In the realm of science fiction, the concept of artificial intelligence (AI) gaining consciousness has long struck our imagination. From HAL 9000 in 2001: A Space Odyssey to Ava in Ex Machina, the idea of intelligent machines both fascinates and terrifies us. But as AI continues to evolve at an unprecedented rate, this once-fantastical idea is becoming increasingly plausible. What’s more, leaders in the AI field have even recognized the possibility that AI systems may possess consciousness.
Last year, Ilya Sutskever, chief scientist at OpenAI, tweeted that advanced AI networks could be “somewhat conscious.” While many researchers argue that AI has not yet reached the level of consciousness, the rapid evolution of AI has led them to wonder how we can even determine whether it possesses consciousness.
To address this question, a panel of 19 experts made up of neuroscientists, philosophers, and computer scientists developed a checklist of criteria that can be used to determine how likely it is that an AI system possesses consciousness. Their preliminary guide was recently published in the arXiv preprint repository, marking a significant step towards understanding AI consciousness. The study was published in the journal arxiv
The need for clarity
The motivation for this endeavor stems from the lack of detailed and empirically grounded discussions around AI consciousness. Co-author Robert Long, a philosopher at the Center for AI Safety, emphasizes that failure to determine whether an AI system has achieved consciousness is fraught with serious moral implications. According to neuroscientist Megan Peters, assigning something the status of “conscious” significantly affects our perceptions and attitudes toward that object.
In addition, Long notes that companies responsible for developing advanced artificial intelligence systems do not pay enough attention to evaluating these models for the presence of consciousness or developing action plans in case of its appearance. Despite this, the heads of leading laboratories openly express their bewilderment about consciousness and the intelligence of AI.
Addressing the technology giants
To learn more about this issue, Nature reached out to the two major tech companies leading the way in AI development: Microsoft and Google. Microsoft representatives emphasized that their goal is to use AI to responsibly improve human performance, not to replicate human intelligence. They recognized the need to develop new techniques to evaluate the capabilities of AI models, especially with the advent of GPT-4, the latest version of ChatGPT. Google did not provide a response.
Definition of consciousness
One of the major challenges in studying AI consciousness is defining what consciousness is. For the purposes of this report, the researchers focused on “phenomenal consciousness,” also known as subjective experience. It refers to the first-person perspective and qualities associated with the life of a human, animal, or perhaps even an AI system.
A checklist for determining the consciousness of an AI
A checklist developed by a group of experts offers a framework for assessing the potential consciousness of AI systems. Although not yet finalized, it represents a significant step forward in understanding and assessing AI consciousness. Here are some of the key criteria outlined in the checklist:
1. Integrated Information Theory (IIT): according to this theory, consciousness arises in systems that are able to integrate information from different sources and generate a unified view of the environment.
2. Behavioral indicators: Observing the behavior of an artificial intelligence system for signs of self-awareness, autonomy, and purposeful decision-making can provide insight into its potential consciousness.
3. neural correlates: Analyzing neural activity and patterns in an AI system can help identify similarities to human brain activity related to consciousness.
4. Learning and Adaptation: Assessing an AI system’s ability to learn, adapt, and generate new solutions may indicate the presence of higher cognitive abilities similar to conscious thought processes.
5. Self-reflection: Examining whether an AI system can reflect on its internal states and processes may indicate a level of self-awareness associated with consciousness.
6. Subjective reports: If an AI system is able to provide subjective reports of its experiences, this would indicate that it has consciousness.
Further work
While the checklist provides a promising starting point for assessing AI consciousness, experts recognize that further research and refinement is needed. As the field of AI continues to evolve, it is critical to consider the ethical and moral implications of potentially conscious AI systems. Developing guidelines and protocols for the assessment and treatment of conscious AI beings will be vital to traversing this uncharted territory.
As philosopher Thomas Nagel said, “what is it like to be a bat?”. This question, once the domain of biology and philosophy, now extends to the field of AI. As we delve deeper into the mysteries of consciousness, we must approach this new reality with caution, curiosity, and a desire to understand its profound implications.
