-expression omitted by the AI, carried various “wrong” evaluations, negatively impacting the final score. Some cases of “partial absence” or “partial redundancy” were probably due to the AI’s tendency to summarize or expand the perceived signals too much. All in all, AI looks to be fairly skilled in perceiving the main signals, but not so much in detecting nonverbal signals and precise time segmentation.

Section M (Meaning) revealed quite a high performance of the MLLM in capturing the literal and indirect meanings of the two interactants: indirect meanings, including allusions, presuppositions, and implicatures, were quite precisely detected, even resulting in a significant number of “better choices”, with AI’s interpretations often more effective and synthetic than the H. For instance, concerning indirect meanings, AI showed a superior ability to identify the speaker’s communicative intent, achieving a “better choice” (BC) classification in several key instances. A notable example in Table 2 is at Line 3 of Section M (Verbal), where the H analysis failed to provide a full interpretation, while AI offered a highly synthetic and effective one: “I am benevolent enough to understand, but competent enough to see the bigger picture”. In a similar case, comparing the indirect meaning of a gaze item, AI correctly interpreted a sign of surprise as an underlying challenge: “I am challenging your characterization of me”. Furthermore, AI exceled in providing synthetic and more descriptive literal interpretations: for a literal meaning of a gesture, H simply noted “Him” by H, whereas the AI interpreted it as “I am balancing two opposing parties”, providing the “better choice”. Such cases confirm that the AI was particularly successful in translating subtle communicative cues into clear, concise, and contextually rich interpretations.

The small number of “wrong” and the high number of “identical” and “limited relevance” cases point to a fair skill of the AI to catch pragmatic nuances of the interaction. Only where, as found by Section S of the comparison, the signals was not detected, did this absence impact on the global interpretation, resulting in a low score.

Section C (Classification) of the comparison table assessed the final and most relevant aspect of the analysis: AI’s capacity to correctly classify the communicative moves as ritual, logos, ethos (our merits and their demerits), and pathos (expressed and induced emotion). Here, Gemini 2.5 Pro showed mixed results. It was very skilled at recognizing expressed and induced emotions, even singling out some like compassion, worry, anger, and annoyance. In the logos category, it often correctly found the logical elements, even if they were sometimes too generic. The classification of the interactants’ ethos moves, instead, provided contrasting results: Gemini distinguished positive from negative judgments, but did not always catch the criterion of the judgment (e.g., competence vs. dominance). Major flaws occurred in the ritual category, which was often omitted. A final problem is that the AI often did not attribute any LEP category to the Interlocutor’s nonverbal signals, whose classification instead is sometimes relevant for the LEP analysis.

Table 4. Rating of Gemini 2.5. Pro’s analysis.

Comparison	n.
Identical	28
Absence	37
Partial Absence	5
Redundancy	42
Partial Redundancy	10
High Discrepancy	11
Minor Discrepancy	22

summarizes the results of the comparison and evaluation of the AI vs. H performance.

Notwithstanding a fair number of “Identical” and “Minor Discrepancy” cases, the two annotations often yielded very distant results. However, beyond revealing formal discrepancies between the AI and H, this does not distinguish them from a qualitative point of view, for example, a case of absence may be a bare omission, or else instead a sensible choice of the AI, if that signal is in fact irrelevant. Therefore, if we also take into account whether any difference is simply “wrong”, “limited”, or “no relevance”, or even a “better choice” than the H (

Table 5. Evaluation.

Evaluation	n.
Wrong	35
Limited Relevance	25
Not Relevant	15
Better Choice	51

), it appears that the “wrong” cases (35) were far fewer than all other ones (91), namely 27.7% out of all discrepancies, while the “better choices” (51) were as high as 40.47% of them.

