Perceived Legitimacy Matters: Building Public Trust and Acceptance of AI-Generated News Images Through Strategic AI Disclosure

The Transparency Dilemma and Disclosure Strategies of AI Use

While generative AI has emerged as a useful tool for visual storytelling in time-sensitive and resource-constrained contexts, it raises concerns about misinformation and journalistic ethics (Dan et al., 2021; von Sikorski & Hameleers, 2025). Transparency is often seen as a remedy to navigate the tension between misinformation and responsible AI-assisted journalism (Diakopoulos & Koliska, 2017). Despite a lack of formal guidelines, long-standing journalistic norms emphasize the need for transparent and responsible AI use in newsrooms (Karlsson, 2010).

However, transparency alone does not necessarily resolve this tension. In some cases, disclosing AI use can even backfire. For instance, Toff and Simon (2024) found that audiences consistently rated AI-labeled news articles as less credible. Similarly, Wittenberg et al. (2025) found that labeling AI-generated images significantly reduced the perceived credibility of presented content, regardless of what types of labels (i.e., explaining how content was made or highlighting content’s potential to mislead) were used. In advertising contexts, Baek et al. (2026) found that disclosing AI-generated images significantly reduced credibility perceptions and led to unfavorable attitudes toward advertisements. These findings point to a disclosure dilemma: although transparency may fulfill ethical standards, it can also inadvertently erode public trust.

In response, scholars call for more strategic disclosure strategies that move beyond treating “disclosure” as a binary (i.e., either with a disclaimer of “generated by AI” or absent) and consider how the framing of disclosure shapes audience perceptions (K. de Fine Licht & J. de Fine Licht, 2020). Rather than advocating for a “full transparency” approach, K. de Fine Licht and J. de Fine Licht (2020) propose that providing justifications for the decision to use AI (i.e., transparency in rationale) can reduce public backlash. That is, disclosure should not merely state that AI was used but also provide context and justification for its use. Strategic disclosure, particularly when it includes clear rationales, can enhance public understanding, mitigate skepticism, and potentially build trust (J. de Fine Licht et al., 2014).

Building on this perspective, we focus on one specific journalistic application: the use of AI-generated news images in disaster reporting. As a major component of news production, visuals help convey the immediacy and urgency of disasters and activate public attention (Zelizer, 2010). However, news organizations show concerns about using generative visual AI, such as potential mis/disinformation and reputational risk (Thomson et al., 2025). To address this, our study seeks to examine whether attributed motives for using AI-generated news images influence audience responses, offering insights into how newsrooms can better communicate AI use and mitigate potential backlash.

Drawing on attribution theory (Weiner, 2006), we conceptualize strategic AI disclosure in terms of internal versus external attributions. Attribution theory posits that individuals have a basic psychological need to understand the causes of various issues and events as a way to adjust to and control their environment (Heider, 1958). People tend to attribute events to either internal factors (e.g., egoistic or self-serving motives) or external causes (e.g., altruistic motives) (Heider, 1958; Jones & Davis, 1965), among which factors such as assumed intentionality, perceived locus of control, and social desirability of the behavior influence the types of attributions made (Ramasubramanian, 2011; Weiner, 1986). In journalistic contexts, news stories often use framing strategies that imply causes and consequences to help their audiences make sense of various social issues and shape their subsequent reactions and attitudes (Coombs, 2007; Iyengar, 1990; Ramasubramanian, 2011). For example, Lu (2024) found that news coverage of the opioid crisis using external attributions increased public support for perpetrator-oriented punitive policies.

Although prior research focused mainly on how news frames shape responsibility attribution, AI disclosure is essentially a similar message that utilizes framing devices to imply motivations and causes of the social entity. Different framing of AI disclosure messages can provide subtle cues that facilitate certain causal interpretations and responsibility attributions (Entman, 1993; Ji et al., 2025). Disclosures that emphasize self-serving motives foreground internal causes, whereas disclosures that stress altruistic motives and social benefits highlight external causes. Prior research suggests that when negative outcomes are attributed to external motives, actors are less blamed. When the same outcomes are attributed to internal motives, they are seen as less justified and may receive more blame (Jeong, 2009; Weiner, 2006).

