Within Survey Estimates
Why Different Expert Groups Disagree on AI Doom Probabilities
Comparing how general AI researchers and specialist safety experts assign different probabilities to existential AI risk.
On this page
- Broad AI research community trends
- Specialist safety and governance group estimates
- Drivers of disagreement and variance
Page outline Jump by section
Introduction
One of the clearest findings from AI doom surveys is not that experts agree on a particular probability of catastrophe, but that different expert populations consistently give different answers. Researchers who work on mainstream machine learning, researchers focused on AI safety and alignment, governance specialists, forecasters, and sceptics often look at the same underlying technology and arrive at sharply different estimates of existential risk.
This matters because debates about p(doom) are often presented as disagreements about a single number. In reality, they are frequently disagreements between communities with different training, incentives, conceptual frameworks, and assumptions about future AI systems. Survey evidence repeatedly shows that who gets asked can substantially affect the resulting forecast. Understanding those population differences helps explain why published estimates range from near zero to well above 50%, even among people with technical expertise. [AI Impacts Wiki]wiki.aiimpacts.orgai risk surveysAI Impacts WikiSurveys of experts on levels of AI Risk9 May 2023 — We know of six surveys of AI experts and two surveys of AI safety/gove…
Broad AI Research Communities Tend to Give Lower Risk Estimates
The largest recent evidence comes from AI Impacts’ 2023 survey of 2,778 researchers who had published at major AI conferences and journals. The survey sampled a broad cross-section of active AI researchers rather than focusing only on people working on existential risk or alignment. [AI Impacts]youtube.comAI Impacts Survey - The key implications, with Katja Grace…
Even within this wider population, concern about catastrophic outcomes was not rare. Depending on the wording used, the median respondent assigned around a 5% probability to outcomes comparable to human extinction or permanent severe disempowerment from advanced AI. Large minorities assigned probabilities of 10% or higher. AI Impacts Blog [JAIR]jair.orgThousands of AI Authors on the Future of AIby K Grace · 2025 · Cited by 195 — Depending on how we asked, between 38% and 51% of responden…
However, these headline figures can create a misleading impression if readers assume that all AI researchers think similarly. The distribution was extremely wide. Some respondents placed risk near zero, while others assigned very high probabilities. The average result therefore reflected a mixture of very different worldviews rather than a coherent consensus. [AI Impacts]youtube.comAI Impacts Survey - The key implications, with Katja Grace…
Broad AI-research samples also include many people whose primary work involves areas such as computer vision, natural language processing, robotics, recommendation systems, optimisation, or theory rather than long-term AI safety. For many of these researchers, existential-risk questions are not central to their daily work. Surveys suggest that familiarity with specialised AI-safety concepts varies considerably across the broader research community. [arXiv]arxiv.orgSource details in endnotes.
That distinction becomes important when comparing their responses with those of specialist populations.
Safety and Alignment Specialists Often Report Higher p(doom)
Surveys focused on AI-safety researchers and related experts generally find higher levels of concern than surveys of the broader AI field. AI Impacts’ survey catalogue notes a recurring pattern: populations specifically selected for safety or governance expertise tend to produce higher existential-risk estimates than general AI-research samples. [AI Impacts Wiki]wiki.aiimpacts.orgai risk surveysAI Impacts WikiSurveys of experts on levels of AI Risk9 May 2023 — We know of six surveys of AI experts and two surveys of AI safety/gove…
One reason is straightforward selection. People who choose to spend their careers on alignment, control problems, interpretability, or catastrophic-risk governance often do so because they already believe those issues are important. But selection alone does not explain the whole gap.
A 2025 survey examining disagreement among 111 AI experts found evidence that experts clustered into two broad perspectives. One group tended to view future AI primarily as a controllable tool. The other was more likely to view sufficiently advanced AI as a potentially autonomous and difficult-to-control agent. Those perspectives were strongly associated with different p(doom) estimates. [arXiv]arxiv.orgSource details in endnotes.
The study also found differences in familiarity with key safety concepts. Researchers who were less concerned about existential risk were often less familiar with ideas such as instrumental convergence, goal misgeneralisation, and other concepts frequently discussed within alignment research. The authors argued that disagreement was not simply a dispute about probabilities but reflected different conceptual models of what advanced AI might become. [arXiv]arxiv.orgSource details in endnotes.
From the perspective of many safety specialists, current machine-learning systems should not be treated merely as software tools becoming incrementally more capable. Instead, they are viewed as early examples of systems that could eventually develop strategic behaviour, pursue objectives in unintended ways, or become difficult to supervise effectively at scale. Those assumptions naturally produce higher doom forecasts. [arXiv]arxiv.orgSource details in endnotes.
