What Actually Makes Governments Regulate Technology?
Many people have an intuition about how technology gets regulated that goes like this: journalists write alarming articles, the public gets scared, politicians respond. It’s assumed that negative press leads to public fear, which leads to protective legislation. Disaster to coverage to regulation feels so intuitive that most people never question it. Yet the data tells a much more interesting and complex story.[1]
Over the past couple of months, I analyzed over a decade of global media coverage for two transformative technologies: artificial intelligence and nuclear energy. I made extensive use of the GDELT database, which tracks hundreds of thousands of news articles worldwide. These database entries were combined with Google Trends search data and a hand-coded database of regulatory events spanning 2015 to 2025. At the end, three findings jumped out, and two of them cast doubt on the typical story.
The findings were:
1. Positive coverage of AI actually predicts more regulation, not less. The most enthusiastic period of AI journalism in the dataset coincided with the sharpest acceleration in regulatory activity. This isn’t what the Fukushima model predicts.
2. The most reliable predictor of when governments actually regulate a technology is neither media tone or coverage volume. It’s actually Google search interest, just the boring act of ordinary people typing things into a search bar. This result held across every statistical specification applied.
3. The one directional relationship the data actually confirms runs backwards from the standard story. Meaning, as public interest in AI grows, media coverage becomes more positive. Journalists appear to follow audience enthusiasm at least as much as they shape it.
Tracking Two Technologies Over One Decade
Why analyze both nuclear energy and AI technologies? Both technologies are high risk technologies, and both technologies are seeing recent surges in funding and development. Comparing the two technologies helps us contrast two different strains of news coverage (positive and negative) to see if news sentiment influences regulation around high risk technologies specifically.
AI coverage over the past decade has been relentlessly positive. The average tone score starts slightly negative in 2015, crosses into positive territory almost immediately, and climbs steadily through 2025.
By the end of the sample, the average AI article is up from roughly −0.2 in early 2015 to over +2 by late 2025. That’s a swing of roughly 1.5 points on a scale where many technologies cluster near zero. There are also a lot of AI articles roughly twelve times more than nuclear energy in a typical month.
Nuclear energy coverage, meanwhile, is a flatline below zero. Persistently negative, seeing no real improvement in over ten years. Even while climate policy has pushed nuclear back into serious consideration for energy technologies, the shadow of events like Fukushima and Chernobyl continues to linger.
The gap in coverage sentiment becomes even more dramatic when you split the data by media type. State-controlled outlets (Xinhua, RT, France24) consistently cover both technologies more positively than independent media. For example, for nuclear energy state outlets average a slightly positive tone (+0.28), while independent outlets average deeply negative (-0.91), a huge gap in sentiment. Notably, governments in France, China, and Russia operate major nuclear programs, and have incentive to frame the technology favorably. Independent journalists are more likely to write what they see, or are otherwise incentivized differently, writing about risk, waste, and accident legacy.
For AI, both state (+1.25) and independent media (+0.72) are positive, with enthusiasm being cross-institutional. State media remains a bit more enthusiastic, consistent with governments positioning AI as a national strategic priority.
The Agenda-Setting Question
The conventional theory of media influence, the agenda-setting theory in the academic literature, says that journalists set the public agenda. Journalists decide what to cover and then the public follows. My own analysis finds something rather more complicated.
AI News Coverage
In terms of AI coverage, it’s tempting to read the data through the classic agenda-setting lens.
When you plot media tone and search interest against each other, you find that AI media tone and public search interest are strongly correlated across the entire decade. Cross-lag correlations are used to test if one of these variables predicts changes in the other better than the reverse, say if media tone better predicts fluctuations in public search interest. The correlations were consistently strong, at time horizons of one to six months (peak r = +0.641 at six months).
However, the cross-correlation function is near-symmetric in both directions. This means the peak correlation alone isn’t a directional finding. Today’s search interest predicts future media tone at almost the same strength as today’s tone predicts future search interest. When a cross-correlation function looks the same in both directions, the most likely explanation is that both are rising together, pulled up by the same underlying force.
