Have Robots Achieved Consciousness?

At what level are you conscious? Staring into the eyes of a comatose loved one, many of us have agonized over whether the patient was conscious of caresses received or whispered prayers. Increasingly we will have answers to such questions, thanks in a large part to a growing understanding of reinforcement learning, Artificial Intelligence and how they pertain to consciousness.
A pioneering study by Tristan Bekinschtein and colleagues published in Nature Neuroscience demonstrated that some patients in vegetative states still retain some basic reinforcement learning capabilities (Tristan A Bekinschtein, 2009). These patients were subjected to puffs of air delivered to their eyes, paired with some predictive stimulus such as bell ringing – a well-known test of basic reinforcement learning abilities. Though unable to formulate words or respond in any self-conscious manner, these patients nevertheless showed the ability to anticipate the air puffs. Such patients, while all but indistinguishable from other people in vegetative states, were nonetheless conscious at some level and this turned out to be highly predictive of which patients would later regain what we think of as full consciousness. This can help to distinguish which patients should remain on life support and which should be mercifully allowed to pass away. It also helps distinguish between categories of consciousness.
Such studies lend support to a multi-tiered theory of consciousness akin to that proposed by Nobel prize winning biologist and neuroscientist Gerald Edelman. According to Edelman, consciousness exists on a spectrum with at least two divisions – a primary consciousness and a secondary consciousness (Edelman, 2003). Primary consciousness is an awareness that includes perception, emotion, and crucially the ability to connect basic stimuli in a manner that facilitates anticipatory decision making. This is known as Reinforcement Learning. In reinforcement learning, some past experience is used to improve decision making over time to optimize a desired result. This primary consciousness may extend down to the very roots on the tree of life. Dogs, birds and even many amphibians demonstrate basic reinforcement learning capabilities. This implies that at the very seat of primary consciousness is the reinforcement learning system, a system that allows us to distinguish between positive and aversive stimuli in such a way as to learn from it and guide decision making around these past experiences. Whether this also entails artificial agents such as robots and computers possessed of reinforcement learning, such as AlphaZero, should be considered to have primary consciousness is open to debate. It is likely to be a hotly contended topic in the coming decades as Robots and AI expand their reinforcement learning tool chest.
But this is not the final word on consciousness. According to Edelman, distinct from primary consciousness is secondary consciousness, that is, an awareness that we are aware. This kind of metacognition is the type we normally associate with human consciousness, for it also presages the ability to create highly detailed narratives about ourselves as agents distinct from our surroundings. It also likely corresponds with a different type of learning system, one involving structured logic and explicit reasoning. Such explicit reasoning involves making models of the world in abstract form. To be conscious as we normally think of it, we create a model of the world with ourselves as a distinct component in it, separate from the stream of positive and negative stimuli that furnishes our awareness. There is an increasing body of evidence showing that this declarative learning of the type we find in "conscious humans" corresponds to some kind of model-based reinforcement learning, in contrast to the model-free reinforcement learning of animals who exhibit only primary consciousness (Samuel J. Gershman, 2017). In terms of decision making, this model-based reinforcement learning parallels what’s known as "System 2 thinking" outlined by the economist Daniel Kahneman in dual process theory. System 2 thinking is deliberate, effortful and logic based as opposed to the fast intuitive decision making characteristic of model-free reinforcement learning. While less is known about the exact mathematical shape that model-based reinforcement learning algorithms take in the human brain, and importantly, how they interact in real-time with their model-free cousin, this is one of the hottest areas of research in the cognitive sciences at the moment. As such we should brace ourselves for forthcoming explanations of human consciousness very much rooted in mathematics.
Inanimate Consciousness
The problem of defining secondary consciousness is itself a nasty one. In recent years, we have been able to obtain some clues to this old riddle. The answer may be locked up in what neuroscientists call a "theory of mind." The brains of humans does not seem to be unitary, but an amalgamation of many competing wants and desires fueled by different "evolutionary algorithms". Of the many modules or algorithms, one seems to be responsible for creating a theory of mind – an ability to speculate about what another individual is thinking or feeling. Indeed, as one recent study has shown, we unconsciously mirror other people’s facial expressions in order to better "sense" what they are feeling.
