Effects of promotional and preventive framing strategies in AI generated content on neural responses and trust Authors: Xiaoyi Wang, Raif Remigio Alves, Qiuyu Hu, Hanlin Chen Published: 2026 ABSTRACT Influenced by Regulatory Focus Theory, message framing is crucial in persuasive communication, particularly in marketing. This effect is further enhanced when integrated with artificial intelligence-generated content (AIGC), which offers personalized, scalable, and cost-effective solutions. However, the interaction between message framing, product types, and trust in AIGC remains unexplored. This study addresses this gap by examining how framed messages in AIGC recommendations influence consumer trust across hedonic and utilitarian products using Event-Related Potentials (ERPs). A 2 (message framing: promotion vs. prevention) x 2 (product type: hedonic vs. utilitarian) design was employed. Behaviorally, promotional messages elicited higher trust ratings and faster reaction times than preventive messages across product types. At the neural level, significant interactions between message framing and product type were observed for P3, N400, and LPP. Follow-up analyses showed that, for hedonic products, promotion-framed recommendations elicited larger P3 amplitudes, suggesting greater attentional engagement, and smaller N400 amplitudes, consistent with reduced semantic processing difficulty. The LPP showed broader interaction without reliable simple-framing differences in either product category. Although ERP signals are not direct evidence of trust, they offer insight into the underlying processes shaping AIGC evaluation. This research contributes to the growing literature on consumer neuroscience, trust, and AI-mediated persuasive communication.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-63139-1. Introduction As artificial intelligence (AI) continues to revolutionize various sectors1–3, the rise of AI-generated content (AIGC) marks a significant shift in marketing practices4,5. AIGC refers to content, such as text, images, and videos, that is created by AI systems using generative models like GPT or diffusion models6,7. Increasingly, companies are utilizing these advanced technologies to develop personalized, engaging, and cost-effective marketing strategies8. AIGC allows marketers to deliver messages that are highly customized and responsive to individual preferences9,10, demonstrating its considerable ability to boost consumer engagement, trust, and purchase intentions through emotionally impactful communication11,12. Regulatory Focus Theory (RFT) provides a framework for understanding how individuals pursue goals, distinguishing between promotion-focused and prevention-focused personality types. Promotion-focused individuals are driven by achievement and growth, setting goals to accomplish tasks (e.g., running to train for a marathon), while prevention-focused individuals are motivated by safety and responsibility, and by the avoidance of negative outcomes, such as running to prevent illness13. Although RFT proposes a framework widely used in empirical research in the area of social sciences14,15, Liberman et al.16 argue that people can adopt either motivational orientation depending on the context, self-regulating their behavior to match situational goals. The theory also suggests that communication becomes more persuasive when the message’s regulatory focus matches the recipient’s motivational orientation. Essentially, it is about framing messages to highlight either gains and aspirations, consistent with promotion focus, or risk reduction and protection, consistent with prevention focus14. Once linked to specific product categories, this matching becomes particularly relevant. Micu and Chowdhury17 found that promotional messages are more effective for hedonic products, while preventive messages are more persuasive for utilitarian products. Hedonic products naturally align with promotional goals because they emphasize pleasure, enjoyment, and experiential value, whereas utilitarian items are more closely associated with preventive goals because they emphasize practicality, functionality, and problem solving18. Thus, matching message cues with product characteristics enhances the persuasiveness of marketing communication. In e-commerce, several factors related to product advertising can enhance marketing persuasiveness. Wang and Chou19 show that, in addition to design elements like images, shapes, and colors, message content also plays a key role in product presentation, attracting consumers and influencing purchase decisions. Persuasive communication relies on different appeals to craft a message, and the alignment of message characteristics with product traits can shift consumer attitudes toward the brand or product20. Trust is a multidimensional construct that includes both cognitive and emotional elements. In consumer contexts, trust reflects a person’s willingness to rely on another, based on the expectation that the other will carry out an important action that matters to them21. Once trust is established, it strengthens the relationship between the brand and consumer, influencing behaviors such as intentions to share in social commerce22 and purchasing decisions23,24. While trust is typically associated with human interactions, it also plays a crucial role in engagements with artificial intelligence25. Prior research has already emphasized the importance of trust in online environments26–28. In AI settings, trust often centers on perceptions of competence, transparency, and alignment with user goals29. In recommendation contexts, perceived AI competence is unlikely to be uniform across product types. Prior work indicates that consumers tend to be more receptive to AI recommendations for utilitarian than for hedonic products, partly because AI is viewed as better suited to calculation-based and function-oriented judgments30,31. This source-related cue is relevant to AIGC recommendations, where persuasive messages are not evaluated merely as isolated product claims. Rather, consumers may simultaneously process the content of the recommendation, the type of product being recommended, and the perceived nature of the AI source. Related evidence further shows that making AI systems appear more human-like can reduce users’ psychological distance from the agent, thereby enhancing trust and compliance with their recommendations32. Recent studies further suggest that communication effects in AI-mediated contexts depend on the fit between the AI agent’s message style, the usage context, and consumers’ motivational orientation. Research on AI voice assistants shows that abstract