Views on AI Existential Risk Before and After a Public Event at Harvard University

Abstract: We report the results of identical pre- and post-event surveys given to attendees of a talk, two-sided conversation, and Q&A centered around the book If Anyone Builds It, Everyone Dies at Harvard University in March 2026, covering perceived probability of AI-caused extinction or severe disempowerment resulting from unimpeded AI development, confidence in those estimates, and global priority. Among the 89 matched participants, the post-event median estimate of the probability of existential risk from advanced AI was 70%, and 96% agreed that mitigating AI existential risk should be a global priority. Although these self-selected respondents' pre-event views were already high (50% and 93%, respectively) relative to results of similar surveys that were previously administered to experts and the general public, the event produced increases on all measures when considering the respondents in aggregate. The magnitudes of increases in risk probability were negatively correlated with prior familiarity with the topic: among attendees with little prior familiarity, 60% shifted upward and none shifted downward, whereas among self-described experts, no respondents shifted upward and 20% shifted downward. Self-reported confidence also increased significantly, and confidence shifts were positively correlated with probability shifts. These findings indicate that a structured public engagement event can meaningfully shift risk perceptions, particularly among newcomers to the topic.

• The paper demonstrates that a public event significantly increased AI existential risk estimates from 50% to 70% (p<0.001) among attendees.
• It employs pre- and post-event surveys with a matched sample (n=89) and nonparametric tests to reveal differential updates based on prior familiarity.
• Results imply that structured public engagement can enhance epistemic certainty and prioritize AI risk mitigation, with experts showing resistance to upward shifts.

Surveying Shifts in Perceptions of AI Existential Risk After a Harvard Public Engagement Event

Introduction

The paper "Views on AI Existential Risk Before and After a Public Event at Harvard University" (2603.27785) provides a rigorous empirical analysis of how structured public engagement shapes risk perception, confidence, and prioritization related to AI existential risk. The methodology leverages pre- and post-event surveys at a university event focused on the existential risk from unimpeded advanced AI development. The survey targeted a diverse audience with varying degrees of prior familiarity, allowing for stratified statistical analysis of belief shifts attributable to the event.

Methodological Framework

A matched sample ($n=89$) of survey respondents provided responses both before and after a talk and dialogue session structured around the book "If Anyone Builds It, Everyone Dies." The survey design included questions on perceived probability of AI-caused extinction or severe disempowerment (binned, scenario-anchored), self-reported confidence, prioritization of AI risk mitigation as a global priority, participant role, and prior familiarity with AI risk arguments. The event format ensured that arguments for and against major AI x-risk theses were presented, and statistical robustness was enhanced by careful analysis of matched pre/post data, binning methodologies, and comprehensive use of non-parametric inference such as Wilcoxon signed-rank tests and likelihood ratios.

Empirical Results

The principal empirical finding is a significant upward shift in perceived AI existential risk among the audience overall ($p<0.001$, LR $3,\!734{:}1$), with a median risk estimate rising from 50% to 70% and the mean from 50.5% to 56.9% post-event. Notably, the increase is highly stratified by prior familiarity: among participants with little prior exposure, 60% increased their probability estimate while none decreased, a monotonic decrease of upward shift corresponds to higher prior familiarity, and among self-reported experts, no respondents shifted upward and 20% shifted downward.

Figure 1: Distribution of probability estimates pre- and post-event, stratified by prior familiarity, showing a monotonic decrease in upward shifts with familiarity and onset of downward shifts among experts.

Self-reported confidence in probability assignments also rose significantly, with mean scale values shifting from 2.72 to 3.02 and a marked increase in high-confidence selections. The positive correlation between probability and confidence shifts (Spearman's $\rho = 0.29$, $p = 0.006$, LR $44{:}1$) implies that movement in estimated probabilities was accompanied by increased epistemic certainty.

Prioritization of AI x-risk mitigation began at very high levels (93% agreement pre-event) and strengthened further post-event (96% agreement, $p = 0.004$, LR $223{:}1$).

Figure 2: Direction and magnitude of individual response shifts for probability, confidence, and prioritization questions aggregated by prior familiarity; downward shifts appear predominantly for experts, upward and static shifts dominate for less familiar groups.

Comparisons with external reference samples reinforce the anomalous baseline risk perception of the event audience: the analogous expert survey in 2023 (Grace et al., 2024) yielded median and mean extinction estimates of 5% and 16.2%, respectively, whereas the Harvard audience started much higher. The prioritization question responses also exceed national and student baseline agreement rates.

Theoretical and Practical Implications

The paper demonstrates that structured public engagement can reliably move Bayesian credences and prioritization judgments in high-stakes AI domains, especially among less knowledgeable or uncommitted participants. The gradient of effect by prior familiarity underscores that established domain experts are robust to single-session interventions and may even update downward when presented with balanced or alternative arguments. This suggests a potential boundary between information-driven belief revision and prior commitment/epistemic insulation in AI governance debates, a finding with direct relevance for communication strategies targeting differing demographics.

From a methodological perspective, the use of kernel density estimation to visualize distributional shifts and the computation of likelihood ratios for effect size reporting establish best practices for future x-risk survey research.

Practically, the results imply that targeted engagement venues can meaningfully influence risk perception, informing policy advocacy and public communication planning. The effect appears constrained by self-selection bias and may not generalize to genuinely naive or adversarial populations.

Limitations and Future Directions

The limitations—acknowledged in detail by the authors—include severe self-selection bias (event focus, voluntary attendance), limited sample granularity (binned probability reports), and lack of robust population representativeness. The size of the expert group ($n=10$) constrains statistical power for subgroup analysis. The survey instruments, by design, cannot fully decouple social desirability effects or demand characteristics. Cross-survey comparison is further complicated by changes in real-world AI capabilities and shifting public discourse.

To advance the field, future research should explore repeated measures across diverse populations, randomized controlled exposure to information interventions, and longitudinal tracking of risk perception trajectories. Deeper qualitative analysis of reasoning processes that underlie belief shifts, especially for expert audiences, remains a crucial open problem. Incorporation of more granular probabilistic elicitation techniques could further sharpen resolution.

Conclusion

The study provides statistically robust evidence that public events can drive upward movement in perceived AI existential risk probabilities and associated confidence, with effect sizes strongly modulated by prior familiarity. Expert audiences display resistance or even mild counter-updating, while newcomers are highly susceptible to event framing. These findings deliver actionable insights for both AI risk communication and the broader empirical study of existential risk cognition.