Governance and Long-Term Risk Communities Occupy a Different Position
A third population often appears in existential-risk discussions: researchers and policy specialists whose work focuses on long-term societal risks, governance failures, international coordination, or technological forecasting.
These groups are not always machine-learning researchers themselves, but they frequently engage with scenarios involving loss of control, racing dynamics between companies or states, and failures of global coordination. Their forecasts can differ from both mainstream AI researchers and technical alignment specialists.
In some cases, governance-focused experts assign relatively high probabilities to catastrophic outcomes even when they are uncertain about specific technical alignment failures. Their concern comes instead from social mechanisms: competitive deployment pressures, military incentives, inadequate monitoring, or the possibility that powerful systems are released before safety measures mature. [80,000 Hours]80000hours.orgHere's why working on AI risks could be the most important use80,000 HoursWhy AI risks are the world's most pressing problemsFeb 24, 2026 — AGI could rapidly transform the world, posing existential r…
This helps explain why two experts may agree on the technical capabilities of future AI systems but still assign different p(doom) values. One may view governance failures as the dominant threat, while another views technical control as the primary challenge.
The Strongest Predictor May Be Underlying Worldview
The most striking pattern across surveys is that population differences often track deeper assumptions about AI itself.
Researchers who expect gradual progress frequently assign lower probabilities to doom. If capabilities advance incrementally, institutions have more time to adapt, governments can regulate more effectively, and safety techniques can improve alongside deployment.
Researchers who expect rapid capability jumps or transformative AI arriving on short timelines tend to assign higher risks. In those models, society may face systems with novel strategic abilities before robust control methods exist. [AI Impacts]youtube.comAI Impacts Survey - The key implications, with Katja Grace…
The disagreement therefore runs deeper than a simple argument over percentages. Different expert groups often start from different answers to questions such as:
- Will future AI remain tool-like or become agentic?
- Will capabilities scale smoothly or discontinuously?
- Can alignment methods keep pace with capability advances?
- Will governments and companies coordinate effectively?
- Are current AI systems evidence for or against future loss-of-control scenarios?
Changes in these assumptions can shift p(doom) forecasts dramatically, even among researchers with similar technical backgrounds. [arXiv]arxiv.orgSource details in endnotes.
Why Population Differences Matter for Interpreting Surveys
Readers often encounter headlines suggesting that “experts believe there is a 5% chance of extinction” or that “AI researchers fear doom.” Such summaries can obscure the most informative finding: there is no single expert view.
A survey of thousands of conference-publishing AI researchers captures one population. A survey of alignment researchers captures another. A survey of governance specialists captures a third. Each group is observing different evidence, using different models, and asking different questions about what future AI systems will be like. [AI Impacts]youtube.comAI Impacts Survey - The key implications, with Katja Grace… [AI Impacts Wiki]wiki.aiimpacts.orgai risk surveysAI Impacts WikiSurveys of experts on levels of AI Risk9 May 2023 — We know of six surveys of AI experts and two surveys of AI safety/gove…
This does not mean expert surveys are useless. On the contrary, the population differences are themselves evidence. They reveal where uncertainty is concentrated and which assumptions drive disagreement. When specialist safety communities consistently produce higher estimates than broader technical populations, that gap becomes an important fact about the state of the debate.
The surveys therefore tell us less about a precise probability of extinction and more about the structure of expert disagreement. The central lesson is not that AI doom has been measured, but that experts who spend their time thinking about advanced AI often diverge in systematic ways depending on which community they belong to, what models they use, and what future they believe they are trying to forecast. [AI Impacts Wiki]wiki.aiimpacts.orgai risk surveysAI Impacts WikiSurveys of experts on levels of AI Risk9 May 2023 — We know of six surveys of AI experts and two surveys of AI safety/gove… 2arXiv
Amazon book picks
Further Reading
Books and field guides related to Why Different Expert Groups Disagree on AI Doom Probabilities. Use these as the next step if you want deeper reading beyond the article.
Superforecasting
Explains why expert groups differ and how forecasting performance varies.
Expert Political Judgment
Directly relevant to differences between expert communities.
eBay marketplace picks
Marketplace Samples
Example marketplace items related to this page. Use the search link to explore similar finds on eBay.
Endnotes
-
Source: arxiv.org
Link: https://arxiv.org/abs/2502.14870 -
Source: arxiv.org
Link: https://arxiv.org/pdf/2401.02843Source snippet
arXivThousands of AI Authors on the Future of AIWe conducted a survey of 2,778 AI researchers who had published peer-reviewed research in...