In this case, that force is probably a decade of genuine AI breakthroughs, from GPT-3 to AlphaFold to ChatGPT. These discoveries simultaneously generated journalistic coverage and public curiosity, without either one clearly preceding the other.
There’s another reason to doubt that journalists lead the public on this issue: the Granger causality test. While a “cross-lagged correlation” simply shows that two things move together with a delay, a Granger test asks a more rigorous question: “Does knowing the past of Variable A give me new information that helps predict Variable B, even after I already know everything about Variable B’s own history?” The test asks whether one variable predicts the other’s future values beyond what it would predict on its own. The outcome of the test is that AI media coverage doesn’t significantly predict future public search interest (p = 0.905), finding nothing in the journalists-lead-the-public direction.
One result does actuallysurvive rigorous correction for multiple comparisons, but it runs the other way: public search interest predicts more positive subsequent AI media coverage (adjusted p = 0.006). As more people start paying attention to AI, journalists write about it more favorably. The audience appears to shape the framing at least as much as the press shapes the audience.
Nuclear News Coverage
For nuclear energy, there’s actually a suggestive signal that public attention leads coverage rather than follows it. This is consistent with the idea that nuclear interest is activated by external shocks, like energy crises or climate reports, as opposed to journalism campaigns.
…But this result doesn’t survive correction for multiple comparisons (adjusted p = 0.199), so it would be wrong to call it confirmed. Yet there’s no compelling evidence, in this data, that journalists are setting the public agenda on nuclear either.
Basically, for both technologies, media coverage and public attention appear to co-evolve, impacting each other in a mutually reinforcing relationship. The data can’t cleanly establish which comes first. What it can say is that once established, the feedback loop (wherever it comes from) is real and that it feeds directly into the regulatory dynamic described below.
Something worth noting is that the relationship between AI news tone and subsequent search interest has become stronger over time. In the early period (2015-2019), the correlation between AI tone and search interest was modest and not statistically significant (r = +0.187, p = 0.161). In the later period (2020-2025), it roughly doubled and became clearly significant (r = +0.364, p = 0.002).
Media coverage and public attention to AI have become increasingly coupled, each amplifying the other more strongly than they did a decade ago. It’s not clear what the mechanism is. It’s possible that this is journalism gaining agenda-setting power or audiences are just becoming more responsive to coverage or simply a maturing feedback loop, the two are now harder to disentangle than ever.
The Core Finding: Search Interest Predicts Regulation
The finding that seems to hold up the best is this: Higher search interest predicts more regulation.
To test for this, a panel regression was run. A panel regression essentially watches the same group (in this case, a technology) over multiple time periods to see how it changes. The goal is to watch how the relationship between the dependent variables and independent variables change over time, controlling for unobserved factors that don’t change over time. The fixed effects panel regression tested if, after controlling for the nature of the evolving nuclear or AI technologies, media variables predicted future regulatory events.
Three potential predictors were tested: media coverage volume, media tone, and Google search interest.
Coverage volume added only a little by itself. Once you control for search interest, the number of articles about a technology has no independent predictive power for regulation.
Media tone has a marginal, short-lived effect. At the one-month time step, positive media tone shows a weak positive association with regulatory activity (p = 0.063 in the Poisson model). At longer horizons, even this fades. Search interest, on the other hand, is rock-solid as a regulatory predictor.
When you look at Google Trends search interest data, it presents a search index metric that tracks search volume popularity over time and location. For every one-unit increase in Google search index, there was an associated 2.6% increase in regulatory events one month later.
A one-unit increase in the Google Trends search index is associated with a 2.6% increase in expected regulatory events one month later (IRR = 1.026, p = 0.008), 3.1% at two months (IRR = 1.031, p = 0.010), and 8.1% at four months (IRR = 1.081, p < 0.001). In other words, every step up the Google Trends popularity scale correlates with a small bump in regulatory activity, peaking at about 4 months later in a process that gathers steam.
The fact that the effect is most prevalent at longer time horizons makes sense intuitively. Legislative drafting takes time. A spike in public interest doesn’t produce a regulatory event next week, it produces one next quarter. The signal compounds as attention sustains itself.
