This is an educational and exploratory tool. These estimates are deeply uncertain. Experts range from <1% to >50% P(doom). Use this to structure your thinking, not as a definitive answer.
Risk Factor Assessment
P(Superintelligence by 2100)
Expert estimates range from <5% (LeCun) to >90% (Tegmark). Median AI researcher estimate (Grace et al. 2024): ~10% chance of AGI by 2027, ~50% by 2047. Consider both technical feasibility and economic incentives.
P(Alignment failure | AGI)
The alignment problem: ensuring AI systems pursue intended goals. Estimates range from ~2% (Altman) to >90% (Yudkowsky). Key uncertainty: whether we can specify human values precisely enough for superhuman systems.
P(Catastrophic | Misalignment)
If a misaligned AGI exists, how likely is it to cause civilisation-level harm? Optimists: 5-20%. Pessimists: 80-99%. Key factors: whether AGI can be contained, whether it has physical-world capabilities, and goal preservation under self-improvement.
P(Malicious Use)
Probability of deliberate catastrophic misuse of AI: bioweapons, autonomous weapons, cyberattacks on infrastructure. Estimates: 2-15% (cautious optimists) to 30-50% (risk-focused researchers). Depends on access controls and regulation.
P(Governance succeeds)
Current state: voluntary commitments (Frontier Model Forum), EU AI Act, US executive orders. Key challenges: enforcement, racing dynamics between nations, and verification of compliance.
What's driving your estimate
Each bar shows how much a ±10 percentage point change in that slider moves your P(doom), relative to the other factors.
Where your estimate sits among key researchers
Expert P(doom) Estimates
Why estimates vary so much: Researchers use different definitions. Some estimate extinction, others civilisational collapse or permanent disempowerment. Some give conditional probabilities (given AGI by 2100), others unconditional. Timeframes range from 30 years to never specified. These are subjective probability estimates, not scientific measurements.
Color coding: red >20%, yellow 1–20%, green <1%.
| Person | Role | P(doom) | Source | Year |
|---|---|---|---|---|
| Yann LeCun | Meta AI Chief Scientist | <0.01% | Twitter/X Jul 2024 | 2024 |
| Sam Altman | OpenAI CEO | >0%, "low but non-zero" | diverse public interviews | 2023 |
| Geoffrey Hinton | 2024 Nobel Laureate, ex-Google | 10–20% extinction/30yr; up to 50% surpassing humanity | BBC Radio 4 Dec 2024 | 2024 |
| Yoshua Bengio | Mila, Turing Award | ~20% | UN AI Safety Summit 2024 | 2024 |
| Dario Amodei | Anthropic CEO | 10–25% | Interview Oct 2023 | 2023 |
| Paul Christiano | US AI Safety Institute | ~50% | 80,000 Hours | 2023 |
| Elon Musk | xAI / Tesla | ~10–20% | "I probably agree with Hinton" | 2024 |
| Eliezer Yudkowsky | MIRI Founder | >95% | Time Magazine | 2023 |
| Roman Yampolskiy | AI Researcher | ~99.9% | diverse interviews | 2023–2024 |
| Dan Hendrycks | Center for AI Safety | >80% | public statements | 2023–2024 |
| Andrew Critch | CARES Founder | ~85% | public statements | 2023 |
| Max Tegmark | MIT / Future of Life Institute | >90% | public statements | 2023 |
| Daniel Kokotajlo | AI Futures Project, ex-OpenAI | ~70% | Time100 AI 2024 | 2024 |
| Jan Leike | ex-Anthropic Alignment | 10–90% | public statements | 2024 |
| Shane Legg | DeepMind Co-founder | ~5–50% | public statements | 2023 |
| Emmett Shear | ex-OpenAI interim CEO | ~5–50% | public statements | 2023 |
| Scott Alexander | Astral Codex Ten | ~25% | blog posts | 2023 |
| Toby Ord | FHI Oxford | ~10% | The Precipice, 2020 | 2020 |
| Vitalik Buterin | Ethereum Founder | ~10% | public statements | 2024 |
| Lina Khan | ex-FTC Chair | ~15% | AEI article Oct 2025 | 2025 |
| Metaculus Community | Forecasting platform | ~3% AI-specific | metaculus.com | 2026 |
| Demis Hassabis | Google DeepMind CEO | >0%, "notable risk" | public interviews | 2023 |
| Benjamin Mann | Anthropic Co-founder | 0–10% | public statements | 2023 |
| Andrew Ng | DeepLearning.AI | very low | public statements | 2023 |
Estimates reflect the researcher's publicly stated view at the time of the source. P(doom) is not a standardised metric — definitions of 'doom', timeframes, and conditions vary by researcher.
Last verified:
What surveys show
| Survey | Year | n | Population | Median P(doom) | Definition |
|---|---|---|---|---|---|
| AI Impacts 2022 Expert Survey | 2022 | 738 | ML researchers (ICML/NeurIPS) | ~5% | Human extinction or permanent severe curtailment by AI |
| Grace et al. 2024 (arXiv:2401.02843) | 2024 | 2,778 | Top-AI conference authors | ~5% (median), 14.4% (mean) | Extinction or permanent disempowerment within 100 years |
| Metaculus Community | ongoing | >1,000 forecasters | Forecasting community | ~3% (AI-specific) | Human extinction by 2100 |
Source for Metaculus: Growiec & Prettner, "The Economics of p(doom)", arXiv:2503.07341, March 2026.
The survey median (~5%) is substantially lower than the median of public expert statements (~20%). This gap reflects selection bias: researchers who make public statements about P(doom) are not a random sample of all AI researchers.