Audiences may apply similar explanatory judgment processes in an attempt to evaluate AI use in the journalistic context (K. de Fine Licht & J. de Fine Licht, 2020). Justifying AI use from external causes, such as protecting victims’ privacy or illustrating inaccessible disaster scenes, is more likely to be interpreted as externally motivated and ethically grounded. Such rationales align with professional standards of journalism and may buffer against the trust penalty of using AI in news (Karlsson, 2010). In contrast, justifying AI use from internal causes, such as reducing the newsroom’s travel costs and accelerating news production, may appear internally motivated and trigger punitive attitudes. Importantly, our design also includes a control condition featuring a real photograph, allowing us to compare legitimacy perceptions of AI-generated images under different disclosure messages against a baseline of traditional photojournalism.

Based on the above, this study investigates how disclosure strategies influence both acceptance of AI in journalism and trust in the news organization. Acceptance of AI in journalism reflects broader audience attitudes and has important implications for predicting support for AI integration in future journalism (Kelly et al., 2023). Understanding how disclosure strategies shape AI acceptance offers important insights into the public’s openness to technological innovation in journalism. Trust in news organizations is also critical, given evidence that disclosing AI use reduced perceived trustworthiness of news content and undermined audience confidence in journalism as an institution (Toff and Simon, 2024). Considering both outcomes allows us to provide a more holistic view of how audiences evaluate AI-generated visuals in journalistic contexts.

Given the above, we hypothesize that framing AI use with an external attribution will yield more positive audience responses, compared to a simple disclaimer (H1) and an internal attribution (H2):

Perceived Legitimacy as the Key Mechanism in AI Disclosure Effects

As news organizations increasingly adopt AI into news production and dissemination, its use in journalism has also resulted in audience backlashes, including reduced public trust, heightened risk perceptions, and lower acceptance of AI in future journalism (Mitova et al., 2025; Morosoli et al., 2025). We argue that perceived legitimacy is a key mechanism linking AI disclosure to these negative outcomes. Legitimacy refers to a generalized perception that an entity’s actions “are desirable, proper, or appropriate within some socially constructed system of norms, values, beliefs, and definitions” (Suchman, 1995). In journalism, perceived legitimacy depends on whether audiences believe newsroom practices align with professional and ethical norms (Tong, 2018). People’s concerns about legitimacy would arise when news organizations take unconventional actions that deviate from established norms (Schilke & Reimann, 2025). For instance, Blassnig et al. (2024) found that when audiences perceive editorial gatekeeping as being replaced by invisible news recommender systems, trust in those outlets declines, as it changes the traditional way media and audiences interact. Similarly, in the context of AI-generated news images, audiences may question the newsroom’s legitimacy if they view such practices as violating journalistic standards of authenticity and ethics.

Within attribution theory, disclosure strategies that imply external or internal attribution motives for AI use can shape perceived legitimacy of such behaviors (Jahn et al., 2020). External attributions that highlight altruistic motives are more likely to be seen as socially responsible, thereby enhancing perceived legitimacy (Jahn et al., 2020). In contrast, internal attributions that signal self-serving motives are likely to weaken perceived legitimacy (Coombs, 2007). Prior organizational research supports this logic. Miotto and Youn (2020) found that altruistic attributions of a company’s sustainable collections increased perceptions of corporate legitimacy, whereas Jahn et al. (2020) showed that reward-driven motives reduced both corporate credibility and organizational legitimacy.

Applied to AI-assisted journalism, we argue that when audiences perceive journalistic AI use as driven primarily by expedience or self-serving goals, they are more likely to form internal attributions and reduce perceived legitimacy. When AI usage is framed as serving public rights or the stakeholder interests, they are more likely to form external attributions, thereby enhancing perceived legitimacy. A simple AI disclaimer (e.g., “This image was generated by AI”) is likely to perform worse than an external attribution because it provides no contextual rationale for AI use. Without an explicit explanation, audiences must infer the newsroom’s motives themselves, often relying on their preexisting dispositions toward AI. Given that prior research has shown that citizens are likely to show resistance to integrating AI into the news cycle (Morosoli et al., 2025), a simple AI disclaimer without any rationales given will result in more negative judgments. The absence of a rationale leaves room for audiences to assume convenience-driven motives and conclude that the newsroom might just choose AI for expedience, which could lower perceived legitimacy.