Six robustness checks were done using a linear model, varying the lag structure and the tone metric. The search interest results held across all of them. with coefficients ranging from +0.020 to +0.036, all significant at p < 0.05. This is about as robust as observational social science gets.
Concretely, public attention is a more reliable predictor of regulatory action than anything journalists write or how they write it.
The Mediation Question: How Does the Chain Work?
So if media tone predicts search interest, which in turn predicts regulation, here’s the next natural question: Does attention mediate the tone-regulation relationship, with positive coverage driving regulation through public attention?
The data pattern is genuinely consistent with this story. Positive media tone strongly predicts public search interest (p < 0.001), and public search interest strongly predicts regulatory events (p < 0.001). When you add search interest as a control, the direct tone-to-regulation path drops from marginal significance (p = 0.053) to clearly non-significant (p = 0.362): a 52% attenuation.
That certainly looks like mediation, but I do want to be careful here. The formal framework for mediation technically requires that the direct path be significant before you add the mediator. In this case the significance before adding the mediator was p = 0.053. This is close, but remains outside the conventional threshold. The attenuation could reflect genuine mediation, or it could just be power loss from adding a correlated predictor to an already-weak effect.
So while the data is consistent with a story where positive coverage → public attention → regulation, they don’t confirm it in the strict statistical sense.
The findings are coherent and replicated in both OLS and Poisson models, and for what it’s worth would be my best guess for the actual mechanism. Still, let’s remain appropriately skeptical.
The ChatGPT Tsunami
The post-ChatGPT period serves as a natural experiment, or at least a vivid illustration, of the dynamics described above.
Before November 2022, AI regulatory events averaged 0.15 per month. After ChatGPT launched, this jumped to 1.00 per month. This coincided with the most positive sustained AI coverage in the entire dataset and the largest spike in public search interest.
This is the whole mechanism we’ve covered (positive coverage fueling public attention, public attention driving regulatory action) compressed into one observable episode. The EU AI Act (agreement on revision to cover general-purpose AI models in December of 2023), the Biden executive order on AI (Nov 2023), and the Bletchley Park summit (Nov 2023) all arrived in the wake of an unprecedented surge in public interest… interest that was itself preceded by overwhelmingly positive media coverage.
Note what didn’t happen: the coverage wasn’t dramatically negative. Some outlets had written “AI may destroy civilization” pieces, but the dominant narrative was enthusiasm and possibility. As has been covered elsewhere, most AI news stories in 2023 were still about corporate use, macrotrends, and scientific research. And regulators responded not by restricting the technology because the public was afraid, but by governing it because the public was paying attention.
What This Means
Two practical implications fall out of this analysis.
For AI governance: If you’re a policymaker trying to anticipate when regulatory pressure will build, watch Google Trends, not headlines. Public search interest is a one-to-three-month leading indicator of regulatory activity, and it’s more reliable than media tone or volume.
Also, don’t assume that positive coverage insulates a technology from regulation. It’s quite possible that positive coverage drives attention which drives governance. The post-2024 shift toward more restrictive AI legislation (coinciding with stabilizing search interest and lower tone volatility), may signal a maturing regulatory cycle where governance becomes institutionalized rather than reactive.
For nuclear energy: The data offers no evidence that journalists are setting the public agenda on nuclear energy technologies. There’s even some suggestive (though unconfirmed) evidence that the direction runs the other way. …Nuclear advocates can’t simply run a media campaign and expect it to build durable public support. Genuine public interest has to be activated first. And this has historically required either crisis (Fukushima, Ukraine energy shock) or sustained policy push (net-zero commitments). The state-versus-independent tone gap also suggests that nuclear’s image problem is concentrated in independent media. Credible independent voices like climate scientists and energy economists endorsing nuclear may carry more weight than state-sponsored campaigns, which audiences are already inclined to discount.
Caveats and Limitations
In the interest of transparency, let me try to be clear about what the analysis can and can’t show.
The GDELT data was sampled every seventh day to manage costs, which means working with monthly aggregates rather than daily precision. While fine for most medium-run dynamics I’m studying, it limits my ability to do precise event studies around specific news shocks.
