On the same page? Experts are mostly, but not always aligned about disinformation in times of generative AI

Implications

In an era when generative AI heightens concerns over disinformation, such as deepfakes and synthetic images (Adami, 2024), policymakers face pressure to design effective mitigation strategies (Griffin, 2025; Whyte, 2020). To inform evidence-based responses, they often rely on expert groups engaged in countering disinformation (Chystoforova & Reviglio, 2025). These experts come from diverse professional backgrounds, raising questions about whether they share a common understanding of AI’s impact. While consulting varied viewpoints is essential to tackle the issue from various angles, significant differences in assessments may lead to inconsistent messaging, and political actors may dismiss expert input as divided and therefore unreliable.

Previous research has examined how academics, fact checkers, and journalists assess AI-driven disinformation, though rarely in direct comparison (Kruger et al., 2024). Academics often see themselves as neutral information providers and emphasize platform design (Altay et al., 2023; Chystoforova & Reviglio, 2025), balancing warnings about deepfake threats with cautions against exaggeration (Chesney & Citron, 2019; Simon et al., 2023). Similarly, fact checkers tend not to overstate the dangers of AI-generated disinformation and see it as just one of many forms of disinformation they encounter (Weikmann & Lecheler, 2023). Given their hands-on approach, they often find new and creative ways to tackle this growing task (van Huijstee et al., 2021). Journalists, on the other hand, appear more alarmist, which may come from insufficient experience in recognizing AI-generated content and a tendency to take a pessimistic approach when reporting about AI (Wahl-Jorgensen & Carlson, 2021). Overall, these findings indicate that diverse professional backgrounds may lead to distinct perspectives, which need to be understood to interpret and implement experts’ policy recommendations adequately. Our study addresses this gap by directly comparing stakeholder perceptions across three dimensions: (1) self-perceived competence in dealing with disinformation in times of generative AI, (2) perceived risks posed by AI-generated disinformation, and (3) the attribution of responsibility in mitigating it.

Our findings indicate that, overall, the surveyed experts feel competent in addressing disinformation in the era of generative AI, with fact checkers standing out as the most confident. This finding is unsurprising given their comparatively greater familiarity with detection tools (Weikmann & Lecheler, 2023). Nonetheless, the significant gap between fact checkers and journalists is noteworthy and consistent with prior research (Gregory, 2021). Experts mostly agree on perceived risks. Importantly, two perspectives coexist: AI-generated disinformation is viewed both as part of long-standing challenges tied to emerging technologies (Simon et al., 2023) and as a potential source of confusion or a threat to democratic debate (Kruger et al., 2024). This suggests that experts do not regard these positions as mutually exclusive. We also found that journalists are significantly more likely to agree that amid AI seeing does no longer mean believing. This aligns with the tone of many news headlines about deepfakes and the generally alarmist stance the profession tends to adopt (Gosse & Burkell, 2020)—even among journalists with disinformation expertise. When considering how this translates into EU-policy frameworks, it seems like they overall rather adopt a pragmatic stance, emphasizing the importance of verification, monitoring, and research. Across experts, there is a strong consensus that social media platforms bear responsibility for mitigating the potential impact of AI-driven disinformation, corroborating previous findings (e.g., Altay et al., 2023). In line with the growing emphasis on platform responsibility, the Digital Services Act (DSA) has risen prominently on policymakers’ agendas in recent years. Interestingly, academics assign significantly more responsibility in countering AI-generated disinformation to news users than fact checkers do, reflecting a common scholarly emphasis on media literacy as a means of countering disinformation (Bulger & Davison, 2018). However, current EU policy frameworks address media literacy rather indirectly, for instance, through commitments made by platforms to “empower users” in the Code of Conduct on Disinformation. This finding also emerges at a time when researchers are increasingly examining the effectiveness of community notes, which Meta designated in January 2025 as its exclusive method of moderating disinformation. Lastly, fact checkers place greater responsibility on themselves, a stance that aligns with both their perceived competence and their frequent role as first responders in identifying and addressing false information (Graves & Mantzarlis, 2020). 