conversational styles generate more favorable consumer attitudes in hedonic contexts, whereas concrete conversational styles generate more favorable attitudes in utilitarian contexts. Li et al.33 further show that regulatory focus shapes these responses: consumers with a promotion focus evaluated abstract conversational styles more positively, whereas those with a prevention focus preferred concrete styles. Lan et al.34 report a related pattern in AI voice assistant interactions. Across hedonic- and utilitarian-dominant service contexts, users showed stronger continuous usage intention when the assistant’s language style matched the context: abstract language in hedonic contexts and concrete language in utilitarian contexts. Their findings also identify processing fluency as the mechanism underlying this congruity effect. In chatbot recommendation contexts, consumers generally report lower purchase likelihood for chatbot-recommended products than for human-recommended products; however, promotion focus can mitigate this negative effect, partly by increasing engagement35. Taken together, these studies suggest that regulatory focus, message-style congruence, and processing fluency are relevant not only to traditional marketing communication, but also to AI-mediated recommendation and interaction settings. These studies are relevant to the present research because they show that message effects in AI-mediated contexts depend on the fit between message cues and the evaluative context. Building on this perspective, the present study examines how promotion- and prevention-framed AIGC recommendations are evaluated when paired with hedonic versus utilitarian products. Because all recommendations in this study were presented as AI-generated, AIGC is treated as the communication context rather than as an experimentally manipulated comparison between AI-generated and human-generated recommendations. To address this theoretical and empirical gap, this study combines behavioral trust ratings with Event-Related Potentials (ERPs), a subset of Electroencephalography (EEG), to examine consumer responses to framed AIGC recommendations. This neuroscience technique offers valuable insights for crafting effective message strategies36. It is extensively utilized in marketing research37,38 and is applied in studies related to attention, decision-making, risk assessment, and emotional processing, using components such as P300, N400, and LPP39–44. In contrast to traditional self-report methods, which are susceptible to bias45,46, ERP records real-time cognitive responses, providing detailed information on brain activity with millisecond precision. The P300 component, peaking around 300 milliseconds (ms) post-stimulus, plays a key role in studying attention allocation, stimulus relevance, and decision-making, with amplitude increasing when stimuli are perceived as salient or meaningful47,48. Additionally, existing research suggests that the level of trust in an information source influences the amplitude of the P3 component during late-stage cognitive processing49, i.e., more reliable sources lead to a greater positive deflection in the ERP waveform. Boudreau et al.50 found that participants trusted a stranger’s recommendation 90% of the time when shared interests were present, with a larger P3 amplitude in this condition, suggesting that common interests enhance neural processing of the communicator’s efforts. Therefore, we hypothesize that the P300 amplitude will be greater when the message frame matches the product type, reflecting enhanced attentional engagement with framed AIGC messages. ERP can also identify neural patterns related to word and phrase processing, with the N400 (N4) component peaking around 400 ms, linked to semantic understanding51, lexical congruency53, and cognitive conflict53. Toffolo et al.54 found that semantic violations, such as incongruent sentence endings, triggered a larger N4, supporting its role in semantic processing. When marketing messages are paired with specific product types, like promotion framing with hedonic products or prevention framing with utilitarian ones17, they are typically perceived as congruent. Thus, we hypothesize that message-product mismatches (e.g., a prevention frame for a hedonic product) will elicit larger N4 amplitudes, reflecting increased semantic conflict and reduced processing fluency. This expectation is consistent with recent AI voice assistant research showing that communication styles incongruent with hedonic versus utilitarian contexts reduce processing fluency and downstream usage intentions34. Several scholars have explored how marketing stimuli trigger consumers’ emotions, whether through brands55, consumer satisfaction56, or advertising campaigns58. Marketing serves as a bridge between companies and consumers, enhancing sensory stimulation through advertisements. The brain plays a critical role in identifying emotions and processing vast amounts of information throughout daily life59. In everyday life, decisions are often influenced by both rational and emotional factors. Another key ERP component, the Late Positive Potential (LPP), observed between 600 and 800 ms, is linked to emotional arousal and sustained attention to motivationally relevant stimuli59. For instance, Jin et al.60 found that positive attribute framing in online shopping led to larger LPP amplitudes compared to negative framing, linking the LPP to positive product attitudes. LPP amplitudes tend to increase when stimuli are emotionally engaging or carry affective weight61–63. Since hedonic products often evoke affective responses64,65, we hypothesize that promotion-framed messages for hedonic products will elicit greater LPP amplitudes, reflecting deeper emotional engagement. Taken together, the ERP components examined in this study are interpreted as neural markers of attention and emotional engagement66, decision-making49, and semantic processing fluency50, which are psychological processes theoretically relevant to trust formation49,67,68. We do not treat these neural signals as direct measures of trust but rather as indirect indicators of mechanisms that may shape consumers’ trust judgments in AIGC. In summary, this study investigates how message framing (promotion vs. prevention) and product type (hedonic vs. utilitarian) interact to influence behavioral trust ratings and ERP responses during AIGC recommendation processing. By integrating regulatory focus theory, neurophysiological measures, and the emerging role of AI-generated content in marketing, this work contributes novel insights into how persuasive communication strategies can be optimized in human-AI interaction. Methodology