Макс Тегмарк - Life 3.0 - Being Human in the Age of Artificial Intelligence

Another controversial IIT claim is that today’s computer architectures can’t be conscious, because the way their logic gates connect gives very low integration.24 In other words, if you upload yourself into a future high-powered robot that accurately simulates every single one of your neurons and synapses, then even if this digital clone looks, talks and acts indistinguishably from you, Giulio claims that it will be an unconscious zombie without subjective experience—which would be disappointing if you uploaded yourself in a quest for subjective immortality. *6This claim has been challenged by both David Chalmers and AI professor Murray Shanahan by imagining what would happen if you instead gradually replaced the neural circuits in your brain by hypothetical digital hardware perfectly simulating them.25 Although your behavior would be unaffected by the replacement since the simulation is by assumption perfect, your experience would change from conscious initially to unconscious at the end, according to Giulio. But how would it feel in between, as ever more got replaced? When the parts of your brain responsible for your conscious experience of the upper half of your visual field were replaced, would you notice that part of your visual scenery was suddenly missing, but that you mysteriously knew what was there nonetheless, as reported by patients with “blindsight”?26 This would be deeply troubling, because if you can consciously experience any difference, then you can also tell your friends about it when asked—yet by assumption, your behavior can’t change. The only logical possibility compatible with the assumptions is that at exactly the same instance that any one thing disappears from your consciousness, your mind is mysteriously altered so as either to make you lie and deny that your experience changed, or to forget that things had been different.

On the other hand, Murray Shanahan admits that the same gradual-replacement critique can be leveled at any theory claiming that you can act conscious without being conscious, so you might be tempted to conclude that acting and being conscious are one and the same, and that externally observable behavior is therefore all that matters. But then you’d have fallen into the trap of predicting that you’re unconscious while dreaming, even though you know better.

A third IIT controversy is whether a conscious entity can be made of parts that are separately conscious. For example, can society as a whole gain consciousness without the people in it losing theirs? Can a conscious brain have parts that are also conscious on their own? The prediction from IIT is a firm “no,” but not everyone is convinced. For example, some patients with lesions severely reducing communication between the two halves of their brain experience “alien hand syndrome,” where their right brain makes their left hand do things that the patients claim they aren’t causing or understanding—sometimes to the point that they use their other hand to restrain their “alien” hand. How can we be so sure that there aren’t two separate consciousnesses in their head, one in the right hemisphere that’s unable to speak and another in the left hemisphere that’s doing all the talking and claiming to speak for both of them? Imagine using future technology to build a direct communication link between two human brains, and gradually increasing the capacity of this link until communication is as efficient between the brains as it is within them. Would there come a moment when the two individual consciousnesses suddenly disappear and get replaced by a single unified one as IIT predicts, or would the transition be gradual so that the individual consciousnesses coexisted in some form even as a joint experience began to emerge?

Another fascinating controversy is whether experiments underestimate how much we’re conscious of. We saw earlier that although we feel we’re visually conscious of vast amounts of information involving colors, shapes, objects and seemingly everything that’s in front of us, experiments have shown that we can only remember and report a dismally small fraction of this.27 Some researchers have tried to resolve this discrepancy by asking whether we may sometimes have “consciousness without access,” that is, subjective experience of things that are too complex to fit into our working memory for later use.28 For example, when you experience inattentional blindness by being too distracted to notice an object in plain sight, this doesn’t imply that you had no conscious visual experience of it, merely that it wasn’t stored in your working memory.29 Should it count as forgetfulness rather than blindness? Other researchers reject this idea that people can’t be trusted about what they say they experienced, and warn of its implications. Murray Shanahan imagines a clinical trial where patients report complete pain relief thanks to a new wonder drug, which nonetheless gets rejected by a government panel: “The patients only think they are not in pain. Thanks to neuroscience, we know better.”30 On the other hand, there have been cases where patients who accidentally awoke during surgery were given a drug to make them forget the ordeal. Should we trust their subsequent report that they experienced no pain?31

How Might AI Consciousness Feel?

If some future AI system is conscious, then what will it subjectively experience? This is the essence of the “even harder problem” of consciousness, and forces us up to the second level of difficulty depicted in figure 8.1. Not only do we currently lack a theory that answers this question, but we’re not even sure whether it’s logically possible to fully answer it. After all, what could a satisfactory answer sound like? How would you explain to a person born blind what the color red looks like?

Fortunately, our current inability to give a complete answer doesn’t prevent us from giving partial answers. Intelligent aliens studying the human sensory system would probably infer that colors are qualia that feel associated with each point on a two-dimensional surface (our visual field), while sounds don’t feel as spatially localized, and pains are qualia that feel associated with different parts of our body. From discovering that our retinas have three types of light-sensitive cone cells, they could infer that we experience three primary colors and that all other color qualia result from combining them. By measuring how long it takes neurons to transmit information across the brain, they could conclude that we experience no more than about ten conscious thoughts or perceptions per second, and that when we watch movies on our TV at twenty-four frames per second, we experience this not as a sequence of still images, but as continuous motion. From measuring how fast adrenaline is released into our bloodstream and how long it remains before being broken down, they could predict that we feel bursts of anger starting within seconds and lasting for minutes.

Applying similar physics-based arguments, we can make some educated guesses about certain aspects of how an artificial consciousness may feel. First of all, the space of possible AI experiences is huge compared to what we humans can experience. We have one class of qualia for each of our senses, but AIs can have vastly more types of sensors and internal representations of information, so we must avoid the pitfall of assuming that being an AI necessarily feels similar to being a person.

Second, a brain-sized artificial consciousness could have millions of times more experiences than us per second, since electromagnetic signals travel at the speed of light—millions of times faster than neuron signals. However, the larger the AI, the slower its global thoughts must be to allow information time to flow between all its parts, as we saw in chapter 4. We’d therefore expect an Earth-sized “Gaia” AI to have only about ten conscious experiences per second, like a human, and a galaxy-sized AI could have only one global thought every 100,000 years or so—so no more than about a hundred experiences during the entire history of our Universe thus far! This would give large AIs a seemingly irresistible incentive to delegate computations to the smallest subsystems capable of handling them, to speed things up, much like our conscious mind has delegated the blink reflex to a small, fast and unconscious subsystem. Although we saw above that the conscious information processing in our brains appears to be merely the tip of an otherwise unconscious iceberg, we should expect the situation to be even more extreme for large future AIs: if they have a single consciousness, then it’s likely to be unaware of almost all the information processing taking place within it. Moreover, although the conscious experiences that it enjoys may be extremely complex, they’re also snail-paced compared to the rapid activities of its smaller parts.