These findings are relevant for both policymakers consulting disinformation experts and academics studying the impact of generative AI. For policymakers it is important to consider the background of the experts they engage with, as differences in expertise can subtly influence perspectives regarding mitigation responsibilities. In turn, they may propose different solutions, such as more consumer-driven in the case of academics, while fact checkers are favoring technology-driven social interventions. At the same time, experts overall agree on their risk assessments and are generally confident in their preparedness to address AI-driven disinformation, further legitimizing their role. For researchers, the findings suggest several avenues for further investigation. One is to explore differences in competence between regular journalists who do not specialize in disinformation and fact checkers, which may reveal a greater divide between these closely related expert groups. Moreover, future research could check whether there is also a divide in actual competence. Even though their responses are grounded in professional knowledge and experience, our study only captures experts’ self-perceptions. Another is to continue tracking expert perceptions over time, particularly given their shared view that AI-generated disinformation is a serious concern likely to cause increasing confusion in the future. While seen as part of a long-standing challenge, it is also seen as a pressing one, and expert voices may grow even more cautionary.

A noteworthy limitation of our study is the uneven and sometimes small group sizes, which we address through conservative analytical approaches and robustness checks (see Appendix A). In addition, we acknowledge that our study is specific to the EU expert ecosystem, which may not reflect conditions in the Global South (Vinhas & Bastos, 2025) or in more authoritarian contexts, where experts may be more hesitant about the role of the state in regulating disinformation (Tully et al., 2022). Moreover, further research is needed to examine how stakeholders’ preferences relate to actual policy, as assessing whether experts’ differing responsibility assignments translate into policy outcomes lies beyond the scope of this study. Interviews with policymakers, for instance, could clarify which inputs are taken into account and whether debates are shaped by more by alarmist narratives or more nuanced perspectives. Overall, we conclude that although expert groups may at times offer conflicting input, they share common ground on the roots and nature of disinformation (see also Altay et al., 2023; Kruger et al., 2024). Differences in proposed actions are best understood as complementary, reflecting the inherently shared responsibility of multiple actors (Young, 2006).

Findings

Finding 1: While overall confidence is high, fact checkers feel more competent to deal with AI-generated disinformation compared to academics and journalists.

Overall, the experts who filled out our online survey—academics (n = 47), fact checkers (n = 29), and journalists (n = 16)—reported a high average level of confidence when it comes to doing their work on AI-generated disinformation, signified by a high mean value (M = 5.49, SD = .98) measured on a scale from 1 to 7. This average score is higher than what had been observed in citizen samples using a similar measurement scale (e.g., Hopp, 2021; Weikmann et al., 2025). To explore differences between the expert groups in our sample, we conducted a Welch’s ANOVA, which revealed a statistically significant difference, F(2, 36.20) = 10.23, p < .001. We followed up with a Games-Howell post-hoc test. As can be seen in Figure 1, fact checkers (M = 6.05, SD = 0.69) felt significantly more competent than both academics (M = 5.27, SD = 0.89, p < .001) and journalists (M = 5.10, SD = 1.29, p = .034). No significant difference was found between academics and journalists (p = 1.00).

Finding 2: Experts are largely aligned regarding how they perceive the risks posed by AI-generated disinformation and emphasize the risk of confusion about what is real and fake.

We conducted Welch’s ANOVA to explore whether the surveyed experts agreed on the risks they believe are posed by AI-generated disinformation by having them rate six statements on a 7-point Likert scale, ranging from “Strongly disagree” to “Strongly agree” (see Figure 2). We could not find significant differences for most variables, demonstrating strong alignment overall; thus, we pooled participants’ responses. Ninety percent of experts agreed that AI-generated disinformation will cause confusion about what is real and fake, with 45% agreeing strongly. In addition, there was agreement that it will undermine democratic debate, indicated by 69% of the sample, who answered above ”Neither agree nor disagree” on the 7-point scale. When asked about the extent to which these risks may be overstated (Simon et al., 2023), there was more balance in the responses. While most agreed, 33% disagreed (lower than 4 on the 7-point scale) that AI-generated disinformation is part of moral panics about new technologies, and 25% disagreed (lower than 4 on the 7-point scale) that it is simply a variation of long-standing issues surrounding mis- and disinformation. When asked whether AI-generated disinformation will disrupt political decision-making, this question had the highest percentage of neutral responses, with 32% neither agreeing nor disagreeing. These findings are reported in the stacked bar plot below (see Figure 2).