Taken together, we expect that disclosure strategies framing AI use as externally motivated are more likely to enhance perceived legitimacy than strategies that offer no rationale or highlight internal attributions.

Perceived legitimacy, in turn, should shape public trust in news organizations and the broader use of AI in journalism. According to organizational legitimacy theory (Deephouse & Suchman, 2008), when audiences evaluate an organization as legitimate and perceive it as aligned with social norms and ethical standards, they would support and trust that organization; when they view it as illegitimate, audiences would question its integrity and reliability, leading to distrust or other negative reactions. Mitova et al. (2025) found that citizens’ trust in the journalistic deployment of AI-powered tools remains relatively low. Similarly, Toff and Simon (2024) found that audiences perceived news labeled as AI-generated as less trustworthy, even when articles themselves are not evaluated as any less accurate. This suggests that people’s distrust and increased scrutiny of AI-assisted journalism may not only concern the accuracy of AI-generated content but also other factors, such as the ethical standards of AI-assisted news production (Culver & Lee, 2019). We therefore argue that reduced perceived legitimacy might undermine trust in the news organization.

The same logic applies to the acceptance of AI in journalism, which reflects audiences’ willingness to support integrating AI into news production. Acceptance of AI in journalism depends not only on perceived usefulness but also on whether the technology is viewed as normatively appropriate (Martin & Waldman, 2023). When AI use is seen as legitimate, audiences should be more willing to accept its role in journalism. Therefore, we propose:

In fact, the mediating role of perceived legitimacy has been found in extant studies on public responses toward AI usage in various contexts. For instance, Schilke and Reimann (2025) found that disclosing AI use to the public could reduce public trust across multiple tasks, such as communication, analytics, and artistry, and the trust erosion was largely explained by damaged legitimacy perceptions. Applying this to AI-assisted journalism, disclosure strategies influence legitimacy perceptions, which in turn shape both public trust in the news organization and acceptance of AI in journalism.

How Attitudes Toward AI Moderate the Effects of Disclosure Strategies

While strategic disclosure is designed to foster public trust and enhance acceptance of AI-generated images in journalism by increasing the perceived legitimacy of AI use, audiences may not uniformly accept the rationales offered by news organizations (Morosoli et al., 2025). We argue that the impact of disclosure strategies is likely contingent on individuals’ underlying attitudes toward AI. Attitudes toward AI, ranging from techno-optimism to skepticism or technophobia, can fundamentally shape how audiences interpret disclosure messages. For instance, individuals with more positive attitudes often view AI adoption as innovative and efficient (Bewersdorff et al., 2025; Morosoli et al., 2025). They may therefore respond more favorably when disclosures frame AI use in socially responsible terms (e.g., external attributions) but still become skeptical when AI is presented as self-serving or when disclosure is minimal (e.g., a simple AI disclaimer). In contrast, individuals with negative attitudes toward AI are predisposed to perceive AI involvement as risky, threatening, or deceptive. For them, even socially responsible framings may not fully mitigate suspicion, and self-serving or minimal disclosures may further erode legitimacy.

These diverging expectations highlight that disclosure strategies are unlikely to affect all audiences in the same way. Yet, existing research has not examined whether and how attitudes toward AI interact with disclosure strategies to shape perceived legitimacy, which subsequently influences trust and AI acceptance. To address this gap, we propose the following research question and present our conceptual model in Figure 1:

OPEN IN VIEWER

Figure 1.

The conceptual model of this study.

Overview of the Study

This study was approved by the Institutional Review Boards at the University of Michigan. To examine the proposed hypotheses and research questions, we conducted a 4 (Disclosure Strategy: Simple AI Disclaimer vs. Internal Attribution vs. External Attribution vs. Real Photo) × 2 (Victim Count: Multiple vs. Single) × 2 (Disaster Context: Volcano Eruption vs. Avalanche) between-subjects experiment.