By reflexively taking the same expression as another person, we gain useful insights about their mental state. This could even be why some old couples come to resemble each other – by mirroring each other’s facial expressions, the wrinkle lines in their face gradually converge upon a similar pattern. Aiding and abetting this process are mirror neurons, neurons that respond equally when we perform an action as to when we see someone else performing an action. It is as if the brain tries on the other person’s actions as we watch them. Through such mechanisms, we form a theory about another person’s intentions and motivations. Why might such a theory of mind be important? In a highly social animal like ourselves, knowing what another human is thinking could be of great survival benefit. Does this person wish us ill, or, are they friendly and a potential ally?
Perhaps more interestingly, we seem to have a theory of mind about ourselves, a part of our brain that tries to make sense and create a story about our own actions. This is one of the leading explanations for secondary consciousness, that is, a theory of mind turned in upon its owner. As often as not, the beliefs that this self-referencing theory of mind comes up with about the things we do is entirely bogus, or at the very least, disingenuous. We are already familiar with the process under a different name, "denial", or confabulation. Interestingly, denial seems to be more the rule than the exception, as we more or less systematically interpret our actions with erroneous explanation. Some neuroscientists have compared consciousness to a lawyer, constantly defending our actions and making up credible though highly spurious reasons for the things we do.
This comparison to a lawyer hints at why we might need such a cantankerous algorithm in the first place, that is, to justify our actions in a highly social environment. Other animals lacking a self-referencing theory of mind seem to feel neither shame nor denial. While there isn’t a clear answer as yet why humans would have evolved such a bizarre evolutionary algorithm, we have room to speculate. One such speculation originates from what’s called the "red queen effect." In evolutionary terms, the red queen effect results when there is a series of cascading adaptations between two organisms that are locked in a survival struggle. As soon as one organism gets the upper hand in their pitched battle, the other evolves a mechanism to counter it, and the process repeats such that both organisms must keep evolving just as fast as they can to keep the balance between them. Hence, the reference to the red queen from Alice in Wonderland, who explained to the young heroine "here it takes all the running you can do to keep in the same place."
How could the red queen effect have possibly given rise to secondary consciousness? If we assume a theory of mind was developed in order to read the minds of others, the better to predict if they were going to help or hinder us, then there would be a survival edge to be gained by foiling this adaptation and finding a method to deceive it. If a conscious theory of mind can construct an alibi for the things I’m doing, even when they are expressly untrue, this could give me a survival edge in a highly social environment. In the game of Cops and Robbers, if you develop a truth detector, then I need to develop a better lie fabricator, and round and round it goes.
At our present juncture, fake news may be spurring a similar kind of red queen race within humanity, leading us towards hybrid forms of consciousness the merge our biological wetware with machine intelligence. Consider that with the aid of computers, the ability to create false narratives using such tools as Photoshop has made it ever more difficult to differentiate truth from fiction. Spinning false narratives has never been easier, and as a result there is increased need for counter weapons to unravel these yarns. Such countermeasures are also likely to take the form of an algorithm, albeit one residing in a computer. For instance there are AI algorithms that are trained to differentiate real images from ones that have been retouched in Photoshop. Increasingly, it may be necessary for individuals to adopt the use of such AI to differentiate fake news from truthful accounts. In such a manner we may find ourselves put under evolutionary pressure to fold such forms of artificial intelligence into our biological wetware, relying upon them to keep us one step ahead of the lie mongers. Where such back and forth measures are likely to end is difficult to say, but the addition of "hybrid" forms of consciousness to our present biology is not out of the question. Already there have been strides in creating a neural lace by the Elon Musk backed company called Neuralink, which could potentially enable such forms of hybrid consciousness in humans, fusing machine and biology into a single entity.
Another source of competitive pressure that could lead to hybrid forms of consciousness is coming from the marketplace. There currently exists considerable competition to gain admittance to top ranked academic institutions , such that students and job seekers will avail of drugs like Adderall and Ritalin to secure the necessary grades and exam scores to gain access to these highly coveted positions. If an applicant whose consciousness was aided by a neural lace or a CRISPR enabled genetic modification showed a propensity to win access to the best schools and jobs, then it is likely that such adaptations would quickly spread throughout the entire population. The red queen effect can give rise to curious and unforeseen consequences and it could be wise to carefully consider the competitive pressures we are presently subjecting ourselves to, lest they paint us into corners we later find to be distinctly dystopian.
