Hugging Face Adds "P(doom)" to User Profiles: Anonymized Community Survey on AI Catastrophic Risk
Hugging Face introduced an optional P(doom) probability indicator on developer profile cards. The setting allows machine learning researchers to publicly share their estimated probability of catastrophic AI risk while aggregating data into a public consensus dashboard.
Hugging Face rolled out a feature on October 6, 2026, allowing developers to display their estimated P(doom)—the subjective probability that artificial general intelligence will cause human extinction or irreversible catastrophe—on their user profile cards.
The feature pairs personal profile customization with a large-scale, ongoing survey of the global machine learning engineering community.
What Is P(doom) and How the Profile Integration Functions
Originally coined in AI alignment forums and academic discussions, P(doom) represents a subjective Bayesian probability ranging from 0% (zero catastrophic risk) to 100% (inevitable doom).
Under Hugging Face Account Settings > Profile & Visibility, users now have access to a dedicated slider:
┌────────────────────────────────────────────────────────────────────────┐
│ Hugging Face P(doom) Configuration │
├────────────────────────────────────────────────────────────────────────┤
│ Personal P(doom) Estimate: [───●────────────────────────] 15% │
│ │
│ Visibility Options: │
│ [x] Display on public profile card (e.g., "@username · P(doom): 15%") │
│ [x] Include in anonymous research data pool │
│ │
│ Research Demographics (Optional): │
│ • Primary Domain: Alignment / Interpretability / Pretraining / Infra │
│ • Hardware Experience: <1 Year / 1-4 Years / 5+ Years │
│ • Employer Type: Frontier Lab / University / Independent Open Source │
└────────────────────────────────────────────────────────────────────────┘
Initial Survey Findings: Median Sits at 14.5%
Within 24 hours of release, more than 42,000 developers and researchers configured their P(doom) values. The aggregated dataset reveals notable distribution patterns across developer sub-groups:
| Research Cohort | Sample Size | Median P(doom) | Interquartile Range |
|---|---|---|---|
| All Participants | 42,100 | 14.5% | 5% – 35% |
| Mechanistic Interpretability Researchers | 3,400 | 32.0% | 15% – 60% |
| Distributed Systems & Infra Engineers | 9,800 | 4.2% | 1% – 10% |
| University Graduate Students (CS/AI) | 14,200 | 18.0% | 5% – 40% |
| Commercial Frontier Lab Engineers | 4,600 | 21.5% | 10% – 50% |
Engineers working directly on hardware infrastructure and CUDA kernels report the lowest catastrophic risk estimates (median 4.2%), while researchers studying interpretability and test-time reasoning loops report the highest (median 32.0%).
Open Research Data Endpoint
To assist academic sociologists and technology policy scholars, Hugging Face established an open dataset and API endpoint:
# Download real-time anonymized survey histogram data
curl -s "https://huggingface.co/api/surveys/pdoom/distribution" | jq .
The feature demonstrates how open-source collaboration hubs can turn theoretical philosophical debates into empirical, community-wide statistical data.