Forecasting and Assessing the Impact of Artificial Intelligence on Society

TL;DR

This paper presents a comprehensive study forecasting the potential impact of artificial intelligence (AI) on society by identifying AI capabilities, postulating future products, and soliciting expert opinions on their applications, implications, and timelines. It highlights the critical need to understand AI's societal effects beyond its technical feasibility, emphasizing the importance of foresight in guiding its development.

What Problem Does This Paper Address?

As AI research progresses, machines are becoming increasingly sophisticated and influential in human affairs. However, there's a significant lack of organized effort to systematically understand both the capabilities and limitations of AI and its potential societal impact. This paper addresses this gap by providing a structured framework for forecasting AI's future and assessing its implications, aiming to inform decision-makers about the potential consequences of AI development.

Methods

The study employed a two-part methodology:

  1. Technical Portion: This involved delineating areas of AI, examining current capabilities, and postulating hypothetical products that could result from future advancements. This was an iterative process, where conceptualizing products led to a deeper investigation of required capabilities.
  2. Judgmental Portion: This used a Delphi study, a systematic forecasting technique, to gather expert opinions. Participants were asked about prototype and commercial dates for the postulated products, as well as the possibility and desirability of their applications and implications. The Delphi process involved multiple rounds of questionnaires with feedback summaries to refine consensus.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • AI's Transformative Potential: The study underscores that intelligent machines will fundamentally alter human lives and society in ways unprecedented in history, moving beyond mere automation to enhance or replace human capabilities in traditionally human domains.
  • The Need for Foresight: While predicting the future is inherently uncertain, especially with rapidly evolving technology, making decisions with long-term implications necessitates understanding AI's potential societal effects. This study aims to provide a basis for such informed decision-making.
  • Key AI Capabilities Driving Products: The research identified core AI capabilities like language understanding, problem-solving, perception, modeling, learning, robotics, and game playing as foundational for developing future AI products.
  • Societal Impact is Multifaceted: The postulated AI products have implications across various societal sectors, including human information processing, security, health, law, commerce, education, and social interaction.
  • Potential for Both Benefit and Detriment: Experts expressed optimism about AI's potential, but also identified specific products with significant risks, such as opportunities for censorship, government monitoring, and the displacement of workers.
  • Challenges in Formalizing Human Judgment: A significant hurdle for AI development is the difficulty in formalizing subjective human concepts like aesthetics, mores, and ethics, which are crucial for products like computer arbiters or creative systems.
  • Expert Consensus and Divergence: The Delphi study revealed a general optimism among experts regarding AI development timelines and desirability, though some products were flagged as potentially dangerous, highlighting areas where likelihood and negative desirability coincided.

Key Results

The study identified 21 hypothetical AI products. Experts provided estimates for prototype and commercial dates, with many products expected to have prototypes by 1985. The Delphi study revealed that:

  • Half of the AI products were considered of "high potential significance."
  • Respondents were generally optimistic about technology development, with many products having median prototype dates before 1985.
  • A consistent lag of 5-12 years was observed between median prototype and median commercial dates for products with early prototypes.
  • Products flagged as potentially dangerous due to high likelihood and detrimental implications included: Automated Inquiry Systems (censorship), Automated Intelligence Systems (government monitoring), Automatic Identification Systems (government monitoring), Mobile Robots (military use), Machine/Animal Symbionts (weapons, value alteration), Computer Psychiatrists (government influence), Robot Tutors (indoctrination), and General Factotums (emotional impact).

Limitations and Future Work

The authors acknowledge several limitations:

  • Extrapolation Difficulty: It has historically been challenging to extrapolate the results of AI research, as much of the work cited involved limited domains, small data bases, and ad hoc techniques. The extendibility of these techniques to "real-life" data bases remains uncertain.
  • Evaluating Progress: Determining significant progress in AI is difficult, as advancements might be analogous to climbing a tree when the goal is to reach the moon.
  • Formalizing Subjectivity: The formalization of subjective human concepts like aesthetics, ethics, and judgment remains a significant challenge, hindering the development of certain AI products.
  • Delphi Methodology Constraints: The Delphi method has challenges related to participant time commitment and the difficulty of accurately calibrating expert self-assessments.

Future work could involve a more detailed examination of the societal implications, potentially through further Delphi rounds or other forecasting methods. The paper also suggests the need for an "Al Ethical Review Board" to address safety, privacy, and ethical concerns, and to communicate potential dangers to decision-makers.

Summary updated

BibTeX
@inproceedings{1973forecasting,
  title = {{Forecasting and Assessing the Impact of Artificial Intelligence on Society}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1973},
}
Figure