MaCro Philosophy

Five reasons we will never have conscious AI

The question of whether (or when) we will create artificial consciousness is one of the most important of our time. Whether or not AI will be conscious will determine our moral responsibility towards it, and, perhaps, its attitude towards us. In this post I give five reasons to believe that we will never create such entities. This is the second installment in a series, the first part of which argued for the possibility of conscious AI. The points here are intended to answer those in the previous article.

By consciousness I am referring to phenomenal consciousness, and by conscious AI I mean an artificial system for which there is something it is like to be that system. This isn't a completely satisfactory definition, but a better one requires its own paper (or perhaps book). For something more precise I defer to Eric Schwitzgebel's reasonable and practical definition of consciousness here.

1. When AI achieves human capabilities it completely surpasses them. Human-level and human-like AI are a myth.

Previously: AI is achieving, and surpassing, human capabilities in many areas. Being conscious is just another human capability to be achieved (or surpassed).
"Every move AlphaGo plays is surprising and is out of our imagination. We can’t read those kind of moves. Some of the moves that AlphaGo played are not based off our human studies." - Stephanie Yin, 1 Dan Professional

When AlphaGo became the strongest Go player on the planet, it didn't just emulate humane players and then play slightly better. It created new ideas. In fact, emulation went in the other direction. In modern Go, everyone is now diving into 3-3 points at the first chance they get (which used to be considered bad), a strategy which AlphaGo introduced to the world. Humans are emulating AI.

AlphaZero, the successor of AlphaGo that can learn perfect information games from self-play alone, also plays alien-like moves. In this short clip, Demis Hassabis of DeepMind (the creators of AlphaZero) enthusiastically explains that this is "chess from another dimension".

Excerpt from Demis Hassabis's talk at the Kinds of Intelligence Symposium

It is hard enough to understand consciousness in humans, and we have access to our own conscious states and an increasing understanding of how the brain works. How then can we make claims about consciousness in systems very alien to us, that operate in completely different ways?

2. Consciousness is unique to evolved biological entities.

Previously: Consciousness is not rare or unique or special. It is found across different animals with different evolutionary histories.
"Sentience is brought into being somehow from the evolution of sensing and acting; it involves being a living system with a point of view on the world around it." - Peter Godfrey-Smith, Other Minds

Slide from Anil Seth

Perhaps consciousness is common in biological intelligences. That does not mean it is common in artificial intelligences. There are fundamental differences between the biological and the artificial. Anil Seth lists five aspects of being a conscious self, which are closely related to being biological entities.

In particular, having a body, having a narrative, and being part of a social structure, all seem to be important components of having a conscious self. These are currently unique to biological systems and not part of traditional AI.

Another possibility is that the underlying physical properties required for consciousness are not present in the systems we build AI out of. Perhaps consciousness requires certain neuronal, atomic, or even quantum states, that are present in us, but not present in the material we will build our AIs out of.

3. Most of what the brain does that we want to emulate is unconscious.

Previously: Our conscious intelligence is our best source for, and will be emulated in, advancements in artificial intelligence.
"an individual [with blindsight] might point quite accurately towards a visual target in their blind field, despite confidently asserting that they had seen nothing!" Danckert and Goodale, Blindsight: A conscious route to unconscious vision.

The brain is our best known example of general intelligence (and consciousness). While we are not even close to unlocking all its secrets, what we do know has inspired many of the currently popular AI techniques (most notably convolutions and neural networks themselves). Yet, even though the brain is conscious, it does most of its work unconsciously. This is most pronounced in cases of blindsight. In these cases, people who report that they have no visual experience (at least in some regions), are still able to respond (above chance) to stimuli in the area. These cases are rare, and come from damage to certain areas of the brain, but the cases that do exist have been well studied and verified. Information is entering the visual system, being processed, and can be acted upon, but the person never becomes conscious of it.

The video shows a blindsight patient navigating obstacles in a corridor that he cannot see. He claims that he feels a need to move in a certain way, but not because he can see any obstacles. If the brain - our best example of consciousness - can achieve most of its complex tasks unconsciously, then AI, which we have no reason to believe is conscious, can certainly do so too.

4. We will create AI that is most useful to us, not that is most like us.

Previously: We will create AI to solve human problems, to navigate our world, and to interact with us in human-like ways.
"Since the first grain silo ... the first computer ... we've always made things bigger, quicker, longer lasting, different than us." Austin Walker - Friends at the Table.

Excerpt from Friends at the Table, Words by Austin Walker, Music by Jack de Quidt

Our history of technological advances - from clothes, to grain silos, to the printing press, to electricity and space travel - is a history of building things to solve problems our biological bodies cannot solve by themselves. These are all things with properties vastly different from our own. We will create AI to solve our problems and interact with us in our world, but we won't need to create AI to be like humans to do this. In fact, we will want AI to be not-like-us, exactly so that it can do the things we can't do ourselves.

5. We will have the choice of what we build. If it is possible to make conscious AI, and we understand how and why, we can choose not to.

Previously: 5. The evolutionary pressure to procreate has morphed from a drive to pass on our genes, to a desire to pass on our unique conscious perspective of the world.
"Many mourned, and said that a child was a terrible idea to begin with, impossible, under the circumstances, and humanity would do well to remember that eventually, every child replaces its parent." Catherynne M. Valente - Silently and Very Fast

The precautionary principle suggests that one should exercise caution whenever there is uncertainty about possible outcomes, and some of them have a high risk. There are complex questions to address about the ethics and ramifications of introducing new conscious entities to the world (artificial or not). As such, it makes sense to avoid doing so, at least until we understand the situation better. The arguments presented here suggest that this should be achievable.

Ultimately, as long as we keep research on consciousness connected to research on AI, it will be possible to steer the course of research in the appropriate direction.

Find the previous installment here.

In both these posts I've favoured introducing ideas over giving solid arguments. In the next installment I will compare both sides to see which is the most convincing when taken a little more seriously.

Blog Homepage