Participants first completed a pretreatment survey measuring their attitudes toward AI. They were then shown a mock breaking news article attributed to CBS News, describing either a recent avalanche or volcanic eruption and its associated victims. The two disaster issues were used as a form of stimulus sampling, which improves construct and external validity (Wells & Windschitl, 1999). Based on the identifiable victim effect (Slovic, 2007), the number of victims depicted in the AI-generated news image was manipulated to show either a single victim or five victims, and we aimed to examine whether the number of victims depicted in AI-generated disaster news images would influence moral judgments of AI use. However, victim count did not produce significant main effects or interactions with disclosure strategy on any outcome variable (all ps > .05). Given space constraints, we do not report these results in the main text. Means and standard deviations for all experimental conditions, including victim count, are reported in the Supplemental Materials.

Participants were randomly assigned to see versions of the news article that varied along the following dimensions: (a) disclosure strategy: the article included either a real photo (no AI disclosure), an AI-generated image with a simple disclaimer (no justification), an AI-generated image attributed to an internal reason (i.e., cost-saving), or an AI-generated image attributed to an external reason (i.e., protecting the victim’s privacy); (b) Victim count: the article depicted either one or five victims; and (c) Disaster context: the article described either a volcanic eruption or an avalanche. Immediately after reading their assigned article, participants completed a post-treatment survey measuring their attitudes toward the use of AI-generated images in the news article. At the end of the survey, all participants were debriefed and informed about the purpose of this study and the experimental manipulations.

Design and Stimuli

We created the stimuli in May 2025, comprising three elements: news articles, AI-generated images, and disclosure strategies. The articles were adapted from two real news articles about volcanic eruptions and avalanches, but were modified to reflect breaking news scenarios where firsthand reporting was infeasible. Each original article was rewritten into a 110-word news story with fabricated details (e.g., author, date, and location) to fit our research context. To manipulate the victim count, we created two versions of each article, one describing a single victim and the other describing multiple victims, while keeping all other content consistent.

We utilized Sora, a text-to-image model developed by OpenAI, to generate news images tailored to the news article. The first author iteratively crafted and refined prompts based on each article to produce realistic disaster images. When outputs were unsatisfactory or contained any flaws, the author revised and resubmitted the prompts. This process was repeated until high-quality images were obtained. All article versions, image generation prompts, and images were included in the Supplemental Materials.

To develop appropriate AI disclosure strategies, we conducted a pilot study on June 1, 2025, to identify internal and external attributions that align with public perceptions. Specifically, we sought to determine which reasons for using AI-generated images in disaster news were perceived as altruistic (external) or self-serving (internal). We began by drafting five candidate justifications. Then, we recruited 100 participants via Prolific and randomly assigned them to rate the perceived self-serving intentions of each justification. The results suggest that the justification, protecting the victims’ privacy, was perceived as the least self-serving (M = 2.36) justification, whereas cost-saving was rated as the most self-serving (M = 4.03) reason. Paired t-test confirmed these differences were statistically significant (t = −10.05, p < .001). Thus, we selected “protecting the victim’s privacy” as external attribution and “cost-saving” as internal attribution. Figure 2 illustrates the sample news article and accompanying images, each with different disclosure messages. More details about the pilot study and selection process are included in the Supplemental Materials.

OPEN IN VIEWER

Figure 2.

The avalanche disaster news article included an AI-generated image accompanied by different disclosure strategies. The versions featuring multiple victims or the real photo (control) are shown on the right side. See Supplemental Materials for all stimuli.