The study covers only two technologies. The AI-versus-nuclear contrast is suggestive, but confirming whether the media-public relationship differs systematically by technology maturity would require extending the analysis to other domains, such as GMOs, cryptocurrency, autonomous vehicles, facial recognition, etc.
The regulatory events database is hand-coded and covers headline actions (the EU AI Act, US executive orders, major licensing decisions) better than it covers administrative guidance, agency rule-making, or subnational legislation. It’s a lower-bound estimate of regulatory activity, and as a result it’s probably noisier than is optimal.
And of course, the biggest caveat is that this is just observational data. Panel fixed effects can absorb time-invariant differences between technologies… but I can’t rule out time-varying confounders like global economic shocks that could independently drive both media coverage and regulatory activity. A cleaner causal design (instrumental variables, difference-in-differences around exogenous policy shocks) would strengthen the conclusions.
Regarding the agenda-setting question specifically: the cross-lag correlations between media tone and public search interest are symmetric (strong in both directions), so they can’t establish directionality on their own. The Granger causality tests, which can test directionality, find only one surviving result: public search interest predicts AI media tone (not the reverse). The relationship between coverage and attention is probably a mutually reinforcing feedback loop rather than a clean one-way causal arrow. A longer time series, and running the analysis on more technologies, would help clarify whether directional patterns emerge under different conditions.
…However, I’m still fairly confident in the search-interest-predicts-regulation finding. It holds across every specification, in both count and linear models, with growing coefficients at longer lags that match the substantive logic.
Data and methods note: This analysis uses the GDELT Global Knowledge Graph (GKG) hosted on Google BigQuery, Google Trends search indices, and a hand-coded regulatory events database. The study period is February 2015 to November 2025 (130 months). Statistical methods include vector autoregression with Granger causality (BIC-selected lag order, Benjamini-Hochberg FDR correction for 28 pairwise tests), panel regression with entity fixed effects (Poisson GLM and OLS), mediation analysis (Baron & Kenny procedure), and robustness checks across six specifications. All Granger p-values are reported as both raw and FDR-adjusted throughout.
Who Sets the Agenda? (A decade of AI, Nuclear, and the limits of media influence)
Link post
What Actually Makes Governments Regulate Technology?
Many people have an intuition about how technology gets regulated that goes like this: journalists write alarming articles, the public gets scared, politicians respond. It’s assumed that negative press leads to public fear, which leads to protective legislation. Disaster to coverage to regulation feels so intuitive that most people never question it. Yet the data tells a much more interesting and complex story.[1]
Over the past couple of months, I analyzed over a decade of global media coverage for two transformative technologies: artificial intelligence and nuclear energy. I made extensive use of the GDELT database, which tracks hundreds of thousands of news articles worldwide. These database entries were combined with Google Trends search data and a hand-coded database of regulatory events spanning 2015 to 2025. At the end, three findings jumped out, and two of them cast doubt on the typical story.
The findings were:
1. Positive coverage of AI actually predicts more regulation, not less. The most enthusiastic period of AI journalism in the dataset coincided with the sharpest acceleration in regulatory activity. This isn’t what the Fukushima model predicts.
2. The most reliable predictor of when governments actually regulate a technology is neither media tone or coverage volume. It’s actually Google search interest, just the boring act of ordinary people typing things into a search bar. This result held across every statistical specification applied.
3. The one directional relationship the data actually confirms runs backwards from the standard story. Meaning, as public interest in AI grows, media coverage becomes more positive. Journalists appear to follow audience enthusiasm at least as much as they shape it.
Tracking Two Technologies Over One Decade
Why analyze both nuclear energy and AI technologies? Both technologies are high risk technologies, and both technologies are seeing recent surges in funding and development. Comparing the two technologies helps us contrast two different strains of news coverage (positive and negative) to see if news sentiment influences regulation around high risk technologies specifically.
AI coverage over the past decade has been relentlessly positive. The average tone score starts slightly negative in 2015, crosses into positive territory almost immediately, and climbs steadily through 2025.
By the end of the sample, the average AI article is up from roughly −0.2 in early 2015 to over +2 by late 2025. That’s a swing of roughly 1.5 points on a scale where many technologies cluster near zero. There are also a lot of AI articles roughly twelve times more than nuclear energy in a typical month.