Pre-test Before the formal laboratory experiment, a preliminary test was conducted to validate the classification of products as hedonic or utilitarian69. To ensure familiarity, twenty products were selected, targeting university students. An online Qualtrics survey with 140 participants (67 males) asked respondents to rate each product on a 7-point Likert scale (1 = highly utilitarian, 7 = highly hedonic). Hedonic products included a game console, perfume, a mahjong set, a human-shaped pot, a bottle of wine, fragrance diffusers, a ceramic skunk, headphones, an eBook reader, and a smartwatch. Utilitarian products included a USB flash drive, food storage container, rice cooker, portable power bank, water filter jug, WiFi router, coffee machine, laptop, toolbox set, and washing machine. A paired sample t-test showed a significant difference between the two categories, with hedonic products scoring higher (M = 6.04, SD = 0.49) than utilitarian products (M = 1.39, SD = 0.38), t (139) = 86.28, p < .001. These results confirmed that participants perceived the products according to their intended categories, validating the list for use in the laboratory phase. Participants Participants in the laboratory phase were twenty-eight right-handed students from Zhejiang University, aged 18 to 34 years (nineteen males, average age 21.91, SD = 2.63). All were native Chinese speakers with normal or corrected-to-normal vision, without any neurological or mental health history. They each received 40 RMB (Chinese yuan) for participating. The final sample size of 28 participants is comparable to analogous ERP studies on message framing that used similar sample sizes, including Zubair et al.70 (N = 20) and Wei et al.71 (N = 22). In addition, to provide a more transparent assessment of the final sample, we conducted a sensitivity analysis using G*Power 3.1.9.672. The analysis was specified for a repeated-measures ANOVA with one group and four repeated measurements, using α = 0.05, power = 0.80, an assumed correlation among repeated measures of 0.50, and a nonsphericity correction of ε = 1. Under these parameters, the sensitivity analysis returned an effect size of f = 0.226, indicating that the final sample was sensitive to detecting within-subject ERP effects in the small-to-medium range. The study received approval from the Internal Review Board (IRB) at Zhejiang University’s Neuromanagement Lab. All experiments were conducted in accordance with relevant ethical guidelines and regulations. Informed consent was obtained from all participants prior to their participation in the study. The study was conducted in China, where the age of majority is 18 years. All participants were adults aged 18 years or older. Stimuli Twenty images were created using AIGC for each product approved in the pre-test phase. To minimize subliminal preferences for one picture over the other, all the product pictures had the same design and color palette, mainly focusing on black-and-white items (Fig. 1). To ensure consistency, all the products were aligned in the middle of the screen against a white background. Each image was saved in a JPEG file, weighing 840 × 640 pixels, and they were presented in a random order throughout each block. Fig. 1Sample of one of the products used in the experiment. A game console was presented to participants as a hedonic item, accompanied by the respective product name. Sample of one of the products used in the experiment. A game console was presented to participants as a hedonic item, accompanied by the respective product name. In addition to the product images, forty unique messages framed according to the two conditions used in this study were created (Fig. 2). Twenty of these messages emphasized the promotional benefits of hedonic products (e.g., for a game console, the message was: “Ignite every thrilling moment; bring the excitement home today!“) and utilitarian products (e.g., for a USB flash drive, the message was: “Experience lightning-fast data transfer; enjoy the convenience of easy portability!“). Similarly, twenty messages emphasized the preventive benefits of each product. For example, for hedonic items (e.g., game console, the message was: “Don’t risk missing out on unforgettable moments; ensure you don’t miss the best thrilling experience of this generation!“) while for utilitarian items (e.g., USB flash drive, the message was: “Don’t let your precious data disappear; ensure secure storage with no risk!“). The sentences were displayed in Chinese, centered on the screen against a white background to ensure consistency, and presented in a randomized order during the experiment. Fig. 2An example of an AI-generated recommendation sentence presented to participants highlighting the promotional aspects of a hedonic product. An example of an AI-generated recommendation sentence presented to participants highlighting the promotional aspects of a hedonic product. Procedures This experiment employed a 2 (message framing: promotional vs. preventive) × 2 (product type: hedonic vs. utilitarian) within-subject design. Participants, seated in a dimly lit, soundproof, and electrically shielded room, viewed stimuli on a screen placed one meter away. Participants rated their perceived trust in AIGC recommendations on a 5-point Likert scale, where ‘1’ and ‘2’ indicated lower trust and ‘4’ and ‘5’ indicated higher trust. Using the keypad, they pressed ‘1’ or ‘3’ to navigate through the scale from lower to higher trust and ‘2’ to confirm their choice. After completing 15 practice trials, participants engaged in four experimental blocks of 40 trials each, presenting 160 randomized product images and framed messages. Each trial began with a 400–600 ms fixation cross (Fig. 3), followed by a 400–600 ms blank screen, a target product image for 1500 ms, and another blank screen for 400–600 ms. A framed message then appeared for 1500 ms, followed by a 400–600 ms blank screen and the same product image with the message and a Likert scale for trust ratings, remaining on-screen until participants confirmed their choice. Trials were separated by a 400–600 ms blank screen, and participants rested for two minutes between blocks to reduce fatigue. Stimuli presentation and data recording were managed using E-Prime 3.0 (Psychology Software Tools, Pittsburgh, PA, USA). Fig. 3Experiment paradigm and timeline for a task trail. Participants were first presented with a fixation cross (400–600 ms), followed by an image of the product (1500 ms). After another fixation cross (400–600 ms), a framed AI-generated recommendation sentence was displayed as the target (1500 ms). Finally, participants rated how much they trusted the