Participants

We conducted the main study on June 5, 2025, recruiting a total of 1,381 English-speaking participants based in the United States from Prolific. We chose Prolific because it can provide comparatively high-quality data, including stronger attention and comprehension performance than other crowdsourced platforms (Peer et al., 2021). Each participant received an average compensation of $2.5. Participants were randomly assigned to one of the sixteen experimental conditions described previously. Individuals who had participated in the pilot study were not eligible for the main study. To ensure data quality, we applied multiple exclusion criteria. First, we removed participants who either failed at least one of two attention check questions or reported issues during the survey, resulting in 1,298 valid responses. Next, we excluded participants with unusually long survey completion times, since unusual long durations could indicate that respondents became distracted or disengaged partway through the study. Following the outlier detection approach proposed by Leys et al. (2013), we excluded cases with completion times exceeding three median absolute deviations (MADs) above the median. This resulted in a final sample of 1,171 participants.

Among the participants, 45.35% identified as male and 52.86% identified as female, and 1.79% preferred not to disclose. The average age of participants was 40.13 (SD = 13.98). The median education level was a bachelor’s degree, and the median annual household income fell within $70,000 to $79,999. The racial or ethnic background of participants was as follows: White 63.88%, Black 24.77%, Asian 4.01%, Hispanic or Latine 3.84%, American Indian or Alaska Native 0.85%, Native Hawaiian or other Pacific Islander 0.17%, Other racial or ethnic types 1.79%, and 0.68% preferred not to disclose.

Measures

All items were measured using a 5-point scale ranging from 1 = strongly disagree to 5 = strongly agree. The mean and standard deviation for each condition, full-scale items, and correlation plots for all measured variables are provided in the Supplemental Materials.

Manipulation Checks

For the manipulation of disclosure strategies, we asked participants exposed to either external or internal attribution about their perceptions of the newsroom’s motive for using an AI-generated image (1 = entirely altruistic, 5 = entirely self-serving). For the victim count, participants were asked whether the news image included at least one person (0 = no, 1 = yes), and if they said yes, we further asked them to indicate whether the image showed more than one person (0 = no, 1 = yes). For the manipulation of disaster context, all participants identified the disaster described in the article (1 = Volcanic Eruption, 2 = Hurricane, 3 = Wildfire, 4 = Avalanche).

The one-way ANOVAs results showed that participants in the internal attribution condition were significantly more likely to perceive the justification as self-serving (M = 3.46, SD = 1.22) than those in the external attribution condition (M = 3.12, SD = 1.23), F(1, 592) = 11.22, p < .001, partial η² = .019, suggesting our manipulation of disclosure strategies was successful. Moreover, our chi-square tests showed that 95.4% of participants in the multiple-victim condition reported seeing more than one person, while 95.4% in the single-victim condition reported seeing only one, χ²(1) = 751.73, φ = .910, p < .001, which validates our manipulation of the victim count. For the disaster context, our chi-square results indicated that most participants in the volcano condition identified it as a volcanic eruption (98.9%), and most in the avalanche condition selected avalanche (99.3%), χ²(3) = 1,099.93, p < .001, suggesting this manipulation was also successful.

Effects of Disclosure Strategy on Trust, AI Acceptance, and Perceived Legitimacy

Our H1−H4 proposed that compared to the simple disclaimer and internal attribution, external attribution would increase trust, AI acceptance, and perceived legitimacy. Indeed, there was a significant main effect of the disclosure strategy condition on trust in the news organization, F(3, 1,155) = 24.58, p < .001, η² = .06. Post hoc pairwise comparisons showed that participants who viewed real photos (control condition) reported significantly higher trust (M = 3.34, SD = 0.89) than those shown AI-generated images with a simple disclaimer (M = 2.67, SD = 1.15; p < .001), an external attribution (M = 2.82, SD = 1.11; p < .001), and internal attribution (M = 2.67, SD = 1.08; p < .001). However, trust levels did not significantly differ among the three AI disclosure conditions, providing no support for H1a and H2a. In summary, we found that regardless of how AI usage was disclosed, AI-generated images consistently reduced public trust in news organizations.

