Source-linked AI summary
The Transformative Potential of Artificial Intelligence
Ross Gruetzemacher, Jess Whittlestone
TL;DR
The paper addresses poorly defined concepts of human-level and general AI that may not capture the societal impacts most relevant to advanced AI. It analyzes transformative AI through historical literature and proposes three levels of societal transformation. The resulting framework is intended to improve discussion, preparation, and resource allocation concerning different advanced-AI futures.
Problem
Terms such as HLAI and AGI are poorly defined and focus attention on humanlike or general capabilities, while interpretations of transformative AI remain ambiguous.
Method
The paper reviews literature on economic history and technology-driven societal change to distinguish three levels of transformative AI and define practically irreversible change.
Results
The paper proposes a framework distinguishing three levels of possible societal transformation from AI and argues that TAI—potentially practically irreversible change across society—is neglected in current discussions.
Takeaways & Limitations
The proposed levels give researchers, strategic planners, and decision makers a framework for understanding possible futures, preparing for impacts, and allocating resources.
Takeaways & Limitations
The framework’s treatment of potential impacts is incomplete, including possible effects from AI-driven surveillance, lethal autonomous weapons, and reinforcement learning beyond productivity gains.
Abstract
from arXiv · showhide
The terms 'human-level artificial intelligence' and 'artificial general intelligence' are widely used to refer to the possibility of advanced artificial intelligence (AI) with potentially extreme impacts on society. These terms are poorly defined and do not necessarily indicate what is most important with respect to future societal impacts. We suggest that the term 'transformative AI' is a helpful alternative, reflecting the possibility that advanced AI systems could have very large impacts on society without reaching human-level cognitive abilities. To be most useful, however, more analysis of what it means for AI to be 'transformative' is needed. In this paper, we propose three different levels on which AI might be said to be transformative, associated with different levels of societal change. We suggest that these distinctions would improve conversations between policy makers and decision makers concerning the mid- to long-term impacts of advances in AI. Further, we feel this would have a positive effect on strategic foresight efforts involving advanced AI, which we expect to illuminate paths to alternative futures. We conclude with a discussion of the benefits of our new framework and by highlighting directions for future work in this area.
1. Introduction
Recent AI progress has intensified concern about societal impacts, while terms such as HLAI and AGI remain focused on humanlike or general capabilities. The paper introduces transformative AI as a broader lens and develops distinctions to clarify possible societal changes.
- Recent advances in deep learning have raised concerns about AI applications and societal impacts among researchers, policy professionals, and the public.
- HLAI, HLMI, and AGI emphasize humanlike or sufficiently general capabilities, although advanced AI could produce dramatic societal changes without human-level cognition.
- Transformative AI captures concern about a broad spectrum of advanced systems with substantial potential for societal impact, but its usage remains ambiguous.
- This ambiguity limits clear understanding, anticipation, forecasting, and communication about possible future AI scenarios.
- The paper reviews existing TAI definitions and literature on economic history and technology-driven change before proposing three levels of AI transformation.
- The framework is intended to clarify discussions of advanced AI, potential impacts, research priorities, and futures research amid a wide range of plausible futures.
2. Existing Definitions of Transformative AI
Existing definitions of transformative AI vary and are often ambiguous, especially about the scale of change implied. Some invoke comparisons with historical revolutions without specifying how those comparisons should be interpreted.
- Four existing definitions of TAI are presented, but all are described as somewhat ambiguous.
- Definitions differ over whether transformation means radical changes to welfare, wealth, or power or extends beyond a narrow task toward superintelligence.
- Definitions comparing TAI with agricultural or industrial revolutions remain unclear about what it means for AI-driven change to be comparable to the industrial revolution.
- Business, government, and policy organizations also use transformation informally to describe AI’s societal impact.
3. Transformative Societal Change in History
Historical research distinguishes transformative technologies by the breadth and extremity of their societal effects, with the agricultural and industrial revolutions representing especially extensive changes. This literature motivates analyzing AI transformation through practically irreversible change and explicit dimensions of impact.
- Different types of transformation: Historical literature identifies multiple types and levels of technology-driven societal transformation, rather than a single uniform pattern.Examples include general-purpose technologies, narrower domain transformations, long waves, and major revolutions.
- General-purpose technologies: General-purpose technologies are widely used across many applications and generate spillover effects, supporting economic growth across sectors.Electricity is a commonly cited example, while GPT-related productivity gains may emerge only after complementary innovations.
- Narrow but extreme impacts: Some technologies can be transformative through extreme effects in a narrow domain without qualifying as general-purpose technologies.Nuclear weapons are presented as transforming warfare and international relations despite more limited economic breadth.
- Major revolutions: The agricultural and industrial revolutions produced unusually broad and extreme changes, including settled civilizations, mechanized production, population growth, and rising quality of life.The industrial revolution is also described as a phase transition initiating self-sustaining and accelerating economic growth.