P(doom) Calculator
Global catastrophe probability
AI Alignment Risk
Technical failure probability
Explore →AI Job Displacement
Your personal career risk
Explore →A doom calculator estimates the probability of catastrophic outcomes from advanced AI systems. In AI safety research, this probability is called P(doom) — the likelihood that artificial general intelligence (AGI) causes permanent, large-scale harm to humanity.
Unlike most calculators, there is no single formula. P(doom) is a subjective probability that depends on your assumptions about AI development timelines, alignment difficulty, and governance effectiveness. This doom calculator walks you through each assumption and combines them into a personal estimate.
Published estimates range from near 0% (Yann LeCun) to over 90% (Eliezer Yudkowsky). The median across 50+ public estimates is approximately 20%. See the full AI Risk Survey 2026 →
Yudkowsky
MIRI founder — most pessimistic major researcher
Hinton
10–20% extinction/30yr (BBC Dec 2024); up to 50% AI surpassing humanity (BNN Jun 2024)
Median researcher
2022 AI researcher survey median P(doom)
LeCun
Yann LeCun (Meta) — most optimistic major researcher
The 100× Disagreement Matters
When experts disagree by two orders of magnitude (<1% vs >99%), the takeaway is not that one side is right — it's that we face genuine uncertainty about unsolved problems. P(doom) is a tool for structuring uncertainty, not a prediction. The most rational response is to support AI safety research while acknowledging the limits of our current knowledge.
Why Experts Disagree by 100×
P(doom) estimates range from <1% to >99% because they depend on assumptions about unsolved open problems: whether the alignment problem is tractable, how quickly AI capabilities will advance, and whether governance mechanisms can keep pace with technology.
These are not empirical disagreements resolvable by more data — they are partly philosophical judgements about the nature of intelligence, human institutions, and the behaviour of optimisation processes at scale.
Educational tool: This calculator structures your uncertainty — it does not produce a ground truth. P(doom) is not a scientific metric with a knowable true value. Use it to explore your own assumptions.
Before you set P(AGI): Every major AI timeline prediction from the past 60 years was wrong — from Simon (1958) to Yudkowsky (2021). This systematic over-optimism matters for your estimate. Read our analysis of why forecasts keep failing →
What Drives Estimates Up or Down
Factors that increase P(doom)
Rapid capability scaling outpacing safety, race dynamics between labs or nations, difficulty of the alignment problem, inadequate governance, and dual-use misuse potential of powerful models.
Factors that decrease P(doom)
Capability plateau before dangerous levels, alignment proving tractable, international cooperation, significant lab safety investment, democratic oversight, and narrow AI remaining dominant longer than expected.
What you can do
Engage with AI governance debates. Support responsible AI institutions (ARC, MIRI, Anthropic safety teams, Centre for AI Safety). Stay informed. The civic response is more impactful than individual behaviour changes.
Frequently Asked Questions
Formula & Sources
Bayesian Decomposition of P(doom)
P(doom) = P(AGI) × P(misalign|AGI) × P(catastrophe|misalign) + P(misuse) − overlapP(AGI)— Probability of artificial general intelligence by 2100P(misalign|AGI)— Probability of value-misalignment given AGIP(catastrophe|misalign)— Probability of irreversible catastrophe given misaligned AGIP(misuse)— Probability of catastrophic deliberate misuse (bioweapons, etc.)
Framework adapted from MacAskill, 'What We Owe The Future' (2022); Ord, 'The Precipice' (2020)
"I think it's quite likely — like 10 to 20 percent — that we'll end up with a world that is really bad for humans."
Sources & References
- Ord, T. (2020): The Precipice. Hachette Books. P(doom from unaligned AI) ≈ 10% this century.
- MacAskill, W. (2022): What We Owe The Future. Basic Books. → whatweowethefuture.com
- Grace, K. et al. (2024): "Thousands of AI authors on the future of AI." arXiv:2401.02843. Survey of 2,778 AI researchers: median P(extremely-bad outcome) ≈ 5%. → arXiv
- AI Impacts (2022): "2022 Expert Survey on Progress in AI." 738 ML researchers. Median P(extinction or severe curtailment) ≈ 5%. → AI Impacts
- Metaculus (2026): Community forecasts on AGI timelines and existential risk. ~3% AI-specific extinction by 2100. → Metaculus
- Bostrom, N. (2014): Superintelligence: Paths, Dangers, Strategies. Oxford University Press.
- Hinton, G. (2024): BBC Radio 4 (Dec 2024): 10–20% extinction within 30 years. BNN Bloomberg (Jun 2024): up to 50% AI surpassing humanity.
- Yudkowsky, E. (2023): MIRI. P(doom) > 95%. → Time Magazine
Why estimates differ by 10,000×
- Different definitions of 'doom': Extinction vs. permanent disempowerment vs. civilisational collapse.
- Different timeframes: Some "within 30 years" (Hinton), others "within 100 years" (Ord, AI Impacts), others unconditional.
- Conditional vs. unconditional: P(doom | AGI is built) vs. unconditional P(doom) — mathematically different numbers.
- Selection bias: Survey medians (~5%) are consistently lower than public statements (~20%).
- Alignment difficulty: The core dispute — whether aligning capable AI is tractable — explains most of the variance.
Related Tools
Estimate probability of AI technical failure and unaligned outcomes.
Estimate your role's automation risk and financial runway.
Calculate your emergency preparedness budget.
How long to earn a billionaire's wealth at your salary?
Why expert forecasts keep failing — and what it means for P(doom).
For informational purposes only — not financial, medical, or legal advice. Results are estimates; use at your own risk. Full terms