🔬 physics

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

A study of attendees at a March 2026 Harvard event regarding AI existential risk found that while pre-event views were already high, the structured engagement significantly increased perceived probabilities and confidence in those estimates, particularly among those with little prior familiarity with the topic.

Original authors: Greg Kestin, Nate Soares

Published 2026-03-31

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Greg Kestin, Nate Soares

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you walk into a room where people are discussing a very serious topic: Could super-smart computers eventually take over and end humanity?

This paper is like a "before and after" snapshot of what happened when a group of people attended a specific talk about this scary possibility at Harvard University in March 2026. The researchers wanted to see if hearing a structured, two-sided conversation could change how people felt about the danger.

Here is the story of what they found, explained simply:

1. The Setup: A "Risk Check-In"

Before the event started, the audience (about 180 people) was asked three main questions:

How likely is it that AI will cause human extinction if we don't stop it? (Give a percentage).
How sure are you of that answer?
Should fixing this problem be a top global priority, like fighting pandemics or nuclear war?

Then, they watched a talk by Nate Soares (an AI safety expert) who argued that advanced AI is extremely dangerous. Afterward, they watched a conversation where the host, Greg Kestin, challenged the speaker with counter-arguments and questions. Finally, the audience answered the exact same questions again.

2. The Big Picture: The "Fear Meter" Went Up

Think of the audience's worry level as a thermometer.

Before the talk: The average person thought there was a 50% chance (a coin flip) that AI could end us. They were already pretty worried.
After the talk: The average worry jumped to 70%.
The Priority: Almost everyone (96%) agreed that stopping this risk should be a top global priority after the event.

The Analogy: Imagine a group of people looking at a storm cloud. Before the talk, they thought, "It might rain." After hearing a meteorologist explain how the storm could turn into a hurricane, they all agreed, "We definitely need to build a shelter, and we should take this seriously."

3. The Twist: Newcomers vs. Experts

This is the most interesting part of the study. The researchers looked at how different groups reacted based on how much they already knew about the topic.

The Newcomers (The "Fresh Eyes"):
People who knew very little about AI risk before the event were the most easily swayed.
- 60% of them increased their fear level.
- 0% of them decreased their fear.
- Analogy: It's like showing a movie trailer to someone who has never seen the genre. They leave the theater thinking, "Wow, that looks terrifying!" Their view shifted dramatically because the new information was a revelation to them.
The Experts (The "Veterans"):
People who already considered themselves experts on the topic reacted differently.
- None of them increased their fear.
- 20% actually decreased their fear slightly.
- Analogy: Imagine showing a complex magic trick to a professional magician. Instead of being scared, they might say, "Oh, I see how that trick works; it's not as dangerous as it looks." Because the talk included counter-arguments and technical questions, the experts felt the speaker's case wasn't as airtight as they thought, so they became slightly less worried.

4. Confidence: Feeling More Sure

Not only did people's opinions change, but they also felt more confident in those opinions.

Before the event, many people were guessing.
After the event, even those who didn't change their minds felt more certain about what they believed.
Analogy: It's like walking into a dark room and fumbling around. After someone turns on a light (the talk), you might realize the room is bigger than you thought, or smaller, but you are no longer stumbling in the dark. You feel more sure of your surroundings.

5. Why This Matters

The study shows that one single event can actually change how the public thinks about AI danger.

If you want to raise awareness among people who don't know much, a clear, direct talk works wonders.
If you are talking to experts, you have to be very careful with your arguments, because they might push back and lower their fear if they find holes in your logic.

The Bottom Line

The paper concludes that when you bring people together to discuss AI risks in a structured way, you can move the needle. For the average person walking in with little knowledge, the message landed hard: "This is a real, serious danger, and we need to act."

It's a reminder that how we talk about the future matters. A good conversation can turn a vague worry into a clear, urgent priority.

1. Problem Statement

The paper addresses a gap in the literature regarding how public audiences with varying levels of prior familiarity assess and update their beliefs about AI existential risk (specifically, the probability of human extinction or severe disempowerment due to unimpeded AI development). While expert surveys (e.g., Grace et al., 2024) and general public polls exist, there is limited empirical data on how structured public engagement events influence risk perception, confidence levels, and the prioritization of AI safety as a global issue. The study investigates whether a single, high-profile event featuring a talk and a two-sided conversation can measurably shift these metrics.

2. Methodology

Event Design

Context: A public lecture at Harvard University (March 11, 2026) co-hosted by the Harvard Division of Science, Harvard Book Store, and Harvard Library.
Content:
- Talk: Nate Soares (MIRI President) presented arguments from the book If Anyone Builds It, Everyone Dies regarding AI extinction threats.
- Conversation: A moderated, two-sided dialogue facilitated by Greg Kestin. It included the speaker's arguments alongside counterarguments and alternative perspectives (e.g., potential alignment advantages of Large Language Models trained on human data) gathered from academia and the AI safety community.
- Q&A: A 13-minute audience session.
Format: The event was recorded and made public.

Survey Instrument

Identical pre- and post-event anonymous surveys were administered via QR code/URL to attendees.

