Table 1.
Experiment design.
Fig 1.
Perceptions of the Quality of the Blogs Fig 1.
* Notes: * p < 0.1, ** p < 0.05, *** p < 0.01. This figure reports on the perceptions of blog quality measured using 10 attributes of high quality blogs [45,52–54]. Red and blue lines are coefficients and 95% confidence intervals from regressions without and with controls, respectively. Perceptions are measured using a 5-point Likert scale ranging from ‘strongly disagree’ (0) to ‘strongly agree’ (4) that the blog is high-quality with respect to that attribute. The 10 attributes are: blog had an appropriate amount of detail, a clear rationale, a catchy title, good visual presentation, an appropriate tone, was of adequate length, was easy to understand, and provided clear, relevant, and sufficiently detailed recommendations. The outcome in Panel a) is the simple average across 10 blog attributes, and the outcome in Panel b) is the weighted average across 9 attributes (weighted by respondents’ reported relative importance of attributes – tone was not rated). The outcome in Panel c) is the unweighted average for the same 9 attributes as in Panel b) for robustness. Means are reported on the left hand side and coefficients from a regression of the measure of quality on indicator variables for AI-generated blogs, AI-reported blogs, and their interaction are reported on the right hand side. We control for wave and strata (country, seniority, and gender) fixed effects. Standard errors are clustered at the individual level (the unit of randomization).
Fig 2.
Heterogeneity in Perceptions of the Quality of the Blogs Fig 2.
* Notes: * p < 0.1, ** p < 0.05, *** p < 0.01. This figure reports on the perceptions of blog quality measured using 10 attributes of high quality blogs [45,52–54] across four dimensions of heterogeneity. Panel a) reports on the grade level of the blog, Panel b) on whether the respondent has previously used AI, Panel c) on whether the respondent trusts AI to write high-quality blogs, and Panel d) by gender (female = 1). Red (attribute = 0), blue (attribute = 1), and black (difference) lines are point estimates and 95% confidence intervals. Perceptions are measured using a 5-point Likert scale ranging from ‘strongly disagree’ (0) to ‘strongly agree (4) that the blog is high-quality with respect to that attribute. The 10 attributes are: blog had an appropriate amount of detail, a clear rationale, a catchy title, good visual presentation, an appropriate tone, was of adequate length, was easy to understand, and provided clear, relevant, and sufficiently detailed recommendations. We control for respondent characteristics, wave, and strata (country, seniority, and gender) fixed effects. Standard errors are clustered at the individual level (the unit of randomization).
Fig 3.
Intended Engagement with the Blog Fig 3.
* Notes: * p < 0.1, ** p < 0.05, *** p < 0.01. This figure reports on the likelihood of the respondents’ engagement with the blog. Red and blue lines are coefficients from regressions without and with controls, respectively, and include the point estimate and 95% confidence interval. Engagement is measured across five actions: whether they are likely to re-read the blog, share the blog with others, look up studies cited in the blog, look up relate studies, or contact the authors. The responses follow a Likert scale ranging from 0 to 4, with 0 indicating ‘very unlikely’ and 4 indicating very likely’. The outcome in Panel a) is the average rating, Panel b) is the standardized score of the first principal component, and Panel c) is the number of actions where the respondent indicated they are ‘likely’ or ‘very likely’ to take that action. We control for wave and strata (country, seniority, and gender) fixed effects. Standard errors are clustered at the individual level (the unit of randomization).
Fig 4.
Beliefs About Others’ Intended Engagement with the Blog Fig 4.
* Notes: * p < 0.1, ** p < 0.05, *** p < 0.01. This figure reports on the respondents’ beliefs about others’ likelihood of engagement with the policy brief. Red and blue lines are coefficients from regressions without and with controls, respectively, and include the point estimate and 95% confidence interval. Engagement is measured across five actions: whether respondents believe that others are likely to re-read the blog, share the blog with others, look up studies cited in the blog, look up relate studies, or contact the authors. The responses follow a Likert scale ranging from 0 to 4, with 0 indicating ‘very unlikely’ and 4 indicating very likely’. The outcome in Panel a) is the average rating, Panel b) is the standardized score of the first principal component, and Panel c) is the number of actions where the respondent indicated that others are ‘likely’ or ‘very likely’ to take that action. We control for wave and strata (country, seniority, and gender) fixed effects. Standard errors are clustered at the individual level (the unit of randomization).