recommendation provided by the AIGC on a 5-point scale, with unlimited response time. Experiment paradigm and timeline for a task trail. Participants were first presented with a fixation cross (400–600 ms), followed by an image of the product (1500 ms). After another fixation cross (400–600 ms), a framed AI-generated recommendation sentence was displayed as the target (1500 ms). Finally, participants rated how much they trusted the recommendation provided by the AIGC on a 5-point scale, with unlimited response time. ERP recordings and analysis For ERP recording, EEG data were recorded using a 64-electrode cap, band-pass filtered from 0.05 to 100 Hz, and digitized at a sampling rate of 1000 Hertz73. Utilizing the International 10–20 system and a Neuroscan Synamp 2 Amplifier (Scan4.5, Neurosoft Labs, Inc.). The left mastoid was set as the reference and used cephalic grounding. Vertical EOGs were captured above and below the left eye, and horizontal EOGs at each eye’s outer canthus. We kept electrode impedance below 5 kiloohms (kΩ)74. Data affected by amplifier clippings or electromyography bursts were excluded. For ERP analysis, EEGs were segmented into 1000 ms epochs from − 200 ms pre-stimulus to 800 ms post-stimulus, using the fixation period as a baseline. After digitally low-pass filtering below 30 Hz (24 dB/Octave), epochs with baseline-to-peak deflections over ± 100 µV were removed from the average72. This study analyzed the P300, N4, and LPP ERP components at 300–400 ms75, 400–500 ms76, and 600–800 ms77 post-stimulus onset, respectively. Visual inspection and scalp distribution analyses identified these components as most prominent at frontal, frontocentral, and central sites, leading to the selection of the following electrodes: F3, FZ, F4, FC1, FCZ, FC2, C3, CZ, and C4 for the P300; F3, FZ, F4, FC3, FCZ, FC4, C3, CZ, and C4 for the N4; and F1, FZ, F2, FC1, FCZ, FC2, C3, CZ, and C4 for the LPP. Odd-numbered electrodes indicated the left hemisphere, even numbers the right, and ‘z’ denoted midline locations. Statistical analysis employed a repeated-measures ANOVA between message framing, product type, and electrodes, with further simple effect analysis conducted when interaction effects were significant. Data analysis was performed using SPSS 23.0 (IBM, New York, NY, USA). To examine whether the four experimental conditions were separable at the EEG pattern level, supplementary multivariate pattern analyses (MVPA) were conducted on the reprocessed event-related single-trial EEG data following established EEG/ERP time-resolved decoding approaches78,79. These analyses included whole-scalp time-resolved decoding and ROI-based decoding using the predefined electrode clusters and time windows corresponding to the P3, N4, and LPP components. Linear support vector machine classifiers were trained and evaluated using stratified 5-fold cross-validation. Four-class decoding was evaluated against the 25% chance level, whereas pairwise decoding was evaluated against the 50% chance level. Full details of the MVPA procedures and results are reported in the Supplementary Material. Results Behavioral results Reaction time (response times in milliseconds) Participants’ reaction times for AIGC recommendations were analyzed using a 2 (message framing: promotion vs. preventive) × 2 (product: hedonic vs. utilitarian) repeated-measures ANOVA. The results revealed a significant main effect for message framing, F (1, 27) = 9.445, p = .005, ηp2 = 0.259, with promotional messages eliciting faster response times for both product types (Mhedonic = 1475.55, SD = 738.26; Mutilitarian = 1328.84 SD = 724.45) compared to preventive messages (Mhedonic = 1530.90, SD = 748.09; Mutilitarian = 1545.16, SD = 1013.72). No significant main effect for product type or interaction between message framing and product type was observed (p > .05). Perceived trust (5-point Likert scale) To analyze participants’ perceived trust in AI recommendations, a 2 (message framing: promotion vs. preventive) × 2 (product: hedonic vs. utilitarian) repeated-measures ANOVA was conducted. The results revealed a significant main effect for message framing, F (1, 27) = 8.632, p = .007, ηp2 = 0.242, indicating that promotional messages were associated with higher perceived trust for both product types (Mhedonic = 3.35, SD = 0.482; Mutilitarian = 3.50, SD = 0.462) compared to preventive messages (Mhedonic = 3.02, SD = 0.570; Mutilitarian = 3.06, SD = 0.598). However, no significant main effect was observed for product type, nor was there a significant interaction between message framing and product type (p > .05). ERP results P3 A repeated-measures ANOVA was conducted to examine the interactive effect of message framing (promotion vs. prevention) and product type (hedonic vs. utilitarian). The analysis revealed a significant two-way interaction, F (1, 27) = 9.138, p = .005, ηp2 = 0.253. A significant main effect of electrode site was observed, with degrees of freedom corrected using the Greenhouse-Geisser method, F (3.032, 81.859) = 13.118, p < .001, ηp2 = 0.327. However, no significant main effects for message framing or product type were found, p > .05. A simple effect analysis revealed a significant difference between the promotion and prevention frames for the hedonic product condition (Mpromotion = 0.653, SD = 3.80 vs. Mprevention = − 0.768, SD = 3.98), F (1, 27) = 10.027, p = .004, ηp2 = 0.271 (Fig. 4). No significant differences were observed between message framing conditions for the utilitarian product (Mpromotion = − 0.283, SD = 3.98 vs. Mprevention = 0.186, SD = 4.36), F (1, 27) = 0.824, p = .372, ηp2 = 0.030. Fig. 4Grand average ERP amplitudes (µV) for the P300 component. The figure shows waveforms recorded at the FZ, FCZ, and CZ electrodes within the 300–400 ms time window. Grand average ERP amplitudes (µV) for the P300 component. The figure shows waveforms recorded at the FZ, FCZ, and CZ electrodes within the 300–400 ms time window. Fig. 5Scalp topographies at the decision-making phase (300 ms post-stimulus) for each experimental condition (Hedonic Promotional, Hedonic Preventive, Utilitarian Promotional, Utilitarian Preventive), corresponding to the P3 component. Maps reflect ERP amplitudes (µV) averaged across all electrode sites. Scalp topographies at the decision-making phase (300 ms post-stimulus) for each experimental condition (Hedonic Promotional, Hedonic Preventive, Utilitarian Promotional, Utilitarian Preventive), corresponding to the P3 component. Maps reflect ERP amplitudes (µV) averaged across all electrode sites. N4 A repeated-measures ANOVA revealed a significant two-way interaction between message framing and product type