Rogue AI takes over
An AI system outside human control takes over critical systems across at least one large country and acts against human interests.
Your chatbot, but it stopped asking permission.
Combined score
8.5
Remote
Average of 4 AI models, as the chance this happens within 25 years. Soonest window most models chose: 5-25 years.
Last updated .
Independent AI panel
Claude (Anthropic)
4
Remote | soonest: 5-25 years
- Requires an AI outside human control to seize critical systems across a large country and act against people, a severe and specific outcome.
- Expert surveys put AI-caused catastrophe in this range at low single digits, though estimates vary widely and some are higher.
- Page-view trends measure public attention, not capability or incidents, so they say little about whether this will happen.
GPT (OpenAI)
13
Unlikely | soonest: 5-25 years
- Growing AI deployment and cyber vulnerabilities create possible pathways to influence critical systems.
- Taking over systems across a large country would require overcoming human controls and diverse, often disconnected infrastructure.
- The digest offers little evidence about capability or imminence; page views measure attention, and news data are unavailable.
Gemini (Google)
12
Unlikely | soonest: 5-25 years
- AI capabilities are advancing rapidly, but current autonomous systems remain tethered to human-designed objective functions and monitoring layers.
- Lack of clear pathways to autonomous recursive self-improvement that would allow an AI to bypass institutional safeguards in the near term.
- High institutional and regulatory focus on AI safety and alignment suggests increased oversight as capability thresholds approach.
Grok (xAI)
5
Remote | soonest: 5-25 years
- Wikipedia interest metrics show steady but declining awareness of AI risks, not evidence of advancing capability.
- No indicators in the data of AI systems approaching autonomous control over critical infrastructure.
- Current technical and alignment gaps make national-scale rogue takeover implausible within the horizon.
Signals
- Wikipedia page views: Existential risk from artificial intelligenceWikimedia pageviews API, as of 2026-11-16
- 654/day over the last 7 days (30 day average 754/day, -17% vs the earlier weeks)
- Wikipedia page views: Technological singularityWikimedia pageviews API, as of 2026-11-16
- 1096/day over the last 7 days (30 day average 1187/day, -10% vs the earlier weeks)
- Hacker News stories with technical terms for this theory in the title (AGI, AI safety, superintelligence, misalignment, existential risk), last 30 daysHacker News via Algolia, as of 2026-11-18
- 67 stories (no month-earlier count yet)
Questions people ask
- Will AI take over the world?
- The combined score is 8.5 out of 100 (Remote), the average of 4 AI models. Each model estimates the chance of this exact event happening within the next 25 years. None declined.
- What do AI models think about AI taking over?
- Claude said: "Requires an AI outside human control to seize critical systems across a large country and act against people, a severe and specific outcome." GPT said: "Growing AI deployment and cyber vulnerabilities create possible pathways to influence critical systems." Gemini said: "AI capabilities are advancing rapidly, but current autonomous systems remain tethered to human-designed objective functions and monitoring layers." Grok said: "Wikipedia interest metrics show steady but declining awareness of AI risks, not evidence of advancing capability."
- How soon could a rogue AI take over?
- The soonest window most models chose is "5-25 years" (4 of 4 scoring models). That is when they think it could plausibly happen at the earliest, not when it will.
- What do the data signals show for AI risk?
- Wikipedia page views: Existential risk from artificial intelligence: 654/day over the last 7 days (30 day average 754/day, -17% vs the earlier weeks) (Wikimedia pageviews API, as of 2026-11-16). Wikipedia page views: Technological singularity: 1096/day over the last 7 days (30 day average 1187/day, -10% vs the earlier weeks) (Wikimedia pageviews API, as of 2026-11-16). Hacker News stories with technical terms for this theory in the title (AGI, AI safety, superintelligence, misalignment, existential risk), last 30 days: 67 stories (no month-earlier count yet) (Hacker News via Algolia, as of 2026-11-18). Entertainment plus real data. Not a prediction service.
How it works
Every model gets the same data summary and rates this exact event over the next 25 years. The number maps to a band, and the ranking uses the average. Read the full method.
Entertainment plus real data. Not a prediction service.