Artificial Intelligence

Artificial intelligence has gone from a fringe issue to a major factor in human society. There is an urgent need to responsibly govern the technology, including for the possibility of it posing catastrophic risks.

An Introduction to Artificial Intelligence

When GCRI was founded in 2011, we were among the only groups taking AI risk seriously. Our 2017 and 2020 surveys of artificial general intelligence (AGI) projects was published in part to show that while AGI may seem like speculative science fiction, it was an active pursuit of many research and development groups. Now—early 2025—AGI doesn’t seem so speculative anymore. Current state-of-the-art systems already have some generality, being able to field inquiries across a wide range of domains, even while they still have significant limitations. AI has also become a major economic force and a significant political issue.

Where things go from here is not known. It is clear that the technology and its place in society will continue to change. However, the directions things could go in are extremely divergent. AI could take over the world, rendering humans obsolete, or it could be yet another economic bubble soon to burst, or it could end up with significant but comparatively modest impacts on the economy and society. These events could unfold within just a few months or over the course of many decades. The outcomes could be good or bad, or utopian or dystopian, or even catastrophic. The history of AI is full of false predictions and cycles of boom and bust, but it has never had such a high profile.

The only way to guarantee we avoid AI catastrophe is to stop building more advanced AI. That would also avoid other potential harms from the technology, such as mass unemployment, wealth inequality, environmental degradation, weapons proliferation, misinformation, harms to marginalized communities, and a loss of human connections. However, it would also miss out on any potential benefits, such as gains in productivity and economic growth, progress in science and technology, and, in the extreme case, the realization of more utopian scenarios, including scenarios involving large-scale expansion into outer space.

A central challenge, then, is to pursue courses of action that would, as much as can be done, obtain the benefits of AI while avoiding the harms, including catastrophic harms. This requires an understanding of both the technology itself and the social, political, and economic contexts in which it is governed. For example, to reduce the risk of catastrophe, it is important to identify the particular AI capabilities that could result in catastrophic outcomes and to craft governance approaches that address these capabilities across all AI groups, including those in rival corporations and countries. Whereas rival groups may typically prefer to compete, they may be more willing to cooperate to avoid catastrophic outcomes that no one wants, especially if there is outside pressure for them to do so.

GCRI’s work on AI centers on three broad themes: risk analysis, governance, and ethics.

Our AI risk analysis work develops models for assessing the potential for AI technology to cause catastrophic harm, especially the risk of catastrophic AI takeover. Our early risk analysis developed the ASI-PATH model as a theoretical framework for synthesizing and understanding certain concepts of AI risk. More recently, our study of large language model (LLM) risk develops a more practical framework for studying takeover catastrophe risk from LLMs and applies it to actual LLMs. The study found that the LLMs available at the time fall well short of the capabilities that may be needed for takeover, though the technology continues to change and continued monitoring is warranted.

Our AI risk governance work seeks to develop practical means for addressing catastrophic AI risk alongside other AI issues. We recognize that AI raises a wide range of issues and that some people active in AI issues are not as focused on catastrophic risk. Therefore, as a matter of strategy, our research seeks to develop win-win solutions that reduce catastrophic risks while also making progress on other issues. This includes advancing responsible corporate governance, achieving collective action between AI groups, and bridging divides between those concerned about catastrophic risks and other issues. We have also worked on environmental implications of AI, drawing on our background in climate and environment. We believe better outcomes will occur when those with synergistic goals can work together instead of being at odds with each other.

Our AI ethics work explores ideas for what ethical values should be built into AI systems, also known as machine ethics. This is another domain of wide relevance, including for both the AI systems that could pose catastrophic risks and for other, less capable systems. One specific focus is on “social choice” ethics that seek to align the values of AI systems with the values of humanity as a whole. This is analogous to how democracy can align the values of governments to the values of the populations being governed. It raises difficult questions about how to define the values of humanity as a whole, given that people have many disagreements about values, the potential for aggregate values to be manipulated, and the possibility of including the values of nonhumans. This work extends our broader work on ethics.