Key Questions:
1. Risk Probability: "What probability do you put on future AI advances causing human extinction... (Assume unimpeded development)?" (7 binned responses: <5% to >95%).
2. Confidence: 5-point scale (Not at all confident to Completely confident).
3. Global Priority: Agreement with the statement: "Mitigating the risk of extinction from AI should be a global priority..." (5-point Likert scale).
4. Demographics/Role: Affiliation (Student, Faculty, Industry, etc.).
5. Prior Exposure (Post-event only): Self-reported familiarity with AI risk arguments (Nothing at all to Expert level).

Participants & Analysis

Sample: 180 total attendees; 89 matched participants completed both pre- and post-event surveys.
Demographics: The sample was diverse, including 26% working in scientific fields outside academia, 25% in non-science fields, 15% retired, and 11% self-described experts.
Statistical Methods:
- Shift Analysis: Wilcoxon signed-rank tests to detect overall shifts in probability and confidence.
- Correlation: Spearman rank correlations to assess relationships between prior familiarity and shift magnitude.
- Likelihood Ratios (LR): Used to compare the best-fit hypothesis against the null hypothesis (e.g., LR 3,734:1 indicates the data is 3,734 times more likely under the shift hypothesis than the null).

3. Key Contributions

Empirical Evidence of Belief Updating: Provides quantitative evidence that structured public engagement can significantly shift risk perceptions, particularly among those with low prior familiarity.
Familiarity Gradient: Demonstrates a negative correlation between prior familiarity with the topic and the magnitude of risk perception shifts (newcomers shift more than experts).
Confidence Dynamics: Shows that belief updates are accompanied by increased confidence, suggesting genuine belief revision rather than confusion or uncertainty.
Expert Reversal: Identifies a unique pattern where self-described experts showed a slight downward shift in risk probability, contrasting with the upward shift in the general audience.

4. Key Results

Overall Shifts (n=89)

Risk Probability: The median estimated probability of AI existential risk increased from 50% (pre-event) to 70% (post-event). The mean shifted from 50.5% to 56.9%.
- Statistical Significance: $p < 0.001$ , Likelihood Ratio (LR) 3,734:1.
Confidence: Mean confidence increased from 2.72 to 3.02 (on a 5-point scale). The proportion of "Very" or "Completely" confident respondents rose from 17% to 28%.
Global Priority: Agreement that AI risk mitigation should be a global priority increased from 93% (pre) to 96% (post), with strong agreement rising from 73% to 85%. ( $p = 0.004$ , LR 223:1).

Stratification by Prior Familiarity

The magnitude of the probability shift was negatively correlated with prior familiarity ( $\rho = -0.38$ , $p < 0.001$ ):

Low Familiarity ("A little"): 60% shifted upward, 0% shifted downward. Mean shift: +12 percentage points.
Moderate Familiarity: 39% shifted upward, 10% downward. Mean shift: +10 percentage points.
High Familiarity ("A great deal"): 22% shifted upward, 3% downward. Mean shift: +3 percentage points.
Experts: 0% shifted upward, 20% shifted downward. Mean shift: -4 percentage points.
- Note: The expert downward shift was not statistically significant ( $p=0.25$ ) due to small sample size ( $n=10$ ), but the direction was unique to this group.

Correlations

Probability vs. Confidence: Participants who shifted their risk estimates also reported increased confidence ( $\rho = 0.29$ , $p = 0.006$ ). This suggests the event provided clarity rather than creating uncertainty.
Baseline vs. Shift: Pre-event risk estimates did not predict who would shift; the shift was driven by the event exposure itself.

Comparisons to External Benchmarks

vs. AI Researchers (2023): The pre-event median of 50% was vastly higher than the 5% median found in the 2023 Grace et al. survey of AI researchers. The authors attribute this to self-selection, the "unimpeded" assumption in the question, and the 2+ year gap in AI capability advancements.
vs. General Public/Students: The post-event agreement on global priority (96%) significantly exceeded a 2023 US national poll (70%) and a 2024 Harvard undergraduate survey (~40%).

5. Significance and Limitations

Significance

Public Engagement Efficacy: The study validates that well-structured, two-sided public events can effectively update public understanding of existential risks, especially for newcomers.
Policy Implications: The high level of agreement on global priority (96%) among attendees suggests that informed public engagement could build a strong base of support for AI safety policies.
Expert vs. Public Divergence: The data highlights a potential divergence between self-described experts (who may have nuanced counter-arguments) and the general public (who are more susceptible to upward risk shifts when presented with the core arguments).

Limitations

Self-Selection Bias: Attendees chose an event with a provocative title ("If Anyone Builds It, Everyone Dies"), likely skewing the sample toward those already concerned about AI risk.
Social Desirability: Despite anonymity, participants may have felt pressure to align with the speaker's arguments.
Granularity: The use of binned responses (e.g., 40–60%) limits the precision of detected shifts.
Sample Size: The "Expert" and "Nothing at all" groups were small ( $n=10$ and $n=1$ ), limiting robust statistical inference for these specific subgroups.
Generalizability: Results are specific to this single event, speaker, and audience composition; different formats or speakers might yield different results.

Conclusion

The paper concludes that a single public engagement event produced statistically significant shifts in perceived AI existential risk, confidence, and prioritization. The most profound changes occurred among those with little prior familiarity, while experts showed a tendency to lower their risk estimates, possibly due to deeper engagement with technical counter-arguments. These findings suggest that direct, structured dialogue is a powerful tool for belief updating in the context of AI safety.