A further positive result stemming from the evaluation is that the AI did not seem particularly biased in its analysis. One problem often pointed out about MLLMs is that they are subject to biases induced by their knowledge base. Here, instead, Gemini’s analyses quite overlapped those of the human: even in cases of redundancy, where it highlighted something that the human did not, this does not contradict what may be implied by the rest of the human’s analyses. In general, Gemini does not seem to be biased toward one or another position, but was fairly impartial between the two interactants: for example, it was not particularly pro-American, although being issued in the States, and kept its analysis on a descriptive and not evaluative level.

Finally, thanks to this evaluation, we were able to tune a “final prompt” that can be used from now on to have an MLLM analyze a political debate in terms of multimodal LEP analysis (see Supplementary Materials, 4. Final prompt).

6. Conclusions

This study aimed to assess whether a multimodal large language model (MLLM) could be instructed, through prompt engineering, to analyze political discourse and political debates according to a multimodal LEP framework, identifying elements of logos, ethos, and pathos across words, voice, gesture, facial expression, gaze, and posture. Focusing on the Trump–Zelensky meeting held at the White House on 28 February 2025, we first conducted a human LEP-based analysis and then instructed Gemini 2.5 Pro using two different prompts. Through iterative testing and refinement, an optimized prompt was developed, and the model’s performance was evaluated using a benchmarking framework comparing AI and human analyses.

The results indicate that an MLLM can be successfully instructed to perform a theoretically grounded multimodal LEP analysis of political discourse, and in some cases, can match or even outperform human analysis. This finding is encouraging, given that comprehensive multimodal analyses of political speeches and debates are notoriously time-consuming and cognitively demanding, often limiting human research to short excerpts. An AI-assisted approach has the potential to extend such analyses to longer speeches or larger corpora, making them more feasible for research, journalism, and political communication analysis.

It is important, however, to clarify the scope and generalizability of these results. While the empirical findings cannot yet be generalized in a strong sense—given that the study examined a single case using a single model—the methodological approach is highly transferable. The iterative process of prompt design, testing, benchmarking against human analysis, and systematic evaluation can be replicated across different cases, contexts, and analytical frameworks.

From a technical standpoint, the experiments were conducted using Google AI Studio, primarily for reasons of accessibility and ease of use. This choice was pragmatic rather than methodological: the same procedures could be implemented through custom interfaces, APIs, or dedicated virtual environments, potentially allowing for more complex multimodal pipelines. At the same time, the use of an accessible interface highlights an important advantage, namely that advanced forms of discourse analysis through prompt engineering can be even made available to researchers without specialized expertise in machine learning or software development.

A further limitation concerns the exclusive use of a proprietary, state-of-the-art model. While this choice provides insight into the upper bounds of current MLLM performance, it also raises issues of reproducibility, dependency on commercial platforms, and data privacy. Given the rapid evolution of open-source multimodal models, it is plausible that comparable performance could soon be achieved with models running locally and adapted through fine-tuning or instruction tuning. Such developments would be particularly relevant for sensitive domains such as political discourse analysis, where data protection and privacy are crucial.

Of course, the work on LLMs as a tool for discourse analysis generally raises relevant ethical issues. Although Gemini in our case did not show much subject to bias, this is always a risk, which may have severe outcomes given the apparent “authoritativeness” that an MLLM can inspire. Other ethical concerns, although not so much for public contents like in our case, may be salient for verbal or visual contents exposing vulnerable subjects. To avoid such risks, it is important for the AI to always be supervised and monitored by human interpretation.

In fact, this work points toward a more interactive perspective on human–AI collaboration in discourse analysis. Rather than treating the MLLM solely as an analytical tool, future developments may explore richer dialogic interactions in which the model acts as an epistemic partner. Through iterative exchanges, the AI may highlight aspects of the discourse that the human analyst had not initially considered, thereby enriching interpretation, suggesting new analytical angles, and supporting a reflexive analytical process. In this sense, the contribution of MLLMs lies not only in their efficiency and scalability, but also in their potential to expand the interpretive space of multimodal discourse analysis.