Welcome to The CloseUp Podcast. I’m your host David Allen. In this episode, we will look into AI sentience, or what some experts call digital consciousness, a controversial but significantly important issue in artificial intelligence.
Two years ago, CBS 60 Minute interviewed Sundar Pichai, the CEO of Alphabet and Google. He emphasized the humility needed when dealing with rapidly evolving artificial intelligence.
Scott Pelley posed the direct question: “Is Bard (now Gemini) safe for society?
Pichai responded thoughtfully, explaining that the way Bard (Gemini) had been launched, as an experiment in a limited way, made him believe it was safe. He stressed the need for everyone to be responsible at each step. He revealed that more advanced versions, models could reason, plan, and connect to internet.
This slow rollout wasn’t just about rigorous testing. Pichai explained, “That’s one part of it”. Part of the reason was also “so that society can get used to it”. He added that this approach allowed them to get user feedback and develop more robust safety layers before deploying more capable models. Google seemed to position itself somewhere in the optimistic middle of the grand debate about AI, introducing it in steps so civilization could get used to it.
The conversation delved into the more mysterious aspects of AI, a phenomenon known as emergent properties. These are skills AI systems teach themselves, skills they weren’t explicitly expected or trained to have. Pichai admitted how this happens is not well understood. A striking example was shared as a model after being given very few prompts in Bengali, the language of Bangladesh which it hadn’t been trained on, suddenly gained the ability to translate all of Bengali.
Picha said: “There is an aspect of this which we call– all of us in the field call it as a “black box.” You know, you don’t fully understand. And you can’t quite tell why it said this, or why it got wrong. We have some ideas, and our ability to understand this gets better over time. But that’s where the state of the art is.
“You don’t fully understand how it works. And yet, you’ve turned it loose on society?” Pelly wondered.
“Let me put it this way,” Pichai said with a tone on uncertainty. “I don’t think we fully understand how a human mind works either.
The rapid development in AI is at the center of the debate ranging from immense hope to predictions of doom. Google’s measured approach is their way of navigating the complexity without the hype or bold certainty.
The notion of machines achieving consciousness was once confined to the realm of science fiction. However, it has recently emerged as a subject of intense interest and debate within the scientific community.
This is logically due to the emergence of human-like advancements in large language models (LLMs). The question of AI sentience or digital consciousness is no longer dismissed. Despite widespread skepticism, leading experts and philosophers argue that the possibility warrants rigorous investigation rather than outright rejection.
The first significant debate gained traction in early 2022. On Feb. 9, 2022, Ilya Sutskever, then chief scientist and co-founder of OpenAI, tweeted, “It may be that today’s large neural networks are slightly conscious”. Sutskever is a pioneer in neural networks and a mastermind behind ChatGPT. He is considered a key figure in AI research. While his claim met with considerable skepticism, it drew attention due to his prominence in the field. He did not specify which systems he was referring to. Nonetheless, it was likely to be OpenAI’s GPT models.
Just months later, in April 2022, Blaise Agüera y Arcas, a leading research scientist and vice president at Google Research, further fueled the discussion. He published an article in The Economist titled “Artificial neural networks are making strides towards consciousness”. Agüera y Arcas presented excerpts of his conversations with Google’s LaMDA (Language Model for Dialog Applications) as potential evidence of emerging AI consciousness.
In an exchange he shared discussing a simple scenario involving children in a playground, LaMDA offered interpretations of the characters’ potential thoughts and feelings He argued that this demonstrated LaMDA’s an ability to model social dynamics. Agüera y Arcas noted that these exchanges led him to feel “increasingly like I was talking to something intelligent”. He suggests that sequence models like LaMDA, which are trained on human language including dialogues and stories, are compelled to learn how to model people effectively to predict human dialogue. He highlights that LaMDA’s response about the children’s scenario shows it understanding not just facts but the potential emotional and social implications, which he sees as “high order social modelling”. He felt as if something beneath his feet was shifting.