Some recent experiments involve what might be considered "inanimate consciousness," including synthetic agents that possess a theory of mind. A theory of mind has been one measure used in differentiating artificial forms of intelligence from human intelligence. To understand this difference there is a simple experiment called the ultimatum game. In the ultimatum game, one player receives a hundred dollars (or other unit of value) and must decide how they wish to split it with a second player. The second player, can either reject the offer in which case both players get nothing, or can accept it in which case they receive the money according to the division specified by player one. If player one is perfectly rational and lacks a theory of mind, it will offer far less than humans, as it will assume the other player will accept any offer greater than zero because receiving something is always better than receiving nothing. Or so runs the logic. Lacking a theory of mind, player one doesn’t take into account that the other player is likely to feel slighted if only offered a single dollar and therefore reject the proposal outright. While a purely rational agent would be better off accepting the dollar than rejecting it, humans will frequently reject such low ball offers. This has been a thorn in the side of economists and psychologists because the entire dogma of utility theory is based upon the presupposition that people maximize utility and would accept such low ball offers the way a computer playing the game would. Interestingly, in autism, a disease frequently described as "mind blindness" because the afflicted often show a marked inability to read the minds of others by observing facial features, patients tend to behave much more like the purely rational agents than "normal" people. In such games – autistic people will expect others to accept low ball offers and be fine accepting them themselves. This evidence from autism supports the idea that humans are using a theory of mind when they operate in many strategic environments and this influences their decision making.
One reason advanced for humans rejecting low ball offers is that we are accustomed to playing iterated prisoner dilemma games, that is, games that are repeated many times and possess a solution in which, when both players cooperate, they are better off than had they both acted selfishly. If as a player in the ultimatum game, I assume we are living in a world with iterated prisoner dilemma situations, then I have a motive to punish you for making a selfish offer so that the next time we play you will not make a low ball offer again. How might one go about creating an algorithm that uses a theory of mind to guess about the others person’s strategy and thus achieve an outcome more similar to a human one?
One of the groups making progress on this is the OpenAI foundation. In collaboration with researchers at Oxford University, the OpenAI team created a reinforcement learning algorithm that possessed a theory of mind about another player when updating its own strategy in the ultimatum game (Jakob N. Foerster, 2018). The secret ingredient in their recipe was adding a term in the reinforcement learning equation that captures changes in a second player’s strategy, that is – a theory of the other players mind. While the math behind this can get a bit hairy, the principle is simple enough – if I know that you are learning too, then I need to take into account changes in your strategy when I formulate my own. Their approach seemed to work – in an iterated prisoner dilemma game, the algorithm learned reciprocative strategies that far more resemble human players than the strategies of purely selfish agents. This could well mark an important step towards creating artificial secondary consciousness. Assuming for a moment that consciousness is a theory of mind turned in upon itself, then it would be a logical first step to model a theory of mind turned outwards, as in the OpenAI experiment. This certainly seems to be a precursor in nature to more complex theories of mind. For instance, dogs arguably possess a simple theory of mind.
You can observe dogs scanning your features and gestures, trying to guess if you are about to refill their food bowl or take them for a walk. While I know of no such attempts to modify reinforcement learning to include a theory of mind about one’s own actions, it may not be far off. While initially such nascent forms of machine consciousness would hardly be recognizable as similar to our own, perhaps in time they would develop the colorful psychological nuances like personal narratives, shame, denial and projection that characterize our own conscious minds. It is important to note that even were this to happen, it does not presuppose a synthetic consciousness’s would pose a danger to humans. Their goal states might still be pre-determined and their actions limited to virtual tasks in virtual worlds.
As we examine the implications of AI and reinforcement learning it is important to organize and prioritize it’s repercussions. While an uprising of conscious robots might pose an existential threat to humans in the distant future, there are likely to be more proximate challenges, for instance, those resulting from increased job automation leading to ballooning wealth disparities. Only after navigating these near term hurdles are we likely to encounter the sensational horrors betokened in so many science fiction novels and films.