If Artificial Intelligence Attains Consciousness: Methods for Detection
“If Artificial Intelligence Attains Consciousness: Methods for Detection”
Researchers could employ a checklist based on six theories grounded in neuroscience to facilitate the assessment. The concept of artificial intelligence gaining consciousness, a theme often explored in science fiction, is progressively moving beyond the realm of fantasy due to the swift advancements in artificial intelligence (AI). This idea has even been acknowledged by prominent figures in the AI field. In the previous year, for instance, Ilya Sutskever, the chief scientist at OpenAI, the organization responsible for the ChatGPT chatbot, suggested that certain state-of-the-art AI networks might possess a degree of “slight consciousness.”
While many experts assert that AI systems have not yet reached the threshold of consciousness, the rapid evolution of AI has sparked contemplation: how would we recognize their consciousness if it were to emerge?
A Multi-Decade Debate on Consciousness Concludes — With Philosophers 1, Neuroscientists 0
In response to this inquiry, a collaborative team of 19 professionals spanning neuroscientists, philosophers, and computer scientists has developed a set of criteria, presented in the form of a checklist, which could indicate a high probability of an AI system possessing consciousness. This provisional guide was recently released on the arXiv preprint repository prior to undergoing peer review. The motivation behind this effort, as co-author Robert Long, a philosopher at the Center for AI Safety in San Francisco, California, explained, was the perceived lack of comprehensive and empirically grounded discussions concerning AI consciousness.
The team underscores the ethical ramifications of failing to identify whether an AI system has achieved consciousness. If an entity were labeled as “conscious,” according to co-author Megan Peters, a neuroscientist at the University of California, Irvine, it would substantially influence how humans perceive the treatment that entity should receive.
Long further suggests that despite remarks from leaders of prominent AI research laboratories about pondering AI consciousness and sentience, there appears to be insufficient effort directed by AI companies toward evaluating models for consciousness or formulating plans for addressing such an eventuality.
The team reached out to major technology firms advancing AI, namely Microsoft and Google. Microsoft’s representative stated that the company’s AI development focuses on enhancing human productivity responsibly, rather than attempting to replicate human intelligence. The introduction of GPT-4, the most advanced iteration of the publicly available ChatGPT, has highlighted the need for new methodologies to assess the capabilities of these AI models for the broader benefit of society. Google did not respond to inquiries.
Get Alyaqobi Hasan’s stories in your inbox
Join Medium for free to get updates from this writer.
Defining Consciousness
A significant challenge in exploring consciousness in AI lies in defining the concept itself. For the purposes of this report, the researchers concentrated on ‘phenomenal consciousness,’ also known as subjective experience. This refers to what it feels like to be a person, an animal, or even an AI system (should one of them indeed possess consciousness).
The Quest for Alternative AI Assessment Beyond the Turing Test
Numerous neuroscience-based theories provide insights into the biological underpinnings of consciousness, yet consensus on the definitive theory remains elusive. To tackle this issue, the researchers employed a variety of these theories, aiming to ascertain AI consciousness by evaluating whether AI operations align with the principles of these theories. This strategy, they argue, is more effective than relying solely on behavioral tests, as AI systems have become highly adept at mimicking human behaviors.
Renowned neuroscientist Anil Seth commends the team’s theory-intensive approach, noting that it highlights the necessity for well-tested theories of consciousness. The authors assumed that consciousness is linked to how systems process information, regardless of their composition — be it neurons, computer chips, or another substrate. This approach, referred to as computational functionalism, also assumes that neuroscience-based theories of consciousness, investigated through brain scans and other techniques in humans and animals, can be extended to AI.
Could Cultivated Brain Tissue Attain Consciousness?
Based on these assumptions, the team selected six theories and extracted a list of indicators of consciousness. One such theory, the global workspace theory, posits that humans and other animals utilize various specialized systems, termed modules, to carry out cognitive tasks like seeing and hearing. These modules operate independently yet in parallel, sharing information by integrating into a unified system. Evaluating whether a particular AI system manifests an indicator derived from this theory would involve analyzing the system’s architecture and information flow, as explained by Long.
Seth appreciates the clarity of the team’s proposal, highlighting its thoughtfulness and transparent assumptions, even though he may disagree with some of these assumptions. The authors stress that their paper is not a definitive guide for assessing AI systems’ consciousness. They invite fellow researchers to collaborate in refining the methodology. While the criteria could already be applied to existing AI systems, the report indicates that large language models like ChatGPT, for instance, may demonstrate indicators associated with consciousness as outlined by the global workspace theory. Nonetheless, the work does not yet posit that any existing AI system is a strong candidate for consciousness.
References
Butlin, P. et al. Preprint at https://arxiv.org/abs/2308.08708 (2023).