We found a similar pattern for acceptance of AI in journalism. There was a significant main effect of disclosure strategy, F(3, 1,155) = 3.66, p = .012, η² = .009. Participants in the real photo condition reported greater AI acceptance (M = 2.93, SD = 1.21) than those in the simple disclaimer condition (M = 2.59, SD = 1.27; p < .001). However, AI acceptance did not differ significantly between the control and external attribution (M = 2.81, SD = 1.30; p = .264), or between the control and internal attribution (M = 2.74, SD = 1.27; p = .071). Participants in the external attribution condition reported significantly higher AI acceptance than those in the simple disclaimer condition (p = .035), while there were no significant differences between the simple disclaimer and internal attribution (p = .143) or between the external and internal attribution (p = .498). H1b got supported, but H2b was rejected.

We found a significant main effect of disclosure strategy on perceived legitimacy, F(3, 1167) = 73.49, p < .001, η² = .160. Participants in the control (real photo) condition reported the highest perceived legitimacy (M = 4.00, SD = 0.73), significantly higher than those in the simple disclaimer (M = 2.78, SD = 1.20; p < .001), external attribution (M = 3.05, SD = 1.21; p < .001), and internal attribution (M = 2.86, SD = 1.19; p < .001) conditions. External attribution was perceived to be more legitimate than the simple disclaimer (p = .003) and internal attribution (p = .030). Both H3 and H4 were supported.

Perceived Legitimacy Mediated the Effects of Disclosure Strategies on Trust and AI Acceptance

To test H5, we ran hierarchical regression models. For trust in the news organization, Model 1 (without legitimacy) was significant, F(5, 1165) = 16.07, p < .001, adjusted R² = .060. Adding perceived legitimacy in Model 2 significantly improved model fit, F(6, 1164) = 137.50, p < .001, with adjusted R² increasing to .412. Perceived legitimacy strongly predicted trust (b = 0.588, SE = 0.022, p < .001). H5a was supported. For AI acceptance, Model 3 was not significant, F(5, 1165) = 2.22, p = .050, adjusted R² = 0.005. However, including perceived legitimacy in Model 4 significantly improved the model fit, F(6, 1164) = 145.8, p < .001, with adjusted R² increasing to .426. Perceived legitimacy strongly predicted AI acceptance (b = 0.746, SE = 0.026, p < .001), supporting H5b. Full regression results are provided in Table 1.

OPEN IN VIEWER

Table 1.

Hierarchical Regression Results.

	Trust in news organization		Acceptance of AI in journalism
	Model 1	Model 2	Model 3	Model 4
Disclosure strategy (ref. Real photo)
Simple disclaimer	−0.661*** (0.089)	0.057 (0.075)	−0.342** (0.106)	0.568*** (0.086)
External attribution	−0.510*** (0.089)	0.044 (0.074)	−0.121 (0.107)	0.582** (0.085)
Internal attribution	−0.665*** (0.088)	0.009 (0.074)	−0.190 (0.105)	0.664*** (0.085)
Disaster: volcano eruption (vs. avalanche)	−0.133* (0.062)	−0.058 (0.049)	0.022 (0.074)	0.117* (0.056)
Single victim (vs. multiple victims)	0.039 (0.062)	−0.019 (0.049)	−0.004 (0.074)	−0.078 (0.056)
Perceived legitimacy		0.588*** (0.022)		0.746*** (0.026)
Constant	3.384*** (0.076)	1.024*** (0.108)	2.922*** (0.090)	−0.073 (0.123)
Observations	1,171	1,171	1,171	1,171
R 2	0.065	0.415	0.009	0.429
Adjusted R2	0.061	0.412	0.005	0.426

*

p < .05; **p < .01; ***p < .001.

We then tested H6, which predicted that perceived legitimacy would mediate the relationship between disclosure strategies and both trust (H6a) and AI acceptance (H6b). Using the real photo as the reference, all three AI disclosure strategy conditions produced significant negative indirect effects on trust via perceived legitimacy: simple disclaimer (b_indirect = −0.718, SE_indirect = 0.058, 95%CI = [−0.832, −0.602]), external attribution (b_indirect = −0.554, SE_indirect = 0.054, 95%CI = [−0.659, −0.451]), and internal attribution (b_indirect = −0.673, SE_indirect = 0.056, 95%CI = [−0.785, −0.566]). When using the simple disclaimer as the reference, the indirect effect of external attribution was significantly positive (b_indirect = 0.164, SE_indirect = 0.059, 95%CI = [0.048, 0.279]), whereas the indirect effect of internal attribution was non-significant (b_indirect = 0.044, SE_indirect = 0.057, 95%CI = [−0.066, 0.159]). Using internal attribution as the reference, the indirect effect of external attribution on trust through perceived legitimacy was significantly positive (b_indirect = 0.119, SE_indirect = 0.057, 95%CI = [0.006, 0.229]).