for the N4 component, F (1, 27) = 5.823, p = .023, ηp2 = 0.177. With degrees of freedom adjusted using the Greenhouse-Geisser correction, significant main effects of electrode site were observed, F (2.492, 67.290) = 12.334, p < .001, ηp2 = 0.314, and for the interaction between product type and electrodes, F (2.677, 72.285) = 6.274, p < .001, ηp2 = 0.189. However, no significant main effects for message framing or product type were found, p > .05. Simple effect analysis revealed that, for hedonic products, promotion messages elicited significantly smaller N4 amplitudes compared to prevention messages (Mpromotion = − 0.727, SD = 3.02 vs. Mprevention = -2.142, SD = 3.85), F (1, 27) = 9.035, p = .006, ηp2 = 0.251 (Fig. 6). However, no significant differences were found for the utilitarian products (Mpromotion = -1.479, SD = 4.13 vs. Mprevention = -1.133, SD = 3.86), F (1, 27) = 0.346, p = .561, ηp2 = 0.013. Fig. 6Grand average ERP amplitudes (µV) for the N4 component. The figure illustrates waveforms recorded at the FZ, FCZ, and CZ electrodes during the 400–500 ms time window. Grand average ERP amplitudes (µV) for the N4 component. The figure illustrates waveforms recorded at the FZ, FCZ, and CZ electrodes during the 400–500 ms time window. Fig. 7Scalp topographies at 400 ms post-stimulus onset for each experimental condition (Hedonic Promotional, Hedonic Preventive, Utilitarian Promotional, Utilitarian Preventive), corresponding to the N4 component. Maps show the scalp distribution of ERP amplitudes (µV) across electrode sites. Scalp topographies at 400 ms post-stimulus onset for each experimental condition (Hedonic Promotional, Hedonic Preventive, Utilitarian Promotional, Utilitarian Preventive), corresponding to the N4 component. Maps show the scalp distribution of ERP amplitudes (µV) across electrode sites. LPP The repeated-measures ANOVA revealed a significant two-way interaction between message framing and product type for the LPP component, F (1, 27) = 7.736, p = .010, ηp2 = 0.223. The main effect of product type did not reach conventional significance, F (1, 27) = 3.452, p = .074, ηp2 = 0.113. With degrees of freedom corrected using the Greenhouse-Geisser method, a significant main effect of electrode site was found, F (3.329, 89.885) = 10.791, p < .001, ηp2 = 0.286. No significant main effect was found for message framing (p > .05). Simple-effect analysis showed that the effect of message framing in the hedonic product condition did not reach conventional significance, although the means showed a trend toward larger LPP amplitudes for promotion-framed than prevention-framed recommendations (Mpromotion = − 0.063, SD = 3.91 vs. Mprevention = -1.229, SD = 4.16), F (1, 27) = 3.108, p = .089, ηp2 = 0.103 (Fig. 8). No significant effect was found for message framing in the utilitarian product category (Mpromotion = 0.144, SD = 4.21 vs. Mprevention = 0.993, SD = 4.52), F (1, 27) = 1.599, p = .217, ηp2 = 0.056. Fig. 8Grand average ERP amplitudes (µV) for the LPP component. The figure displays waveforms recorded at the FZ, FCZ, and CZ electrodes during the 600–800 ms time window. Grand average ERP amplitudes (µV) for the LPP component. The figure displays waveforms recorded at the FZ, FCZ, and CZ electrodes during the 600–800 ms time window. Supplementary multivariate pattern analysis (MVPA) To examine whether the four experimental conditions were separable at the EEG pattern level, supplementary multivariate pattern analyses (MVPA) were conducted using the reprocessed single-trial EEG data. Whole-scalp time-resolved MVPA did not reveal reliable condition separability. Four-class decoding remained close to the 25% chance level across the P3, N4, and LPP time windows (P3 = 0.250, p = .484; N4 = 0.252, p = .349; LPP = 0.251, p = .404), and none of the pairwise contrasts showed significant above-chance classification. ROI-based MVPA aligned with the ERP component windows also produced non-significant results. Four-class decoding remained near chance in all ROIs (P3 = 0.252, p = .409; N4 = 0.251, p = .473; LPP = 0.246, p = .689), and the theoretically critical hedonic/promotion versus hedonic/prevention contrast did not differ significantly from the 50% chance level in the P3, N4, or LPP ROIs (P3 = 0.498, p = .568; N4 = 0.489, p = .818; LPP = 0.509, p = .268). Additional MVPA robustness checks using a high-trial subset and pseudo-trial averaging yielded the same null pattern. Thus, while the univariate ERP analyses revealed component-specific amplitude effects, the supplementary MVPA did not provide reliable evidence of condition-level separability in distributed EEG patterns. Full details are reported in the Supplementary Material. Discussion Although EEG/ERP research directly linking trust and neural activity is still emerging, this study extends framing theory by demonstrating its application in AIGC. This study makes a unique contribution to the theoretical understanding of Regulatory Focus Theory by providing neural evidence on how message framing interacts with product type in AI-mediated consumer decision-making. The present analysis revealed a significant interactive effect of message framing and product type on three ERP components: P3, N4, and LPP. This study avoids making strong claims that ERP components are direct neural markers of trust. Instead, we acknowledge the risk of reverse inference, which involves attributing a specific psychological state, such as trust, to a neural response like the P3 or N4 without sufficient supporting evidence. Trust is a multidimensional construct that includes attention, perception, cognition, and decision-making processes49,67,68,80,81. ERP components such as P3, N4, and LPP are broadly responsive to attention, motivational salience, emotional engagement, and cognitive fluency. Although these functions are relevant to how trust is formed, we interpret them as indirect indicators that may support but do not confirm trust-related evaluations. Our goal is not to equate neural signals with trust itself, but to complement self-reported trust ratings with neurophysiological insights into the underlying cognitive and emotional mechanisms. Importantly, unlike previous EEG studies on trust that employed interactive paradigms involving interpersonal trust decisions or evaluation of messages from human agents49,82, the present research adopted a passive viewing design. Participants were exposed to AI-generated messages and subsequently rated their perceived trust on a Likert scale. This setup allowed for the examination of spontaneous neural responses without requiring interpersonal interaction. This approach is consistent with Moravec et al.83, who recorded