How could AI be an existential risk?

A systematic review and synthesis of AI existential catastrophe scenarios

Led by

In collaboration with

What is this review?

This systematic review was led by the Existential Risk Observatory, in collaboration with the MIT AI Risk Initiative and the Future of Life Institute.

Government and academic assessments have determined that AI may pose an existential risk. An existential risk threatens to destroy humanity’s long-term potential, whether through extinction or another unrecoverable outcome. The pathways to such outcomes may differ substantially, and measures designed to prevent one scenario may fail to address, or even increase, the likelihood of others. Although researchers have proposed numerous scenarios for how AI could cause catastrophic and existential harm, few studies have systematically collected and compared them.

To address this gap, we conducted a systematic review, screening 3,184 documents and identifying 95 AI existential risk scenarios. We classified these scenarios into eight scenario families. Three describe power-seeking AI systems disempowering humanity: Unilateral, Dispersed, and Coordinated Takeover. In a fourth family, Incremental Handover, humans irrevocably delegate authority to AI systems that do not seek power. Two families describe AI systems enabling permanent power concentration in a narrow group of humans: Entrenched Totalitarianism and Economic Stratification. The last two families describe how AI systems expand humans’ destructive capabilities to an existential level: Catastrophic Invention and Automated Warfare.

Each family is a sequence of causal steps through which an existential catastrophe could occur. Together, the eight families form a framework that makes the assumptions behind different scenarios easier to examine, enables comparisons across pathways, and helps identify opportunities for intervention. The classification also shows that the literature has concentrated on Unilateral Takeover and Incremental Handover, while Coordinated Takeover, Automated Warfare, and Economic Stratification remain comparatively unexplored. The families provide a foundation for more comprehensive AI risk assessment, incident monitoring, research, governance, and mitigation.

The review is part of the MIT AI Risk Initiative, which aims to increase awareness and adoption of best practice AI risk management across the ecosystem.

Eight families of existential risk scenarios

The review identifies 95 published scenarios in 58 documents and sorts them into eight scenario families.

The eight scenario families (interactive; loads on the published page)

How do the scenario families differ?

The scenario families are context rich and differ in many aspects. The colored groupings showcase some of these and the questions below divide the families further.

The decision tree (interactive; loads on the published page)

The full field guide at aixrisks.com sets out every step in full, the ways each can happen, and all 95 scenarios with their sources.

Explore the full field guide ↗
Abstract 3D terrain with pink and teal contour lines on a black background.

Read the full study

Existential Risk Scenarios from AI: A Systematic Review

Read the paper (PDF) →

Funding

This work was supported by the AI Safety Tactical Opportunities Fund, who did not review the design, collection, analysis, interpretation or reporting of the data.

The MIT AI Risk Initiative and MIT FutureTech are funded by a range of organizations; none of those organizations influenced the design, collection, analysis, interpretation, or reporting of the data.

Acknowledgements

We thank Kaj Sotala (Finnish Center for Safe AI Tutke), Daniel Kokotajlo (AI Futures Project), Joshua Krook (University of Antwerp), Matthew Tokson (University of Utah), and Leslie Thiele (University of Florida) for their responses to the expert consultation, along with four experts who preferred to remain anonymous. Their contributions do not imply endorsement of this work or its conclusions.

Authors

Hugo Save, Peter Slattery, Michael Noetel, Risto Uuk, Martin Veron, David Africa, Peter Barnett, Jonathan Barry, Ze Shen Chin, Antoni Lorente, Malcolm Murray, Simon Mylius, Alexander Saeri, Neil Thompson, Wolfhart Totschnig, Karina Vold, Peter Wallich, Roman Yampolskiy, Kun Zhao and Otto Barten

For full author affiliations, see the paper.

Citation

Save, H., Slattery, P., Noetel, M., Uuk, R., Veron, M., Africa, D., Barnett, P., Barry, J., Chin, Z. S., Lorente, A., Murray, M., Mylius, S., Saeri, A. K., Thompson, N., Totschnig, W., Vold, K., Wallich, P., Yampolskiy, R., Zhao, K., & Barten, O. (2026). Existential risk scenarios from AI: A systematic review. https://aixrisks.com

Collaborate with us

We work with governments, companies, and researchers to build better tools for understanding AI risk. Subscribe to our newsletter for updates. Get in touch to explore partnership opportunities.

Feedback