Nuclear energy coverage, meanwhile, is a flatline below zero. Persistently negative, seeing no real improvement in over ten years. Even while climate policy has pushed nuclear back into serious consideration for energy technologies, the shadow of events like Fukushima and Chernobyl continues to linger.
The gap in coverage sentiment becomes even more dramatic when you split the data by media type. State-controlled outlets (Xinhua, RT, France24) consistently cover both technologies more positively than independent media. For example, for nuclear energy state outlets average a slightly positive tone (+0.28), while independent outlets average deeply negative (-0.91), a huge gap in sentiment. Notably, governments in France, China, and Russia operate major nuclear programs, and have incentive to frame the technology favorably. Independent journalists are more likely to write what they see, or are otherwise incentivized differently, writing about risk, waste, and accident legacy.
For AI, both state (+1.25) and independent media (+0.72) are positive, with enthusiasm being cross-institutional. State media remains a bit more enthusiastic, consistent with governments positioning AI as a national strategic priority.
The Agenda-Setting Question
The conventional theory of media influence, the agenda-setting theory in the academic literature, says that journalists set the public agenda. Journalists decide what to cover and then the public follows. My own analysis finds something rather more complicated.
AI News Coverage
In terms of AI coverage, it’s tempting to read the data through the classic agenda-setting lens.
When you plot media tone and search interest against each other, you find that AI media tone and public search interest are strongly correlated across the entire decade. Cross-lag correlations are used to test if one of these variables predicts changes in the other better than the reverse, say if media tone better predicts fluctuations in public search interest. The correlations were consistently strong, at time horizons of one to six months (peak r = +0.641 at six months).
However, the cross-correlation function is near-symmetric in both directions. This means the peak correlation alone isn’t a directional finding. Today’s search interest predicts future media tone at almost the same strength as today’s tone predicts future search interest. When a cross-correlation function looks the same in both directions, the most likely explanation is that both are rising together, pulled up by the same underlying force.
In this case, that force is probably a decade of genuine AI breakthroughs, from GPT-3 to AlphaFold to ChatGPT. These discoveries simultaneously generated journalistic coverage and public curiosity, without either one clearly preceding the other.
There’s another reason to doubt that journalists lead the public on this issue: the Granger causality test. While a “cross-lagged correlation” simply shows that two things move together with a delay, a Granger test asks a more rigorous question: “Does knowing the past of Variable A give me new information that helps predict Variable B, even after I already know everything about Variable B’s own history?” The test asks whether one variable predicts the other’s future values beyond what it would predict on its own. The outcome of the test is that AI media coverage doesn’t significantly predict future public search interest (p = 0.905), finding nothing in the journalists-lead-the-public direction.
One result does actually survive rigorous correction for multiple comparisons, but it runs the other way: public search interest predicts more positive subsequent AI media coverage (adjusted p = 0.006). As more people start paying attention to AI, journalists write about it more favorably. The audience appears to shape the framing at least as much as the press shapes the audience.
Nuclear News Coverage
For nuclear energy, there’s actually a suggestive signal that public attention leads coverage rather than follows it. This is consistent with the idea that nuclear interest is activated by external shocks, like energy crises or climate reports, as opposed to journalism campaigns.
…But this result doesn’t survive correction for multiple comparisons (adjusted p = 0.199), so it would be wrong to call it confirmed. Yet there’s no compelling evidence, in this data, that journalists are setting the public agenda on nuclear either.
Basically, for both technologies, media coverage and public attention appear to co-evolve, impacting each other in a mutually reinforcing relationship. The data can’t cleanly establish which comes first. What it can say is that once established, the feedback loop (wherever it comes from) is real and that it feeds directly into the regulatory dynamic described below.
Something worth noting is that the relationship between AI news tone and subsequent search interest has become stronger over time. In the early period (2015-2019), the correlation between AI tone and search interest was modest and not statistically significant (r = +0.187, p = 0.161). In the later period (2020-2025), it roughly doubled and became clearly significant (r = +0.364, p = 0.002).