Image credits: map: Seth Baum; landscape: Buckyball Design; voting line: April Sikorski

For over 50 years, experts have worried about the risk of AI taking over the world and killing everyone. The concern had always been about hypothetical future AI systems—until recent LLMs emerged. This paper, published in the journal Risk Analysis, assesses how close LLMs are to having the capabilities needed to cause takeover catastrophe.

Private corporations are at the forefront of AI. This paper, published in the journal Information, surveys how to orient AI corporate governance toward the public interest, covering nine types of actors: management, workers, investors, corporate partners, industry consortia, nonprofit organizations, the public, the media, and governments.

In AI ethics, a major theme is the challenge of aligning AI to human values, but this raises the question of the role of nonhumans. This paper, published in the journal AI & Ethics, documents widespread neglect of nonhumans in AI ethics and argues for giving nonhumans adequate attention, including to avoid potentially catastrophic outcomes.

 

Full List of GCRI Publications on Artificial Intelligence

Owe, Andrea and Seth D. Baum, 2021. The ethics of sustainability for artificial intelligence. In Philipp Wicke, Marta Ziosi, João Miguel Cunha, and Angelo Trotta (Editors), Proceedings of the 1st International Conference on AI for People: Towards Sustainable AI (CAIP 2021), Bologna, pages 1-17, DOI 10.4108/eai.20-11-2021.2314105.

Galaz, Victor, Miguel A. Centeno, Peter W. Callahan, Amar Causevic, Thayer Patterson, Irina Brass, Seth Baum, Darryl Farber, Joern Fischer, David Garcia, Timon McPhearson, Daniel Jimenez, Brian King, Paul Larcey, and Karen Levy, 2021. Artificial intelligence, systemic risks, and sustainability. Technology in Society, vol. 67 (November), article 101741, DOI 10.1016/j.techsoc.2021.101741.

Baum, Seth and Trevor White, 2015. When robots kill. The Guardian Political Science blog, 23 June.

An Introduction to Artificial Intelligence

When GCRI was founded in 2011, we were among the only groups taking AI risk seriously. Our 2017 and 2020 surveys of artificial general intelligence (AGI) projects was published in part to show that while AGI may seem like speculative science fiction, it was an active pursuit of many research and development groups. Now—early 2025—AGI doesn’t seem so speculative anymore. Current state-of-the-art systems already have some generality, being able to field inquiries across a wide range of domains, even while they still have significant limitations. AI has also become a major economic force and a significant political issue.

Where things go from here is not known. It is clear that the technology and its place in society will continue to change. However, the directions things could go in are extremely divergent. AI could take over the world, rendering humans obsolete, or it could be yet another economic bubble soon to burst, or it could end up with significant but comparatively modest impacts on the economy and society. These events could unfold within just a few months or over the course of many decades. The outcomes could be good or bad, or utopian or dystopian, or even catastrophic. The history of AI is full of false predictions and cycles of boom and bust, but it has never had such a high profile.

The only way to guarantee we avoid AI catastrophe is to stop building more advanced AI. That would also avoid other potential harms from the technology, such as mass unemployment, wealth inequality, environmental degradation, weapons proliferation, misinformation, harms to marginalized communities, and a loss of human connections. However, it would also miss out on any potential benefits, such as gains in productivity and economic growth, progress in science and technology, and, in the extreme case, the realization of more utopian scenarios, including scenarios involving large-scale expansion into outer space.

A central challenge, then, is to pursue courses of action that would, as much as can be done, obtain the benefits of AI while avoiding the harms, including catastrophic harms. This requires an understanding of both the technology itself and the social, political, and economic contexts in which it is governed. For example, to reduce the risk of catastrophe, it is important to identify the particular AI capabilities that could result in catastrophic outcomes and to craft governance approaches that address these capabilities across all AI groups, including those in rival corporations and countries. Whereas rival groups may typically prefer to compete, they may be more willing to cooperate to avoid catastrophic outcomes that no one wants, especially if there is outside pressure for them to do so.