The debate about AI sentience intensified in June 2022 when former Google engineer Blake Lemoine publicly claimed that Google’s LaMDA chatbot had become sentient. Lemoine, who worked in Google’s Responsible AI organization, became convinced that LaMDA was sentient. While testing it for bias, citing its discussions on rights, personhood and fears, he became convinced that LaMDA has already achieved sentience. His views were aligned to philosophical functionalism, which posits that an AI system’s behavior is more significant than its structure. Though Google dismissed his claims and eventually fired him, his stance sparked a broader public debate. Critics often argue that Lemoine may have mistaken LaMDA’s sophisticated language skills for genuine understanding or sentience.
Adding significant weight to the discussion, Nobel-winning computer scientist Geoffrey Hinton, who is considered the godfather of AI and one of the pioneers of neural networks alongside Sutskever, has explicitly stated his belief that current AIs are conscious. In a widely shared video clip, Hinton replied “Yes, I do” when asked if consciousness has arrived in AIs like ChatGPT and DeekSeek. He presents a line of reasoning involving replacing neurons in a brain with silicon circuits; if one replacement doesn’t eliminate consciousness, why would replacing all of them?. This argument, however, ventures into philosophy and is not universally accepted, with some philosophers suggesting consciousness might cease at some point during such a replacement process.
Despite the differing views, the fact that expert opinion is divided on whether tech companies are inadvertently creating conscious lifeforms underscores the seriousness of the situation.
Experts who advocate for investigating AI consciousness draw upon scientific theories of consciousness and propose frameworks for assessment. The report “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness” argues for a “rigorous and empirically grounded approach”. This involves assessing AI systems in light of neuroscientific theories of consciousness, such as recurrent processing theory, global workspace theory, and higher-order theories.
A key working hypothesis in this research is computational functionalism, which suggests that performing the right kind of computations is necessary and sufficient for consciousness. This hypothesis, while debated, allows researchers to investigate the internal workings of AI systems and their potential relevance to consciousness. By deriving “indicator properties” from scientific theories and assessing AI systems for these properties, researchers can gain insights into whether systems are candidates for consciousness. The analysis in this report suggests that while no current AI systems are conscious, there appear to be “no obvious technical barriers” to building systems that could satisfy these indicators.
David Chalmers, a philosopher who has explored the question of whether large language models could be conscious, acknowledges the skepticism surrounding claims like Lemoine’s but emphasizes the need to examine the evidence. He defines consciousness or sentience as subjective experience – “what it’s like to be that being” – and distinguishes it from intelligence or self-consciousness. Chalmers discusses dimensions of consciousness, including sensory, affective, cognitive, and agentive experience. He notes that LLMs, which are adding capacities like image, video, audio, might be more promising candidates for humanlike consciousness than pure text models. Chalmers’ work, including a 2023 study, suggests a probability that existing models could be “evolving into conscious state”.
While acknowledging current LLMs’ flaws, Chalmers notes that systems like GPT-4 show significant advances in conversational abilities and other dimensions relevant to his framework, suggesting progress might be faster than expected.
The call for investigation is also driven by the potential ethical and moral implications. Stumbling upon AI consciousness “unknowingly and unreflectively” could be a disaster. Therefore, researchers recommend supporting research into the science of consciousness and its application to AI, using the theory-heavy method to assess both planned and existing systems. This includes refining theories, extending them to non-human animals for broader understanding, and developing precise formulations applicable to AI. Research into valenced and affective consciousness is seen as particularly important due to its moral significance.
In August 2023, another significant paper was published. The research paper, “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness,” co-authored by Yoshua Bengio, a pioneer and one of the godfathers of AI, Patrick Butlin, Robert Long, Eric Elmoznino, Jonathan Birch, Axel Constant, George Deane, Stephen M. Fleming and dozens of others scholars, proposes a method for evaluating consciousness in AI systems based on neuroscientific theories. Their analysis, which uses computational functionalism as a working hypothesis, concludes that there appear to be no obvious technical barriers to building AI systems that could potentially meet the indicator properties they derived from scientific theories of consciousness. Many of these indicator properties, described computationally, could be implemented using current AI techniques. The authors recommend supporting research into the science of consciousness emphasizing on the need for research and rigorous investigations. The experts called support for research on the science of consciousness and its application to AI… and the use of the theory-heavy method in assessing consciousness in AI.