-
Source: jair.org
Link: https://www.jair.org/index.php/jair/article/view/19087Source snippet
Thousands of AI Authors on the Future of AIby K Grace · 2025 · Cited by 195 — Depending on how we asked, between 38% and 51% of responden...
-
Source: arxiv.org
Link: https://arxiv.org/html/2401.02843v1Source snippet
Thousands of AI Authors on the Future of AI5 Jan 2024 — While AI experts' predictions should not be seen as a reliable guide to objective...
-
Source: arxiv.org
Link: https://arxiv.org/html/2502.14870v1Source snippet
arXivWhy do Experts Disagree on Existential Risk and P(doom)...Feb 23, 2025 — The goal of the survey is to understand what drives this m...
-
Source: 80000hours.org
Title: Here’s why working on AI risks could be the most important use
Link: [https://80000hours.org/problem-profiles/artificialSource snippet
80,000 HoursWhy AI risks are the world's most pressing problemsFeb 24, 2026 — AGI could rapidly transform the world, posing existential r...
-
Source: arxiv.org
Link: https://arxiv.org/abs/2401.02843Source snippet
[2401.02843] Thousands of AI Authors on the Future of AIby K Grace · 2024 · Cited by 208 — If science continues undisrupted, the chance o...
-
Source: arxiv.org
Link: https://arxiv.org/pdf/2503.07341Source snippet
arXivThe Economics of p(doom): Scenarios of Existential Risk...by J Growiec · 2025 · Cited by 6 — Field (2025) shows that experts' disag...
-
Source: arxiv.org
Link: https://arxiv.org/pdf/2603.27785Source snippet
Views on AI Existential Risk Before and After a Public...by G Kestin · 2026 — We report the results of identical pre- and post-event sur...
-
Source: 80000hours.org
Link: https://80000hours.org/2024/09/why-experts-and-forecasters-disagree-about-ai-risk/Source snippet
Why experts and forecasters disagree about AI riskSep 16, 2024 — Domain experts in AI estimated a 3% chance of AI-caused human extinction...
-
Source: youtube.com
Link: https://www.youtube.com/watch?v=hliLDNdxkX0Source snippet
AI Impacts Survey - The key implications, with Katja Grace...
-
Source: youtube.com
Title: AI Impacts Survey
Link: https://www.youtube.com/watch?v=xwJx_xqZI3QSource snippet
How Worried Should We Be About AI? - Katja Grace, AI Impacts - DS Pod #290...
-
Source: wiki.aiimpacts.org
Title: ai risk surveys
Link: https://wiki.aiimpacts.org/uncategorized/ai_risk_surveysSource snippet
AI Impacts WikiSurveys of experts on levels of AI Risk9 May 2023 — We know of six surveys of AI experts and two surveys of AI safety/gove...
Published: May 2023
-
Source: aiimpacts.org
Link: https://aiimpacts.org/wp-content/uploads/2023/04/Thousands_of_AI_authors_on_the_future_of_AI.pdfSource snippet
AI ImpactsTHOUSANDS OF AI AUTHORS ON THE FUTURE OF AIby K Grace · 2024 · Cited by 195 — We conducted a survey of 2,778 AI researchers who...
-
Source: blog.aiimpacts.org
Title: 2023 ai survey of 2778 six things
Link: https://blog.aiimpacts.org/p/2023-ai-survey-of-2778-six-thingsSource snippet
AI Impacts BlogSurvey of 2778 AI authors: six parts in picturesJan 4, 2024 — Median respondents put 5% or more on advanced AI leading to...
-
Source: blog.aiimpacts.org
Link: https://blog.aiimpacts.org/p/faq-expert-survey-on-progress-inSource snippet
AI Impacts BlogFAQ: [Expert Survey]({{ 'survey-estimates/' | relative_url }}) on Progress in AI methodologyOct 31, 2025 — These four question variations all got similar answers—in 2...
-
Source: wiki.aiimpacts.org
Title: 2023 expert survey on progress in ai
Link: https://wiki.aiimpacts.org/ai_timelines/predictions_of_human-level_ai_timelines/ai_timeline_surveys/2023_expert_survey_on_progress_in_aiSource snippet
aiimpacts.org2023 Expert Survey on Progress in AI17 Aug 2023 — The 2023 Expert Survey on Progress in AI is a survey of 2,778 AI researche...
-
Source: blog.aiimpacts.org
Title: how bad a future do ml researchers
Link: https://blog.aiimpacts.org/p/how-bad-a-future-do-ml-researchersSource snippet
bad a future do ML researchers expect?May 10, 2023 — For these, the median researcher also gave 5% and 10% answers...
