In May 2023, hundreds of AI scientists published a statement on the reality of AI Existential Risk (X-Risk). Despite this being the latest in a series of warnings, public awareness and concern remain limited. This discrepancy may be attributed to the complexity of the subject and the prevalence of fallacious or ill-informed counterarguments in the public debate. This is a dangerous situation that needs to be addressed urgently.
For that purpose, this paper introduces a comprehensive framework for clarifying the debate on Artificial Super Intelligence (ASI) and its existential risks. The framework comprises a simple and accessible model of AI x-risk, coupled with a methodology for gathering, organising, and analysing arguments related to this model. The methodology involves assembling arguments from various sources, organising them into coherent hierarchical taxonomies, and systematically analysing and synthesising these arguments to reveal areas of contention and consensus. With its hierarchical structure, the framework aims to present a panoramic view of the current discourse, allowing users to understand the current consensus and explore the underlying reasoning at multiple levels of granularity while countering fallacies and bias.
Additionally, the paper proposes the creation of an online collaborative hub. Although not developed within this project, the hub is designed to encourage ongoing dialogue and knowledge sharing among researchers, experts, and the public. It aspires to become a dynamic platform for evolving consensus on AI x-risk. Future work will focus on the development and promotion of this hub, seeking to establish it as an accessible and unbiased entry point into the debate on AI x-risk.
In May 2023, the leaders of the top 3 AI labs along with hundreds of prominent AI scientists signed the following statement¹: “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.” This was the latest in a series of warnings about AI Existential Risk (X-Risk) coming from researchers, experts, philosophers and public figures. However, AI X-Risk still ranks low² in the general public’s list of priorities and is not at the centre of the public debate, which raises the question: why are these warnings not taken more seriously? Those trying to engage objectively with the subject of AI x-risk are faced with a worrying picture. The case for its potential threat³ ⁴ ⁵ ⁶ ⁷ seems compelling, with rational and empirical support, whereas serious counter arguments from within the field are hard to be found. Most of the opposition appears to be focused on ridiculing the concern instead of addressing it rationally⁸ ⁹ ¹⁰ ¹¹ ¹². The complexity of the subject, along with the overwhelming presence of fallacious arguments pushed by motivated actors, contributes to a general lack of awareness about the imbalance between these conflicting viewpoints. This dangerous situation underscores an urgent need for a clear consensus on the reality of AI X-Risk. This project aims to breach this gap by laying the foundations for a lucid, objective view of AI X-Risk.
For this purpose, this paper introduces a comprehensive framework designed to structure and clarify the debate surrounding the creation of Artificial Super Intelligence (ASI) and its potential existential risks. The framework comprises a simple and useful model of AI X-Risk based on four hypotheses about ASI, coupled with a methodology for gathering, organising, and analysing arguments related to these hypotheses. The methodology involves collecting resources, extracting supportive and oppositional arguments, organising them into structured taxonomies, and systematically analysing and synthesising these arguments to reveal areas of contention and consensus. The framework aims to present an unbiased, panoramic view of the current discourse, allowing users to understand the current consensus on each hypothesis and explore the underlying reasoning in depth.
The framework's strength lies in its scientific method-like approach, with a model built upon clear hypotheses, inviting both support and scrutiny. The hierarchical taxonomies of arguments for and against each hypothesis are key. By presenting a synthesis of the debate at every node of the taxonomies, the framework allows users to grasp the consensus with remarkable clarity and depth. Starting from a high-level overview, users can see, for instance, why Hypothesis 1 (H1: ASI will be created in less than 20 years) is considered 'Likely', based on the categorization and assessment of supporting and opposing arguments. If they want to know more, they can drill into, for example, the opposition categories (1. ASI is impossible, 2. ASI is very far away …) and access syntheses at this new level. This structure not only makes the consensus on each hypothesis transparent but also guides users through the reasoning behind it. It demystifies complex debates, breaking them down into comprehensible segments, thereby facilitating a deeper understanding at any desired level of granularity.
Finally, it invites researchers, experts and members of the public on all sides of this debate to contribute via an online collaborative hub, designed to facilitate dialogue, knowledge sharing and provide a dynamic platform for evolving consensus on AI X-Risk. This collaborative hub, while not developed within the scope of this project, is envisioned as a key component for future engagement, and its specifications are outlined in this paper. Future work will focus on the hub's development and promotion, seeking to establish it as an unbiased, inclusive and accessible entry point into this debate.
The framework, alongside with the collaborative hub, fosters engagement by making it easy for individuals to identify where their views diverge from the consensus. The hierarchical structure is not just a tool for understanding—it's also a catalyst for collaboration. As users navigate the taxonomy trees, they can identify specific points of agreement or disagreement with the current consensus, significantly lowering the barrier to entry for contributing to the debate. Individuals can easily find and address the particular node or argument where their insights or counterarguments are most relevant.
<aside> ⚠️ On a more personal and worrying note, I started this project wanting to learn more about the arguments for and against x-risk. More specifically, hoping that there was a serious case against x-risk. My concerns have only been exacerbated as it dawned on me how difficult it was to find actual arguments that address the x-risk model or explain how a super intelligence can be harmless. It is not an overstatement to say that the overwhelming majority of the opposition uses fallacies or simply dodges the questions. It is of the utmost urgency that the world opens its eyes on this reality.
</aside>
The aim of this framework is to provide a centralised way to analyse all the arguments in the debate around AI X-Risk, shielded from fallacies and biases. The central point of this framework is a simple model comprising of four hypotheses that, taken together, imply the reality of existential risk. Databases were created for listing online resources that discuss AI X-Risk, and individual arguments were extracted from these resources. These arguments are organised and classified into two hierarchical taxonomies: a taxonomy of evidence in support of the model, and a taxonomy of evidence falsifying the model, the roots of the taxonomies corresponding to the four hypotheses of the model. Methodologies and guidelines have been created to carry out a systematic analysis of each argument, and a synthesis at each node of the taxonomy. A consensus can then emerge from the bottom of the tree to the hypotheses at the top. Some examples of analyses and syntheses were created.
With the final result in hands, a user should be able to easily access the consensus on each hypothesis, and is able to navigate down the taxonomy trees in order to get a more granular understanding.
<aside> 💡 Example Suppose that the consensus on H3 is “Likely”. A user can see that the support consensus is “Likely” and the opposition consensus is “Unlikely”. If they want more information about the opposition, they can dive deeper down the tree of the opposition taxonomy. They would see that the opposition to this hypothesis breaks down into:
</aside>
<aside> ⚠️ A Note on Bias At this point it is important to point out that this example is representative of the bias of the author, but the vocation of the framework is to be unbiased. Contributors can and should come from any side of the debate, as long as they follow the methodologies and are respectful of the principles of rational debate.
</aside>