Only one test indicated a significant difference across groups, specifically regarding the risk that people will stop believing in anything they see, F(2, 46.44) = 5.98, p = .005, which journalists (M = 5.81, SD = 0.98) rated significantly higher than both academics (M = 4.85, SD = 1.49, p = .015) and fact checkers (M = 4.62, SD = 1.66, p = .011). No significant difference was found between academics and fact checkers (p = .814).

Finding 3: Experts have differing views when asked who is responsible for countering AI-generated disinformation; fact checkers attribute more responsibility to online platforms compared to academics, who in turn hold news users more responsible.

We explored whether the expert groups attributed the responsibility for countering AI-generated disinformation and its societal impact differently. This was partially the case. While there was high consensus that very large online platforms are accountable (62% strongly agreed), fact checkers (M = 6.69, SD = .60) on average agreed significantly more than academics (M = 6.23, SD = .89); p = .026. In addition, fact checkers (M = 6.07, SD = .99) agreed significantly more that fact checkers (i.e., themselves) are responsible when compared to academics (M = 5.38, SD = 1.21); p = .025. Regarding news users, academics (M = 4.74, SD = 1.39) view those as significantly more responsible compared to fact checkers (M = 3.62, SD = 1.54); p = 0.006, but not compared to journalists (M = 4.06, SD = 1.73). We did not find any significant group differences regarding the attribution of responsibility towards journalists, political actors, or academics. Figure 3 visualizes agreement with the questions for all surveyed experts combined.

Methods

To explore expert perceptions regarding generative AI, we conducted an online survey (N = 92) among academics (n = 47), fact checkers (n = 29), and journalists (n = 16) who work towards mitigating disinformation. Participants were drawn from the European Digital Media Observatory (EDMO), a hub co-founded by the European Commission. As members of this network, we leveraged our access to its mailing lists, such as the fact-checking group and various hub-specific lists, as well as personal contacts, to distribute the survey and increase diversity in the sample. Data collection took place between January and May 2025 to ensure broad participation and maximize response rates. Recruiting journalists required additional efforts due to the small pool of this expert group specializing in disinformation. Ethical approval was obtained from the University of Amsterdam (FMG-11584).

Participants reported basic demographic information (gender: female: n = 32,male: n = 59, other: n = 1; age: M = 40.93, SD = 11.63; education level: bachelor’s degree: n = 11, master’s degree: n = 43, PhD: n = 38), indicated the percentage of their work that focused on topics related to disinformation (M = 65.45, SD = 30.71) and how many years of experience they had working on the topic (M = 7.51, SD = 5.76). Respondents were asked to indicate which job descriptions most accurately reflected their current role. One participant indicated in an open answer field that they had multiple roles equally, so they were excluded from the analysis. To ensure participants had a shared understanding, we presented them with a definition of disinformation, namely: “content that (1) contains verifiably false or misleading information, (2) has the potential to cause societal harm, (3) is intentionally disseminated, and (4) serves possible economic or political objectives” (Bleyer-Simon & Reviglio, 2024, p. 3). 

Perceived competence in dealing with AI-generated disinformation (M = 5.49, SD = 0.98, α = .77) was assessed by asking, on a 7-point Likert scale, the extent to which participants were confident in (1) their ability to detect disinformation effectively, (2) their knowledge and skills to conduct work around disinformation, and (3) their ability to distinguish between true and false or misleading information (see Lecheler et al., 2024).

Perceived risks posed by AI- generated disinformation were measured using six items on 7-point Likert scales. These items were inspired by previous research, including studies formulating concrete scenarios of potential risks posed by AI-generated disinformation (Chesney & Citron, 2019) and research suggesting that such risks may be overstated (Simon et al., 2023). As these items do not constitute a scale, no means, standard deviations, or Cronbach’s alphas are reported.

Attribution of responsibility for addressing AI- generated disinformation was measured by asking participants, on a 7-point scale, to what extent they believed the following actors were responsible for countering AI-generated disinformation: (1) very large online platforms (e.g., Meta, Google, X, TikTok), (2) journalists working for legacy media, (3) independent fact checkers, (4) governments and political actors, (5) independent experts and academics, and (6) news users. Again, no scale was constructed for this variable.

In addition to providing descriptive statistics, we conducted significance tests to explore differences between expert groups. Given the relatively small sample size and uneven group distributions, we relied on the more conservative Welch’s ANOVA combined with a Games-Howell post hoc test (see Delacre et al., 2019). We also conducted non-parametric tests as a robustness check, as reported in the Appendix.