Published: May 10, 2023
-
Source: lesswrong.com
Title: ai impacts 2023 expert survey on progress in ai
Link: https://www.lesswrong.com/posts/RkegCmCgjGhskiFvm/ai-impacts-2023-expert-survey-on-progress-in-aiSource snippet
5 Jan 2024 — AI Impacts just released the results from their latest survey! I've found this quite interesting to engage with. Here is the...
-
Source: lesswrong.com
Title: ai impacts survey december 2023 edition
Link: https://www.lesswrong.com/posts/NfPxAp5uwgZugwovY/ai-impacts-survey-december-2023-editionSource snippet
AI Impacts Survey: December 2023 EditionJan 5, 2024 — In Figure 13's question 3, we have 14.4% mean chance of either human extinction or...
Published: december 2023
-
Source: Wikipedia
Title: AI safety
Link: https://en.wikipedia.org/wiki/AI_safetySource snippet
AI safetyAI safety is an interdisciplinary field focused on preventing accidents, [misuse]({{ 'misuse/' | relative_url }}), or other harmful consequences arising from a...
Additional References
-
Source: researchgate.net
Link: https://www.researchgate.net/publication/369063799_AI_Risk_Skepticism_A_Comprehensive_SurveySource snippet
(PDF) AI Risk Skepticism, A Comprehensive SurveyIn this thorough study, we took a closer look at the skepticism that has arisen with resp...
-
Source: linkedin.com
Link: https://www.linkedin.com/posts/adam-chalmers-91001413a_ai-artificialintelligence-aisafety-activity-7302592505810743296-KSppSource snippet
AI risks: Experts' estimates vary widely | Adam Chalmers...3 Mar 2025 — What are the odds that AI leads to catastrophic outcomes for hum...
-
Source: linkedin.com
Link: https://www.linkedin.com/posts/iq-com_why-do-experts-disagree-on-existential-risk-activity-7356043558677630976-pBviSource snippet
AI researchers divided on existential risk: controllable tool or...A survey of 111 AI researchers reveals deep disagreement on existenti...
-
Source: imdb.com
Link: https://www.imdb.com/title/tt38462048/Source snippet
A Survey of AI Experts. Podcast Episode; 2025; 3m. YOUR RATING. Rate. Add a plot in your language.Read more...
-
Source: cset.georgetown.edu
Title: beyond pdoom for ai risk quantifying uncertainty without probability
Link: https://cset.georgetown.edu/publication/beyond-pdoom-for-ai-risk-quantifying-uncertainty-without-probability/Source snippet
P(doom) for AI Risk: Quantifying Uncertainty...As artificial intelligence introduces new risks, some potentially catastrophic or even ex...
-
Source: linkedin.com
Link: https://www.linkedin.com/posts/simontorrance_why-do-experts-disagree-on-existential-risk-activity-7301886948334333952-VwQSSource snippet
"AI as controllable tool" and an "AI as uncontrollable agent" - diverging in beliefs...Read more...
-
Source: chatpaper.com
Link: https://chatpaper.com/zh-CN/chatpaper/paper/110361Source snippet
A Survey of AI Experts. cs.AI23 Feb 2025. Severin Field. Cambridge ERA:AI Fellowship.Read more...
-
Source: facebook.com
Title: In the largest survey yet of AI researchers, a majority say
Link: https://www.facebook.com/newscientist/posts/in-the-largest-survey-yet-of-ai-researchers-a-majority-say-there-is-a-non-trivia/763238515844768/Source snippet
AI advances causing human extinction or similarly permanent and severe disempowerment of the human species?” For this, the median answer...
-
Source: semanticscholar.org
Title: Why do Experts Disagree on Existential Risk and P(doom)?
Link: https://www.semanticscholar.org/paper/Why-do-Experts-Disagree-on-Existential-Risk-and-A-Field/38f774f51534dd13cea1137c4a347b046f741b66Source snippet
do Experts Disagree on Existential Risk and P(doom)? A Survey of AI Experts · Severin Field · Published in arXiv.org 25 January...
-
Source: researchgate.net
Title: 396256646 Thousands of AI Authors on the Future of AI
Link: https://www.researchgate.net/publication/396256646_Thousands_of_AI_Authors_on_the_Future_of_AISource snippet
Thousands of AI Authors on the Future of AIOct 5, 2025 — PDF | In October 2023, 2778 researchers who had published in top-tier artificial...
Published: October 2023
Topic Tree