The proliferation of misinformation online, particularly evident during events like the 2016 U.S. Presidential Election and the COVID-19 pandemic, poses a significant challenge. While the extent of its impact is debated, the need for effective countermeasures is clear. This research focuses on a promising approach: shifting users' attention towards accuracy to improve the quality of news they share. Simply being exposed to misinformation can increase belief in it, and its spread through social media networks amplifies its potential for harm. Previous research has shown a disconnect between users' ability to distinguish true from false headlines and their sharing intentions; even when capable of discerning accuracy, they may still share misinformation. This disconnect, while influenced by factors like animosity towards political opponents and personality traits, suggests that inattention to accuracy plays a critical role. Experiments have demonstrated that prompting participants to consider accuracy – for instance, by asking them to evaluate the accuracy of a headline – reduces this disconnect and leads to improved news sharing quality. This effect has been replicated, various effective prompts identified, and even shown in a field experiment on Twitter. However, questions remained regarding the effect's mechanism (reducing false news sharing or increasing true news sharing), its moderation by individual differences (political ideology, attentiveness), and its persistence over time. This study addresses these gaps through a rigorous meta-analysis.

Existing research highlights the detrimental effects of online misinformation on beliefs and behaviors. Studies have shown that exposure to misinformation increases subsequent belief in false claims, and the networked nature of social media platforms accelerates the spread of these claims. Prior research has also explored the disconnect between accuracy judgments and sharing intentions, finding that even individuals who can identify false information may still share it. This suggests a role for factors beyond deliberate falsehood sharing, such as inattention to accuracy. Several studies demonstrated the potential of accuracy prompts to mitigate this problem by influencing sharing behaviors through increased attention to accuracy. However, these prior studies lacked the scope and rigor to fully establish replicability and generalizability across various contexts and populations.

This study employs an internal meta-analysis of 20 experiments (N = 26,863 participants) conducted by the authors' research group between 2017 and 2020. The key strength of this approach is its ability to mitigate publication bias and p-hacking, common threats to meta-analytic validity. By including all relevant studies regardless of outcome and using identical analyses across all studies, the authors eliminate these biases. The experiments varied in several key aspects: * **Accuracy Prompts:** Several different accuracy prompts were used, including evaluating a neutral headline, assessing the importance of sharing accurate news, presenting social norms about accuracy, showing a PSA video, prompting reasoned versus emotional sharing, and providing digital literacy tips (Table 1 provides details). * **Headline Sets and Topics:** The studies employed different sets of headlines covering both political and COVID-19 related news. * **Subject Pools:** Participants were recruited through various online platforms, including Amazon Mechanical Turk (MTurk), Lucid, and YouGov, offering a range of sample characteristics. * **Individual Differences:** Demographic data (gender, race, age, education), political ideology and partisanship, self-reported importance placed on accuracy, cognitive reflection, and attentiveness were collected to assess moderation effects. Each study involved random assignment of participants to control or treatment conditions. Sharing intentions were measured using Likert scales, and were analyzed using rating-level linear regressions with robust standard errors clustered on participant and headline. Meta-analytic estimates were generated using random-effects models, reflecting the expected variation in effect sizes across studies. Meta-regression was used to examine study-level moderators (subject pool, headline topic, baseline sharing discernment). Headline-level analyses examined the correlation between perceived accuracy and the treatment effect, and individual-level moderation analyses explored the impact of various individual characteristics on the treatment effect.

The meta-analysis revealed a significant overall effect of accuracy prompts on sharing discernment. Accuracy prompts significantly increased the quality of news shared, primarily by reducing the intention to share false news (a 10% decrease relative to control). This effect was consistent across headline topics (politics and COVID-19) and did not decay over successive trials. Moderation analyses showed that: * The effect was significantly larger for participants from MTurk samples compared to more representative samples from Lucid/YouGov. * The effect was smaller for headline sets with higher baseline sharing discernment (i.e., where participants were already more discerning). * The effect was larger for older participants in the more representative samples. * The effect was larger for college-educated participants in both sample types. * There was no significant moderation by gender, race, political ideology, or self-reported value of accuracy. * The treatment effect was significantly larger for politically concordant headlines (i.e., aligning with participants' political leanings), probably due to higher baseline sharing of such headlines. * There was a strong correlation between the treatment effect on a headline's sharing and that headline's perceived accuracy (r = 0.773). The accuracy prompt effect was most effective on headlines perceived as less accurate (Figure 5). * Combining different accuracy prompts resulted in a substantially larger treatment effect, suggesting additive or synergistic benefits of multiple interventions. * The effect on sharing discernment was not unique to the evaluation treatment (asking participants to evaluate a neutral headline), with non-evaluation treatments showing similar increases in sharing discernment. However, the Evaluation prompt was the most consistently studied type of intervention.

The findings strongly support the replicability and generalizability of the accuracy prompt effect. The consistent reduction in sharing of false news across diverse contexts suggests that the effect is driven by a general mechanism, likely the increased attention to accuracy. The lack of robust moderation by demographic or ideological factors is particularly encouraging, suggesting that this intervention could be broadly effective. The correlation between perceived accuracy and the treatment effect further supports this interpretation, showing that the intervention works by targeting headlines perceived as inaccurate. The larger effect sizes observed in MTurk samples might be attributed to higher attentiveness levels in this population; however, cautions were raised concerning the limitations of this sample's representativeness for making strong claims about political differences. The similar efficacy for political and COVID-19 news indicates the generalizability of the intervention to various topics. Furthermore, the persistence of the effect over trials, and the significant increase in the size of the effect with the addition of more prompts, suggests further potential improvements by combining interventions.

This meta-analysis provides strong evidence for the effectiveness of accuracy prompts in reducing the spread of misinformation online. The intervention is replicable, generalizable across various contexts, and does not require platforms to arbitrate truth. The findings highlight the potential for scalable interventions to promote better news sharing quality. Future research should focus on testing the intervention in broader cultural contexts, exploring diverse delivery methods, investigating the underlying mechanisms in detail (possibly using computational models), and examining the interactions with other misinformation interventions.

While the internal meta-analysis effectively minimizes publication bias and p-hacking, its reliance on a single research group's data limits the generalizability across different research teams. The lack of external replication studies reduces the certainty of the generalizability beyond the specific research group. Future research could benefit from more data from external groups and cross-cultural experiments, as well as a more detailed mechanistic exploration of how the prompts operate and how various prompts and interventions might interact.