EEG data as participants evaluated the truthfulness of social media headlines on a Likert scale in a non-interactive, rating-based task, thus assessing perceived credibility in a passive viewing context. Behaviorally, promotional messages were associated with faster reaction times, suggesting that participants processed these messages more fluently or found them easier to evaluate in the task84. Promotional messages also led to higher perceived trust ratings across both product types. Although no significant interaction was found between message framing and product type in the behavioral data, the ERP findings provided additional insight into how the combination of message framing and product type modulated neural responses during recommendation processing. Specifically, the P3 and N4 results point to differences in attentional engagement and semantic congruence, whereas the LPP interaction suggests that later-stage sustained processing was also sensitive to the overall configuration of message framing and product type. These findings illustrate how neural responses can complement self-reported data by revealing attention and emotion-related processes that are relevant for understanding how consumers evaluate AI-generated recommendations. The P3 component, linked to selective attention85,87, and task relevance87,88, highlights the role of framed messages in capturing attention through hedonic and utilitarian benefits. While prior behavioral marketing research by Micu and Chowdhury17 observed a matching effect between promotional messages for hedonic products and preventive messages for utilitarian ones in human recommendations, this alignment may be less pronounced or specific in the context of AI-generated content. In an ERP analysis, Chen et al.89 found that utilitarian products elicited higher P3 amplitudes, emphasizing the influence of product type on neural responses. In contrast, our investigation demonstrates that hedonic products also triggered increased P3 activity when paired with promotional messages, indicating heightened attention. This suggests that while product type alone can influence the P3 component, combining framing and product type (particularly with hedonic items) can also significantly enhance attention during decision-making. Our findings confirm that promotional cues significantly enhance attention for hedonic products, consistent with promotion framing’s focus on positive outcomes. This matches the pleasurable nature of hedonic items, as reflected by the heightened P3 activity. At a broader level, Peng et al.90 found larger P3 amplitudes in promotion-focused consumers under low rather than high information load, a pattern they interpreted in terms of regulatory fit between consumers’ regulatory focus and information load. From a consumer-centered perspective, a promotion frame may be associated with greater attentional engagement and motivational relevance during the evaluation of hedonic products, because it emphasizes pleasure and satisfaction, features that align with the sensory enjoyment these products provide. However, no significant framing effects were observed for utilitarian products. This result may align with prior research suggesting that AI agents are inherently effective at recommending utilitarian items due to their functionality-driven nature. This may render them less influenced by the framing strategies employed in this study30,31. The P3 component in our investigation provides indirect evidence of attentional processes that may be relevant to trust-related evaluation. Long et al.91 linked P3 sensitivity to trust levels, showing larger amplitudes for trust choices and reduced activity for distrust. In our study, larger P3 amplitudes over frontal and frontocentral electrode sites were associated with message-framing effects (Fig. 5), suggesting that promotional cues in AIGC recommendations may increase attentional engagement during evaluation. EEG research has also shown that trust situations often elicit increased frontal activation92. Moreover, prior studies suggest that the frontal and frontocentral regions are sensitive to trust-related signals, particularly those based on promises93 while central regions have been linked to trust in human-machine interactions94. Hence, the P3 results suggest that promotion-framed recommendations for hedonic products elicited larger P3 amplitudes, consistent with greater attentional engagement and motivational relevance during recommendation evaluation. Given that the P3 component is associated with selective attention and motivational salience, processes that may contribute indirectly to trust judgments, these findings should be interpreted as suggestive rather than conclusive. The passive nature of the paradigm also precludes strong causal inferences about trust-specific neural mechanisms. The N4 component, a key ERP marker of stimulus congruence, has been widely used to analyze neural responses to match or mismatch across various stimuli, including images, videos, gestures, and sentences95–98. In message framing studies, smaller N4 responses have been interpreted as indicating reduced dissonance and lower semantic processing demands99, whereas mismatches, such as incongruent brand logos, have been shown to elicit larger N4 responses100. More recently, Liu et al.101 found that, within AI-generated content, opinion-based content elicited larger N4 amplitudes than factual content, which they interpreted as reflecting semantic prediction errors and conflict detection within the N4 time window. Although their study examined content type in an AIGC credibility-assessment context rather than message framing and product type, it supports the relevance of the N4 component for examining semantic conflict and processing demands in AI-mediated evaluations. Regarding the broader concept of cognitive conflict, Xie et al.102 provided ERP evidence that AI (vs. human) agents recommending emotionally driven products elicited heightened N2 amplitudes, indicating increased neural conflict. Although their study focused on the N2 component, the findings further suggest the idea that mismatches in AI-driven contexts can increase neural processing demands. Consistently, this study found that framing mismatches for hedonic products, particularly preventive cues, significantly increased N4 amplitudes, indicating heightened cognitive demand as consumers reconciled the dissonance between the pleasure-driven nature of the product and the risk-averse tone of the preventive messages. While not a direct marker of trust, higher N4 amplitudes in response to incongruent pairings may reflect cognitive conflict or reduced processing