Media coverage and public attention to AI have become increasingly coupled, each amplifying the other more strongly than they did a decade ago. It’s not clear what the mechanism is. It’s possible that this is journalism gaining agenda-setting power or audiences are just becoming more responsive to coverage or simply a maturing feedback loop, the two are now harder to disentangle than ever.
The Core Finding: Search Interest Predicts Regulation
The finding that seems to hold up the best is this: Higher search interest predicts more regulation.
To test for this, a panel regression was run. A panel regression essentially watches the same group (in this case, a technology) over multiple time periods to see how it changes. The goal is to watch how the relationship between the dependent variables and independent variables change over time, controlling for unobserved factors that don’t change over time. The fixed effects panel regression tested if, after controlling for the nature of the evolving nuclear or AI technologies, media variables predicted future regulatory events.
Three potential predictors were tested: media coverage volume, media tone, and Google search interest.
Coverage volume added only a little by itself. Once you control for search interest, the number of articles about a technology has no independent predictive power for regulation.
Media tone has a marginal, short-lived effect. At the one-month time step, positive media tone shows a weak positive association with regulatory activity (p = 0.063 in the Poisson model). At longer horizons, even this fades. Search interest, on the other hand, is rock-solid as a regulatory predictor.
When you look at Google Trends search interest data, it presents a search index metric that tracks search volume popularity over time and location. For every one-unit increase in Google search index, there was an associated 2.6% increase in regulatory events one month later.
A one-unit increase in the Google Trends search index is associated with a 2.6% increase in expected regulatory events one month later (IRR = 1.026, p = 0.008), 3.1% at two months (IRR = 1.031, p = 0.010), and 8.1% at four months (IRR = 1.081, p < 0.001). In other words, every step up the Google Trends popularity scale correlates with a small bump in regulatory activity, peaking at about 4 months later in a process that gathers steam.
The fact that the effect is most prevalent at longer time horizons makes sense intuitively. Legislative drafting takes time. A spike in public interest doesn’t produce a regulatory event next week, it produces one next quarter. The signal compounds as attention sustains itself.
Six robustness checks were done using a linear model, varying the lag structure and the tone metric. The search interest results held across all of them. with coefficients ranging from +0.020 to +0.036, all significant at p < 0.05. This is about as robust as observational social science gets.
Concretely, public attention is a more reliable predictor of regulatory action than anything journalists write or how they write it.
The Mediation Question: How Does the Chain Work?
So if media tone predicts search interest, which in turn predicts regulation, here’s the next natural question: Does attention mediate the tone-regulation relationship, with positive coverage driving regulation through public attention?
The data pattern is genuinely consistent with this story. Positive media tone strongly predicts public search interest (p < 0.001), and public search interest strongly predicts regulatory events (p < 0.001). When you add search interest as a control, the direct tone-to-regulation path drops from marginal significance (p = 0.053) to clearly non-significant (p = 0.362): a 52% attenuation.
That certainly looks like mediation, but I do want to be careful here. The formal framework for mediation technically requires that the direct path be significant before you add the mediator. In this case the significance before adding the mediator was p = 0.053. This is close, but remains outside the conventional threshold. The attenuation could reflect genuine mediation, or it could just be power loss from adding a correlated predictor to an already-weak effect.
So while the data is consistent with a story where positive coverage → public attention → regulation, they don’t confirm it in the strict statistical sense.
The findings are coherent and replicated in both OLS and Poisson models, and for what it’s worth would be my best guess for the actual mechanism. Still, let’s remain appropriately skeptical.
The ChatGPT Tsunami
The post-ChatGPT period serves as a natural experiment, or at least a vivid illustration, of the dynamics described above.
Before November 2022, AI regulatory events averaged 0.15 per month. After ChatGPT launched, this jumped to 1.00 per month. This coincided with the most positive sustained AI coverage in the entire dataset and the largest spike in public search interest.
This is the whole mechanism we’ve covered (positive coverage fueling public attention, public attention driving regulatory action) compressed into one observable episode. The EU AI Act (agreement on revision to cover general-purpose AI models in December of 2023), the Biden executive order on AI (Nov 2023), and the Bletchley Park summit (Nov 2023) all arrived in the wake of an unprecedented surge in public interest… interest that was itself preceded by overwhelmingly positive media coverage.