- Historical magnitude: Figure 1 uses global average GDP and war-making capacity to illustrate the industrial revolution’s sharp and historically unusual change in measures of human progress.Both axes are logarithmic, and the industrial-revolution period is shaded.
- Dimensions of transformation: Transformative change is characterized as practically irreversible and can vary in breadth and extremity across technologies.The framework defines practical irreversibility through long-lasting, extremely costly-to-revoke effects and emphasizes lock-in and path dependence as related ideas.
4. Defining Levels of Transformative AI
The paper proposes three levels of transformative AI, distinguishing potential societal impact while acknowledging uncertainty about which level an AI system may precipitate. It illustrates plausible routes from current capabilities and emerging technologies to these levels.
- Proposed framework: The framework distinguishes three levels of transformative AI, including narrowly transformative AI and radically transformative AI.The supplied table material names the narrow and radical levels, while the paper states that the framework contains three levels.
- Proposed framework: The framework emphasizes potential societal impact because the level of change an AI system will precipitate cannot be determined with certainty in advance.The authors present the levels as a way to characterize possible impacts, not as an inevitable progression.
- Narrowly Transformative AI: Narrowly transformative impacts could arise from widespread use of existing AI capabilities, including surveillance that changes state power, policing, and privacy, or autonomous weapons that alter conflict.These examples are compared with historical impacts such as nuclear weapons, whose effects were more concentrated than those of general-purpose technologies.
- Plausible paths: Offline reinforcement learning could enable end-to-end decision automation across business, healthcare, and robotics, offering a path to broad societal impact.The paper contrasts this potential with supervised learning, which it says has not yet delivered the same real-world value for reinforcement learning.
- Plausible paths: Continued scaling of transformer language models and progress in offline reinforcement learning are presented as plausible paths to a productivity bonus associated with transformative AI.The paper links language advances to practical human-machine interaction through language user interfaces.
- Radically Transformative AI: Radically transformative AI is more speculative and could emerge from systems performing most economically relevant tasks, potentially replacing jobs and accelerating scientific progress without fully human-level intelligence.The paper describes this as a plausible scenario rather than a detailed assessment or prediction.
5. Discussion
The discussion distinguishes transformative AI levels, argues that near-term and broad societal impacts deserve attention beyond human-level AI, and identifies research and foresight priorities. It also emphasizes uncertainty about pathways, historical analogies, and possible consequences.
- Implications for the AI research community: Distinguishing NTAI, TAI, and RTAI highlights that societal transformation may occur before fully general or human-level AI capabilities.TAI concerns practically irreversible change across important domains, while RTAI is associated with transformation comparable to historical agricultural or industrial revolutions.
- Implications for the AI research community: Current or near-future AI could produce practically irreversible changes in important domains, including pervasive economic impacts from advances in offline reinforcement learning.The discussion presents such impacts as a currently neglected topic compared with attention to immediate effects and extreme human-level or superintelligent AI.
- Implications for the AI research community: Potential impacts extend beyond productivity or economics to surveillance, lethal autonomous weapons, authoritarianism, conflict, and other areas of life and society.The discussion explicitly warns that treating offline reinforcement learning’s effects as merely a productivity bonus would be an oversimplification.
- How do different levels of societal transformation relate to each other?: Historical evidence does not clearly identify a single technology as the cause of radical societal transformation, which may instead reflect interacting technology clusters and societal factors.This complicates direct analogies between RTAI and historical revolutions, although AI itself may function as an underlying method generating multiple technologies.
- Implications for futures researchers and practitioners: The framework is intended to shift assumptions that radical AI-driven societal transformation requires anthropomorphic AGI or human-level machine intelligence.It aims to open pathways for considering safe and beneficial RTAI and to support broader exploration of AI futures.
- Over what timeframe could transformative impacts of AI occur?: The framework leaves open whether lower levels of transformation precede higher ones and how quickly RTAI could emerge, motivating scenario preparation and further forecasting research.The authors recommend more rigorous frameworks using forecasting, foresight, and expert-opinion aggregation because claims about AI pathways are subjective and uncertain.
6. Conclusion
The conclusion frames transformative AI through three levels of societal change and argues that TAI is a neglected possibility distinct from both immediate impacts and human-level or superintelligent AI. It presents the framework as useful for preparation, resource allocation, and urgent exploration of plausible AI pathways.
- 6. Conclusion: The paper uses historical technology impacts to distinguish three levels of possible societal transformation from AI.TAI is defined as AI technologies or applications with potential to cause practically irreversible societal and economic change across all society.
- 6. Conclusion: TAI is presented as a neglected topic because existing discussions often focus on immediate AI impacts or the extreme possibility of human-level or superintelligent AI.The paper considers TAI’s emergence over the next decade plausible, potentially through advanced offline reinforcement learning or scaled transformer language models.
- 6. Conclusion: The proposed levels give researchers, strategic planners, and decision makers a framework for understanding possible AI futures, preparing for different transformations, and allocating resources.The conclusion calls for urgent work on plausible paths to TAI and their consequences because development could be rapid.