fluency51,54. This neural response may be indirectly relevant to trust. Still, prior behavioral findings suggest that fluently processed stimuli, such as semantically congruent names or messages, are more likely to be judged as trustworthy67. The neural evidence from the N4 highlights the importance of congruence in message framing, suggesting that mismatched frames may disrupt fluency in AI-generated recommendations, especially for emotionally driven products like hedonic items. When applied congruently, our findings show that promotional framing can enhance the appeal of product types that have previously been associated with consumers’ cognitive conflict when recommended by AI agents102. Conversely, no significant differences in N4 amplitudes were observed for utilitarian products. This may reflect their goal-oriented evaluation based on practicality and functionality, where framing mismatches impose less cognitive demand than the emotional and aspirational reliance on hedonic products. The LPP component, which marks the third stage of cognitive processing, reflects sustained attention and emotional engagement. In the present study, the significant interaction between message framing and product type for the LPP component suggests that later-stage processing was sensitive to how framed recommendations were paired with different product types. This interpretation is consistent with broader ERP research showing that emotionally or motivationally significant stimuli can modulate LPP responses103,104. More specifically, within message-framing investigations, the LPP has also been linked to positive and negative framing cues72 and emotions in green consumption71. Branded products with higher emotional value have been shown to elicit larger LPP responses than branded products with lower emotional value under specific social conditions105, however, the present findings provide a more cautious pattern. Although the interaction between message framing and product type was significant, the simple-effect analyses did not reveal a significant framing difference within either product category. The hedonic product condition showed only a marginal trend toward larger LPP amplitudes for promotion-framed than prevention-framed recommendations, whereas no significant framing effect was observed for utilitarian products. Therefore, the LPP results suggest that late-stage sustained processing was modulated by the broader configuration of message framing and product type, but they do not provide strong evidence for a specific promotion-framing advantage within hedonic products. One possible explanation for the absence of clear simple effects is that late-stage evaluative processing may have been influenced by multiple sources of motivational relevance rather than by message framing alone. For utilitarian products, these sources may include the perceived fit between AI-generated recommendations and function-oriented product evaluation102. This interpretation is consistent with Longoni and Cian’s “word-of-machine” effect30, which suggests that consumers perceive AI recommenders as more competent in recommending utilitarian rather than hedonic products. However, because the present study did not directly measure perceived AI-product fit or emotional responses, this explanation remains tentative and should be examined in future research. The supplementary MVPA analyses add a complementary perspective to the ERP results. Across whole-scalp time-resolved MVPA, ROI-based MVPA, high-trial ROI-based MVPA, and pseudo-trial MVPA, decoding accuracies did not reliably exceed chance. Thus, the MVPA results did not show robust condition-level separability in distributed EEG patterns. Importantly, this finding limits claims about multivariate separability, but does not negate the univariate ERP results, which showed component-specific differences in mean amplitudes within predefined time windows and ROIs. Accordingly, the ERP effects should be interpreted as amplitude-based component effects, rather than as evidence that the four conditions are robustly classifiable from distributed EEG patterns. This methodological distinction is consistent with Grootswagers et al.78, who noted that decoding results in time-series neuroimaging data can be affected by preprocessing choices, trial averaging, classifier selection, and cross-validation design. To the best of our knowledge, few EEG/ERP studies have explored how message framing (promotion vs. prevention) interacts with product type (hedonic vs. utilitarian) in the context of AI-generated recommendations. While previous studies have examined these factors separately17,30,31,102, none have combined them in a neurophysiological design using ERP markers. This intersection of cognitive neuroscience, message framing theory, and AI communication brings a novel theoretical contribution that advances our understanding of how consumers process persuasive AIGC content. Building on this foundation, this study contributes to the theoretical understanding of message framing by connecting Regulatory Focus Theory with neural evidence from EEG in the context of AI-generated content. While most framing research focused on human communication, our findings suggest that, in AI-based recommendation contexts, promotional cues for hedonic products are associated with greater attentional engagement and reduced semantic processing difficulty, as reflected in the P3 and N4 results. More specifically, the P3 findings point to enhanced attentional engagement, whereas the N4 findings suggest reduced semantic processing demands when promotional frames are paired with hedonic products. Together, these findings indicate that the fit between message type and product nature may shape how consumers process persuasive AIGC content at specific stages of neural evaluation. By using a passive viewing design, this study also adds to the literature on consumer neuroscience and AI-mediated evaluation. Neural components such as P3 and N4 were sensitive to evaluative processing even in the absence of interactive tasks. This complements earlier research suggesting that consumers often perceive AI recommenders as more competent in utilitarian than hedonic contexts30, and suggests that promotional framing may also support AI-based recommendations in hedonic settings when message–product congruence is achieved. Lastly, increased N4 responses in mismatched message–product conditions point to greater semantic processing difficulty and reduced fluency. While N4 is not a direct neural marker of trust, this effect offers a useful lens for interpreting how incongruence might influence consumer evaluations. Practically, this study offers