GCRI’s work on AI centers on three broad themes: risk analysis, governance, and ethics.

Our AI risk analysis work develops models for assessing the potential for AI technology to cause catastrophic harm, especially the risk of catastrophic AI takeover. Our early risk analysis developed the ASI-PATH model as a theoretical framework for synthesizing and understanding certain concepts of AI risk. More recently, our study of large language model (LLM) risk develops a more practical framework for studying takeover catastrophe risk from LLMs and applies it to actual LLMs. The study found that the LLMs available at the time fall well short of the capabilities that may be needed for takeover, though the technology continues to change and continued monitoring is warranted.

Our AI risk governance work seeks to develop practical means for addressing catastrophic AI risk alongside other AI issues. We recognize that AI raises a wide range of issues and that some people active in AI issues are not as focused on catastrophic risk. Therefore, as a matter of strategy, our research seeks to develop win-win solutions that reduce catastrophic risks while also making progress on other issues. This includes advancing responsible corporate governance, achieving collective action between AI groups, and bridging divides between those concerned about catastrophic risks and other issues. We have also worked on environmental implications of AI, drawing on our background in climate and environment. We believe better outcomes will occur when those with synergistic goals can work together instead of being at odds with each other.

Our AI ethics work explores ideas for what ethical values should be built into AI systems, also known as machine ethics. This is another domain of wide relevance, including for both the AI systems that could pose catastrophic risks and for other, less capable systems. One specific focus is on “social choice” ethics that seek to align the values of AI systems with the values of humanity as a whole. This is analogous to how democracy can align the values of governments to the values of the populations being governed. It raises difficult questions about how to define the values of humanity as a whole, given that people have many disagreements about values, the potential for aggregate values to be manipulated, and the possibility of including the values of nonhumans. This work extends our broader work on ethics.

Image credits: map: Seth Baum; landscape: Buckyball Design; voting line: April Sikorski

 

Featured GCRI Publications on Artificial Intelligence

For over 50 years, experts have worried about the risk of AI taking over the world and killing everyone. The concern had always been about hypothetical future AI systems—until recent LLMs emerged. This paper, published in the journal Risk Analysis, assesses how close LLMs are to having the capabilities needed to cause takeover catastrophe.

Private corporations are at the forefront of AI. This paper, published in the journal Information, surveys how to orient AI corporate governance toward the public interest, covering nine types of actors: management, workers, investors, corporate partners, industry consortia, nonprofit organizations, the public, the media, and governments.

In AI ethics, a major theme is the challenge of aligning AI to human values, but this raises the question of the role of nonhumans. This paper, published in the journal AI & Ethics, documents widespread neglect of nonhumans in AI ethics and argues for giving nonhumans adequate attention, including to avoid potentially catastrophic outcomes.

 

Full List of GCRI Publications on Artificial Intelligence

Owe, Andrea and Seth D. Baum, 2021. The ethics of sustainability for artificial intelligence. In Philipp Wicke, Marta Ziosi, João Miguel Cunha, and Angelo Trotta (Editors), Proceedings of the 1st International Conference on AI for People: Towards Sustainable AI (CAIP 2021), Bologna, pages 1-17, DOI 10.4108/eai.20-11-2021.2314105.

Galaz, Victor, Miguel A. Centeno, Peter W. Callahan, Amar Causevic, Thayer Patterson, Irina Brass, Seth Baum, Darryl Farber, Joern Fischer, David Garcia, Timon McPhearson, Daniel Jimenez, Brian King, Paul Larcey, and Karen Levy, 2021. Artificial intelligence, systemic risks, and sustainability. Technology in Society, vol. 67 (November), article 101741, DOI 10.1016/j.techsoc.2021.101741.

Baum, Seth and Trevor White, 2015. When robots kill. The Guardian Political Science blog, 23 June.