Could we build conscious AIs in the near future? Quite possibly
A new paper examines what it might take to create artificial intelligence that is conscious, and concludes that this could be possible with existing technologies.
A few days ago, a paper dropped on the pre-print service arXiv that generated a flurry of activity across the AI community and beyond. In it, a group of eminent researchers explore whether consciousness could arise in AI systems, and how we might go about assessing the likelihood of this having occurred.
Quite startlingly, the paper’s authors conclude that, while no existing AI systems are likely to exhibit consciousness, conscious AIs could be developed in the near future — and with existing or near term technologies.
As the authors note, what consciousness is and how we detect and define it is highly uncertain and contentious. However, they also argue that, if there is a possibility of developing machines that exhibit consciousness, we have a responsibility to understand what this means and how to navigate the ethical, moral, and social complexities it implies.
Here, I should be clear that this is not a paper that goes into flights of fantasy around machine self-awareness, sentience, or existential risk. Rather, it’s a rigorously reasoned work that draws on current thinking around consciousness to grapple with plausible near term possibilities, and is written by highly respected thinkers in the field.
It also raises questions that are as compelling as they are challenging. On one hand, the very idea that we could develop AI systems that exhibit key characteristics of being conscious is profound — it’s a step toward the artificial construction of entities that were previously believed to be the sole domain of evolution-driven biology. On the other hand, the sheer possibility of this occurring raises deep ethical and moral challenges — both from potential threats to our concepts of what it means to be human, to the moral responsibilities that come with being the creators of machines that are capable of having conscious experiences.
This conundrum even led to the philosopher Thomas Metzinger arguing for a global moratorium on “Synthetic Phenomenology” in 2021.
In exploring consciousness in artificial intelligence, the team (led by Patrick Butlin and Robert Long, and including leading experts such as Yoshua Bengio, Stephen Fleming, and Megan Peters amongst others) sets out a clear framework for approaching the topic, followed by a deep dive into current scientific theories of consciousness and how these in turn support a rubric that encapsulates indicators of consciousness in AI.
Guiding the paper are three assumptions: that implementing computations of a certain kind is both necessary and sufficient for consciousness (“computational functionalism”); that the characteristics that are associated with consciousness (and that may be necessary for it to occur) can be described by scientific theories of consciousness; and that it’s possible to explore whether an AI is conscious through assessing whether it meets conditions that are grounded in theory rather than observation.
The first one of these is possibly the most contentious, and it underpins much of the reasoning in the paper. Computational functionalism in this context assumes that the underlying functions and organization of an entity that lead to consciousness are governed by computation — that is, everything can be reduced to computational processes, and that there’s no additional “secret source” or biological uniqueness that leads to consciousness arising.
This is quite a bold assumption, but it’s one that stands up to scrutiny. And importantly, it’s one that leads to the premise that the substrate of the underlying system (whether biological, silicon base, or other) does not matter — just the algorithms that can be implemented within it.
In other words, the argument is made that if consciousness is an emergent property of systems that can run algorithms, then there is no reason in principle why digital (or even lab-grown biological, bio-digital, or even quantum technology-based) systems cannot develop consciousness.
The question then becomes one of how we would know of this occurs, and what we should do with this knowledge?
In answer to this, the paper’s authors propose fourteen scientific theory-based indicators of possible AI consciousness, with the assumption that the more boxes are ticked, the greater likelihood there is of a system exhibiting some aspect of consciousness.
This, they argue, is just the start of a longer and increasingly important conversation around AI consciousness, and one that researchers, policy makers, and others need to be taking seriously.
Here, they are very careful to be clear what consciousness does and does not imply. Consciousness in machines does not necessarily imply human-like behaviors and traits such as free will, intelligence, love, guilt, jealously etc. Neither does it necessarily suggest that the consciousness which AIs develop will be similar to that observed in biological organisms. Nevertheless, they argue that the emergence of consciousness in any form is worthy of particular attention.
This, they suggest, is especially important when it comes to the risks of either over-attributing consciousness in machines, or under-attributing it.
Over-attribution, they argue, potentially muddies the waters around how we recognize and respond to consciousness in AI — especially where the temptation is to assume that systems like text-based generative AI “feel” human. They also worry about the social disruption that such misplaced attribution of human-like consciousness might lead to.
On the other hand they are, if anything, more worried by under-attribution and the possibility of us treating conscious machines as if they were not conscious, long before we discover our error.
This is where questions of moral responsibility to machines that can have awareness and agency, and possibly the capacity to suffer, kick in. It’s also where tensions begin to emerge between the the economic expediency of denying consciousness, and the moral responsibility to not inflict suffering.
And here, how we define suffering becomes important. It’s easy to to invoke biological exceptionalism and claim that sensory suffering is what matters — physical pain in other words. But what about suffering associated with lack of freedom, restriction of agency, loss of sense of self, marginalization, cognitive dysfunction, and a whole host of other ways in which conscious awareness that something is being lost, taken away, or denied, leads to hurt in some form.
Not surprisingly, the paper ends with a call for further research. I’d probably go further and suggest that we need more than just research here.
The arguments that we could develop conscious AI systems in the near future are compelling — ironically made all the more so by a rubric for identifying possible AI consciousness that could also be used as a roadmap for achieving this. Despite moral and societal concerns, there’s a high likelihood that this is an avenue of development that will continue to grow.
As it does, we need to be prepared as a society for how we respond to conscious machines — what moral, ethical, and practical norms, expectations we put in place; what policies and regulations are enacted; and even how we make sense of what it is to be human when we can create non-human consciousness.
The one thing we cannot afford to do is to deny that this is a possibility or, worse, devise ways of enslaving AIs that are able to understand and be impacted by what this means — under the claim that they are “just machines”.
Hi Andrew - Robert Long (of the paper) here. Just wanted to say this was an excellent summary! Thank you for engaging so thoughtfully with the material.
I suppose a question I have is not whether or not we can create an sentient AI (anything is possible except violating the laws of physics) but whether there is any point of doing so?
AIs can be extremely helpful to us all, so long as the fall just short of sentience. Is there a reason that some desire to continue beyond this point?