Note what didn’t happen: the coverage wasn’t dramatically negative. Some outlets had written “AI may destroy civilization” pieces, but the dominant narrative was enthusiasm and possibility. As has been covered elsewhere, most AI news stories in 2023 were still about corporate use, macrotrends, and scientific research. And regulators responded not by restricting the technology because the public was afraid, but by governing it because the public was paying attention.
What This Means
Two practical implications fall out of this analysis.
For AI governance: If you’re a policymaker trying to anticipate when regulatory pressure will build, watch Google Trends, not headlines. Public search interest is a one-to-three-month leading indicator of regulatory activity, and it’s more reliable than media tone or volume.
Also, don’t assume that positive coverage insulates a technology from regulation. It’s quite possible that positive coverage drives attention which drives governance. The post-2024 shift toward more restrictive AI legislation (coinciding with stabilizing search interest and lower tone volatility), may signal a maturing regulatory cycle where governance becomes institutionalized rather than reactive.
For nuclear energy: The data offers no evidence that journalists are setting the public agenda on nuclear energy technologies. There’s even some suggestive (though unconfirmed) evidence that the direction runs the other way. …Nuclear advocates can’t simply run a media campaign and expect it to build durable public support. Genuine public interest has to be activated first. And this has historically required either crisis (Fukushima, Ukraine energy shock) or sustained policy push (net-zero commitments). The state-versus-independent tone gap also suggests that nuclear’s image problem is concentrated in independent media. Credible independent voices like climate scientists and energy economists endorsing nuclear may carry more weight than state-sponsored campaigns, which audiences are already inclined to discount.
Caveats and Limitations
In the interest of transparency, let me try to be clear about what the analysis can and can’t show.
The GDELT data was sampled every seventh day to manage costs, which means working with monthly aggregates rather than daily precision. While fine for most medium-run dynamics I’m studying, it limits my ability to do precise event studies around specific news shocks.
The study covers only two technologies. The AI-versus-nuclear contrast is suggestive, but confirming whether the media-public relationship differs systematically by technology maturity would require extending the analysis to other domains, such as GMOs, cryptocurrency, autonomous vehicles, facial recognition, etc.
The regulatory events database is hand-coded and covers headline actions (the EU AI Act, US executive orders, major licensing decisions) better than it covers administrative guidance, agency rule-making, or subnational legislation. It’s a lower-bound estimate of regulatory activity, and as a result it’s probably noisier than is optimal.
And of course, the biggest caveat is that this is just observational data. Panel fixed effects can absorb time-invariant differences between technologies… but I can’t rule out time-varying confounders like global economic shocks that could independently drive both media coverage and regulatory activity. A cleaner causal design (instrumental variables, difference-in-differences around exogenous policy shocks) would strengthen the conclusions.
Regarding the agenda-setting question specifically: the cross-lag correlations between media tone and public search interest are symmetric (strong in both directions), so they can’t establish directionality on their own. The Granger causality tests, which can test directionality, find only one surviving result: public search interest predicts AI media tone (not the reverse). The relationship between coverage and attention is probably a mutually reinforcing feedback loop rather than a clean one-way causal arrow. A longer time series, and running the analysis on more technologies, would help clarify whether directional patterns emerge under different conditions.
…However, I’m still fairly confident in the search-interest-predicts-regulation finding. It holds across every specification, in both count and linear models, with growing coefficients at longer lags that match the substantive logic.
Data and methods note: This analysis uses the GDELT Global Knowledge Graph (GKG) hosted on Google BigQuery, Google Trends search indices, and a hand-coded regulatory events database. The study period is February 2015 to November 2025 (130 months). Statistical methods include vector autoregression with Granger causality (BIC-selected lag order, Benjamini-Hochberg FDR correction for 28 pairwise tests), panel regression with entity fixed effects (Poisson GLM and OLS), mediation analysis (Baron & Kenny procedure), and robustness checks across six specifications. All Granger p-values are reported as both raw and FDR-adjusted throughout.
All data and code used for this analysis is available at this GitHub repo, alongside a full in-depth analysis and report.