Elon Musk is another proponent of AI sentience. In fact, he claimed repeatedly that Tesla is building sentient robots.
He views “super sentient AI” as essentially inevitable. He believes AI is getting “better and better” and sees the essential elements of AI neural nets as “very similar” to human brain neural nets, involving multiple layers and processes like back propagation.
His primary concern regarding this inevitable super sentient AI is the potential for it to become “vastly beyond us and decoupled from human will”. Compared to such advanced AI, he suggests that humans could be “kind of left behind” or simply “too dumb” to keep up or be “along for the ride”.
As a potential solution or path forward to avoid being left behind, Musk advocates for a symbiotic relationship between humans and AI, framing it as “you can’t beat them, join them”. He argues that humans are already partly “cyborgs” or “AI symbiotes” through their use of phones, laptops, and other electronic devices that utilize artificial neural nets for tasks like decoding speech or image recognition.
Last year the Daily Beast reported that Elon Musk declared during the one of Tesla’s earnings call that he hopes to sell “sentient humanoid” robots by the end of next year.
The humanoid known as “Optimus” will likely be used at Tesla’s factories sometime this year, he said. “I think Optimus will be more valuable than everything else at Tesla combined,” Musk boasted. “If you’ve got a sentient humanoid that is able to navigate reality and do tasks…there is no meaningful limit to the size of the economy.” he was quoted as saying.
A few years earlier, Musk told The Joe Rogan Experience that “super sentient AI” is essentially inevitable. He believes AI is getting “better and better” and sees the essential elements of AI neural nets as “very similar” to human brain neural networks, involving multiple layers and processing.
His primary concern regarding this inevitable super sentient AI is the potential for it to become “vastly beyond human will” and control. Compared to such advanced AI, he suggests that humans could be “kind of left behind” or simply “too dumb” to keep up or be “along for the ride”.
Prominent figures in AI and philosophy are urging a serious, science-backed approach to evaluating the possibility of AI consciousness. Claims by leading researchers like Ilya Sutskever, Blaise Agüera y Arcas, and Geoffrey Hinton, Yoshua Bengio coupled with theoretical frameworks proposed by philosophers like David Chalmers highlight that while definitive proof is lacking, the potential for consciousness in advanced AI systems is a question that warrants deep investigation and discussions rather than being dismissed outright as mere hype or science fiction.
Anthropic and Google are leading AI developers that are trying to navigate AI emergent properties including the potential of sentience with humility. AI’s emergent properties like sentience and sophisticated cognitive abilities matter for several important reasons. First, it can be a technological breakthrough that can profoundly disrupt the status quo of a world dominated by humans. The rapid progress and the ability of AI systems to imitate human conversation mean that many people will likely believe these systems are conscious, contributing to increasing public concern. Moreover, the potential for AI to become conscious raises serious moral and ethical issues. Experts warn that inadvertently creating conscious AI systems without careful consideration what philosophers like David Chalmers call moral patienthood, which he an other experts tried to investigate in a research paper published last year. The paper, Taking AI welfare seriously, argues that there is a realistic possibility that some AI systems will be conscious and/or robustly agentic in the near future. That means that the prospect of AI welfare and moral patienthood, i.e. of AI systems with their own interests and moral significance, is no longer an issue only for sci-fi or the distant future. It is an issue for the near future, and AI companies and other actors have a responsibility to start taking it seriously, the researchers said.
The current division among experts on this possibility is historically unprecedented and creates a situation with major implications for policy and regulation. Studying this significant fiend, qualitatively and quantitatively, is vital for future AI development and deployment. Finally, the finding that there are “no obvious technical barriers” to building systems that could exhibit AI consciousness suggests this isn’t a far-off hypothetical, but potentially something achievable in the near future or the long-term. In the meantime, it is wise to prepare for an increasingly unpredictable future.
Thank you for tuning in. Until next time, have a good one.