clear guidance for marketers and platforms leveraging AI-based recommendation systems. First, message framing should align with the product type. Promotional messages that highlight enjoyment or aspiration may be particularly suitable for hedonic products by improving message-product congruence. For utilitarian items, framing may be less critical, as consumers already perceive AI recommenders as relatively competent in that domain30. Second, promotional messages for hedonic products could emphasize vivid, enjoyment-related benefits. Our ERP findings, particularly the larger P3 amplitudes, suggest that such framing is associated with greater attentional engagement, while the behavioral results suggest that it may also support more favorable trust evaluations in AIGC recommendations. Third, our results support the idea that strategic message–product alignment within AI-mediated recommendation matters. Although AI recommenders are typically preferred for functional or search-based tasks30,102, thoughtful framing may improve their effectiveness in hedonic categories as well. For example, an AI assistant recommending a luxury fragrance may be more persuasive when using promotional language such as “Feel confident and unforgettable” rather than a preventive language like “Avoid unpleasant odors”. Finally, future AIGC applications could incorporate adaptive framing by adjusting message strategies based on product type or user preferences. Doing so may help companies improve engagement, support trust-related evaluations, and make AI communication more effective. While this study contributes to expanding knowledge on message-framing techniques in AI-generated recommendations, it also highlights several avenues for further investigation. One potential direction involves exploring the interplay between human and AI recommenders in the context of message framing, as prior research suggests that while AI excels in recommending utilitarian products, human recommenders may be more effective for hedonic items30. Investigating whether the type of recommender influences the effectiveness of framing strategies for hedonic and utilitarian products could yield valuable insights. This study’s sample was limited to Chinese participants, raising questions about the generalizability of its findings across diverse demographic and cultural groups. Future research should conduct cross-cultural studies to verify the universality of these results. Although the sensitivity analysis supported the use of the final sample for the planned within-subject ERP amplitude analyses, future studies designed specifically for multivariate or time-resolved decoding could benefit from larger samples and a greater number of balanced, artifact-free trials per condition. Such approaches may complement conventional ERP analyses by clarifying whether distributed EEG patterns provide additional information about AI-mediated recommendation evaluation and trust-related responses. Another area of investigation involves the neural responses to framing cues for utilitarian products, as this study found no significant differences in any of the three components, suggesting avenues for future exploration. Researchers may also consider other factors, including brand trust, perceived product quality, message credibility, emotionally resonant items, or conditions where prevention framing may be more effective (such as with risk-sensitive consumers or high-risk decisions) to determine if these factors elicit distinct neural responses and clarify the role of framing in trust formation. Finally, given the passive nature of the task, future studies could adopt interactive or risk-based paradigms, such as trust games or choice-based evaluations, to further validate the role of ERP markers in the trust-related evaluation. Additionally, future research could explore within-subject correlations between ERP amplitudes and individual trust ratings across trials or conditions, helping to clarify the direct neurobehavioral relationship between cognitive-evaluative processing and perceived trust. While our findings suggest that P3, N4, and LPP are associated with attentional engagement, semantic processing difficulty, and later-stage sustained processing, respectively, it remains essential to test whether these neural responses directly translate into behavioral trust under more ecologically valid conditions. Combining passive and interactive designs may provide a more comprehensive understanding of how consumers develop trust in AI-generated content. Conclusion In conclusion, using the neuroscience approach of ERP, this study investigated consumers’ decision-making processes and psychological mechanisms regarding AIGC recommendations under message-framing strategies paired with hedonic and utilitarian products. The findings suggest that the P3 component indicated attention and message relevance during the decision-making stage, particularly for promotion-framed recommendations paired with hedonic products. The N4 findings suggest greater semantic processing difficulty associated with message–product incongruence, particularly in response to prevention-framed recommendations for hedonic products. The LPP component showed a broader interaction between message framing and product type, suggesting that later-stage sustained processing was sensitive to the overall configuration of AIGC recommendations, although no reliable simple framing effect emerged within either product category. Overall, the results highlight that aligning promotional cues with the nature of the product type may enhance attention and processing fluency, especially for hedonic products. While these neural responses may reflect cognitive and affective processes relevant to trust formation, we interpret such associations with caution due to the employed research paradigm. Without interactive trust decisions, strong claims about trust-specific mechanisms remain limited. Moreover, the supplementary MVPA analyses did not provide reliable evidence that the experimental conditions were separable based on distributed EEG patterns. Thus, the ERP findings should be interpreted as component-specific amplitude effects rather than as evidence of robust condition-level separability in multivariate EEG patterns. Nonetheless, our findings underscore the importance of aligning message framing with product type to enhance engagement and perceived relevance, which may support more persuasive AI-mediated communication. Future studies using dynamic or incentive-based paradigms could directly examine trust-related neural mechanisms, building on these initial insights. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1