For AI acceptance (H6b), compared to the real photo condition, the simple disclaimer condition (b_indirect = −0.911, SE_indirect = 0.070, 95%CI = [−1.051, −0.775]), external attribution (b_indirect = −0.703, SE_indirect = 0.069, 95%CI = [−0.841, −0.570]), and internal attribution (b_indirect = −0.854, SE_indirect = 0.068, 95%CI = [−0.986, −0.719]) had significant negative indirect effects through perceived legitimacy. Using the simple disclaimer as reference, the indirect effect of external attribution on AI acceptance via perceived legitimacy became significantly positive (b_indirect = 0.208, SE_indirect = 0.073, 95%CI = [0.062, 0.354]), while the indirect effect of internal attribution was non-significant (b_indirect = 0.056, SE_indirect = 0.073, 95%CI = [−0.083, 0.200]). Compared to internal attribution, external attribution had a significantly positive indirect effect on AI acceptance via perceived legitimacy (b_indirect = 0.151, SE_indirect = 0.073, 95%CI = [0.011, 0.300]). Based on the above, H6 was supported: across comparisons, perceived legitimacy significantly mediated the effects of different disclosure strategies on trust and AI acceptance.

Effects of Disclosure Strategies on Legitimacy Were Conditional on Attitudes Toward AI

Turning to moderated mediation (RQ1), we tested whether attitudes toward AI moderated the indirect effects of disclosure strategies on outcomes via perceived legitimacy. Taking trust in the news organization as the outcome and the real photo condition as the reference, the simple disclaimer (index = 0.160, SE = 0.050, 95% CI = [0.063, 0.256]), external attribution (index = 0.269, SE = 0.047, 95% CI = [0.176, 0.360]), and internal attribution (index = 0.214, SE = 0.047, 95% CI = [0.122, 0.304]) showed significant moderated mediation effects on trust. Each AI disclosure condition showed similar patterns across participants with various attitudes toward AI. Specifically, compared to the real photo, all AI disclosures significantly eroded trust through reduced legitimacy, but these indirect effects became weaker among individuals with more positive attitudes toward AI. However, no significant moderated mediation effects were found when comparing different AI disclosure strategies against one another.

We observed a similar pattern when taking the acceptance of AI in journalism as the outcome and the real photo condition as the reference. Again, the simple disclaimer (index = 0.202, SE = 0.063, 95%CI = [0.080, 0.325]), the external attribution (index = 0.339, SE = 0.059, 95%CI = [0.223, 0.449]), and the internal attribution (index = 0.270, SE = 0.060, 95%CI = [0.155, 0.387]) produced significant moderated mediation effects. Compared to the real photo, all AI disclosure strategies reduced AI acceptance indirectly via lower legitimacy, but the magnitude of these indirect effects diminished among participants with more positive attitudes toward AI. Further comparisons among different disclosure strategies highlighted the advantage of the external attribution relative to a simple disclaimer (index = 0.137, SE = 0.071, 95% CI = [0.003, 0.275]). Compared to the simple disclaimer, the external attribution produced significantly more positive indirect effects on trust among participants with neutral or positive attitudes toward AI, whereas the internal attribution showed no significant moderated mediation effects on trust (index = 0.068, SE = 0.073, 95% CI = [−0.073, 0.213]). Using the internal attribution as the reference, the external attribution also showed no significant moderated mediation effects on trust (index = 0.069, SE = 0.068, 95% CI = [−0.068, 0.198]). All statistics for the conditional indirect effects, standard errors, and confidence intervals are included in the Supplemental Materials.