Cognitive scientist Gary Marcus and science-fiction author Ted Chiang recently engaged in a discussion at Town Hall Seattle about the burgeoning challenges associated with the rapid advancements in artificial intelligence (AI). During the event, Marcus highlighted the increasing risks posed by AI technologies, suggesting that regulatory measures may become necessary to ensure safety and accountability. He proposed models such as those used by the Food and Drug Administration (FDA) or the Federal Aviation Administration (FAA) to provide a framework for evaluating AI developments, arguing for a system where new AI models undergo rigorous assessments to ascertain their societal benefits versus potential harms. This implies a need for a structured approach to manage generative AI, which could involve requiring approval before new AI tools are introduced to the market.
An essential aspect of effective AI regulation, according to Marcus, is the implementation of external audits for software companies to evaluate AI performance. He pointed out the troubling use of large language models in job selection processes, noting the biases inherent in these systems and the lack of mechanisms to audit their impact on hiring practices. Marcus advocates for liability laws that hold companies accountable for any significant societal harm caused by their AI systems. This shift toward accountability is crucial in addressing the ethical concerns surrounding AI, as the current lack of oversight enables harmful practices to proliferate unchecked.
In Marcus’s new book, “Taming Silicon Valley,” he outlines the various problematic aspects of generative AI, such as issues with plagiarism, disinformation, deepfakes, and transparency. He remarked on the contradiction between the assurances tech companies provide about their commitment to AI safety and the absence of independent oversight in discussions between government and industry leaders. Highlighting the danger of regulatory capture, where powerful companies influence self-regulatory measures without external scientific input, Marcus emphasized the necessity for independent scientists to be involved in AI regulation. This concern underscores the risk that technological decisions may be made by a few influential figures in the industry without widespread consultation or consideration for potential societal risks.
Furthermore, the conversation touched on the contentious issue of open-source AI. With opinions divided among industry leaders—such as Meta’s Mark Zuckerberg advocating for open-source AI while Geoffrey Hinton, a leading figure in AI ethics, cautioned against it—Marcus stressed the importance of not leaving such significant decisions solely in the hands of corporate executives. He proposed the establishment of a Federal AI Administration or an International Civil AI Organization, drawing a comparison to the rigorous regulation of commercial airlines, which have multiple layers of oversight ensuring passenger safety. He argued that a similarly robust regulatory framework is essential for AI technologies to safeguard society and avoid potential crises stemming from unmanaged AI advancements.
Despite the compelling arguments for increased regulation, Marcus expressed skepticism about the likelihood of achieving these changes in the current political climate. He noted that substantial reforms are unlikely given the prevailing political trends. In his book, Marcus suggested a boycott of generative AI products as a way to induce change, although Chiang raised doubts about the practicality of such an approach, given the pervasive integration of AI in various software applications. Marcus conceded that while a widespread consumer boycott would be difficult, public awareness and pressure can be instrumental in advancing sound policies surrounding AI, similar to the advocacy seen for climate change.
In discussing specific AI applications, Marcus speculated that the performance enhancements associated with large language models might be reaching a plateau, as the ability to harness greater quantities of data becomes constrained. Chiang echoed this sentiment, acknowledging that while AI can revolutionize fields like materials science and biomedicine, the broader applications for reasoning and understanding complex real-world situations remain a challenge. Marcus expressed concern regarding OpenAI’s trajectory toward surveillance applications, suggesting that companies may prioritize data monetization over ethical considerations. Overall, their dialogue underscored the pressing need for regulatory frameworks and thoughtful discourse around the evolving landscape of artificial intelligence, particularly in terms of transparency, accountability, and the broader impact on society.
