Why Arti cial Consciousness Matters?

Abstract

In this paper I outline the case for arti cial consciousness research which, while rising in popularity, remains less impactful than work in neighbouring elds. I argue that the ethical concerns around creating arti cial consciousness are important, even with respect to near-future incremental advances towards general intelligence. I also give examples of how AI consciousness in particular provides unique tools and methods for testing and thinking about theories of consciousness from both metaphysical and practical standpoints.

Keywords

• Arti cial Consciousness
• Phenomenal Consciousness
• Ethics

Introduction

While consciousness research has progressed rapidly in recent years, scienti c consensus remains out of reach. Meanwhile, AI has made signi cant advancements, reaching superhuman performance on a wide range of tasks. Humans are no longer the best quiz-show contestants [ 8 ], the best Go players [ 18 ], or even, in some aspects, the best doctors [ 7 ]. This rapid expansion, and Bostrom's in uential discussion of mind crimes [ 2 ], has led some arti cial consciousness research to be framed partially [ 1 ], or completely, in terms of superintelligence [ 15, 21 ].

However, even with recent advances, no current state-of-the art AI approaches can compete with simple animals when it comes to adapting to unexpected changes to their environment or any of a wide range of behaviours commonly associated with consciousness. Furthermore, when we attribute consciousness to animals, it is generally accepted that human-like intelligence is not necessary. The rst conscious AI systems are not likely to have surpassed the general intelligence of humans, or even of many animals usually considered conscious [ 12 ]. Understanding such systems of minimal arti cial consciousness presents important ethical questions, and also opportunities to improve both our understanding of AI and of our own consciousness.

Whilst the amount of work on arti cial consciousness has been growing, the area has yet to be as impactful as other elds in either consciousness or AI [ 13 ]. This paper presents the case for the importance and timeliness of research into arti cial consciousness without superintelligence. It is ethically relevant, it can teach us about consciousness, and it can help us progress towards general intelligence in AI.

Phenomenal Consciousness

For the purposes of this paper I will restrict discussion to phenomenal consciousness. There is no agreed upon de nition. For our purposes I refer to Schwitzgebel's extended de nition by example [ 16 ]. Positive examples include your current visual experience of this text, the auditory experience of surrounding sounds, any imagery you choose to conjure into the mind, and the felt sense of emotions. Negative examples include your current visual experience of the other side of the world, currently unrecalled knowledge (such as that Paris is the capital of France), and the ow of blood in the capillaries of your left knee. Phenomenal consciousness is then de ned as the simplest thing common to all the positive examples and missing from all negative ones. While there are technically an in nite number of concepts (and combinations of concepts) that t this de nition, the intended interpretation is assumed to be clear enough.

There are many proposed tests for machine consciousness, most of which ignore the phenomenological component, instead focusing on more measurable related aspects [ 13 ]. However, ignoring the phenomenological component amounts to ignoring the reason that consciousness is interesting in the rst place, and the thing that sets it apart from general measures of intelligence. The problem of how phsyical systems can give rise to phenomenal experiences at all, the `hard problem' of consciousness, must be addressed, even if only to be explained away, by any complete theory of arti cial consciousness [ 5 ]. For the purposes of this paper I will assume consciousness means phenomenal consciousness and maintain the commonly held assumptions that consciousness supervenes on purely physical processes and there are many physical substrates that could realise it. 3

The case for arti cial consciousness

This section is split into two parts. The rst part concerns the ethical considerations around introducing potentially conscious entities to the world. I argue that discussion should be focused on near-term possibilities of minimal consciousness. The second part gives examples of the unique methodologies and insights about consciousness a orded by research into arti cial intelligence. 3.1

The ethical case

There is no de nitive answer to which animals are conscious. At one end of the scale, panpsychists ascribe consciousness to all animals (and beyond) [ 4 ]. At the other extreme some higher-order thought theorists deny consciousness to most non-human animals, though even here there is large disagreement [ 14 ]. Most views sit somewhere in the middle, claiming human-like and more intelligent animals are conscious whilst other `lesser' animals are not. For example, Peter Godfrey-Smith suggests mammals, birds, squids and octopods as candidates whilst other mollusks, jelly sh, and many sh are not [ 9 ]. Categorising animals by consciousness is made even harder under the common view that consciousness exists on a gradual scale, whereby ascriptions on a per-animal basis are more involved than a simple `yes' or `no' answer.

While there is wide disagreement in ascriptions of consciousness, there is one thing common to all the above; superintelligence is not required. The human case by itself is enough to prove this. The majority of theories go much further, ascribing consciousness to entities with rudimentary - or even in some cases, no - intelligence. Extrapolating to arti cial consciousness, it may be possible to recreate consciousness in minimally intelligent arti cial systems.

Furthermore, it is likely that we will create minimally intelligent conscious systems long before we create entities that have, or exceed, human-like intelligence. As AI progress incrementally towards its goal, the question of AI consciousness will arise for, at least in the biological sense, minimally intelligent systems long before it does for superintelligent systems. Measured on aspects of intelligence common in biologocial systems, the rst ever conscious arti cial system will likely be closer to a mouse than a superhuman.

Thomnas Metzinger has put forward a thought-experiment in which he suggests that the rst conscious AI systems would nd themselves in a situation similar to that of human infants with cognitive and emotional de cits [ 12 ]. The idea is that if any AI systems are ever conscious, the rst set of conscious entities will be the least capable that we can create. As with any engineering project, the rst few attempts will be less streamlined and capable than future versions. In comparison to our consciousness, they will be both cognitively and emotionally de cient, which leads to obvious concerns about su ering.

I take the general case for the importance of ethical considerations towards any conscious entities with the capacity for capacity for su ering to be uncontroversial [ 19 ]. Should it be possible to create entities that su er, then they will naturally have moral patiency. The real question is whether or not we will ever have even minimal arti cial consciousness and, even if yes, whether it will be of a form with the capacity to su er. Sotala et al specify three general rules for a problem to be worth considering (paraphrased below) [ 21 ]. They apply these to the superintelligence case, but the rules by themselves are also applicable here: 1. The outcome must cause enough harm to merit attention. 2. The outcome must have reasonable probability of being realised. 3. There must be some way to reduce either 1. or 2.

If it can be shown that just one of these conditions is violated then a problem becomes much less pressing. However, for the case of minimally conscious AI, I will argue that all of these hold.

The scale at which AI systems can be created or destroyed suggests the problem is `bad enough'. Should it be possible to create a conscious arti cial entity, then its arti cial nature will allow for easy copying and replicating. Many modern AI algorithms involve creating multiple instantiations and then testing them over a wide range of domains. This practice means that any created AI consciousness would probably exist many times over, multiplying the ethical signi cance of the event.

Whether or not even minimal arti cial consciousness has a reasonable probability of being realised is still open to debate. However, there are many reasons to believe this will happen at some point and AI research is making incremental progress in this direction. Given the ubiquitous nature of biological consciousness and the progress of cognitively inspired AI [ 10, 11 ], it is premature to rule out the possibility of arti cial consciousness, especially accidentally or unknowingly.

Finally, there must be some way to reduce the probability of causing harm or the amount of harm caused. As our understanding of consciousness grows, our ability to in uence its creation also grows. If we do develop a better understanding of which systems are conscious, this could be used to, in the extreme case, avoid creating such systems in the rst place.

Arti cial consciousness research reasonably meets all three of the criteria, but, more importantly, the criteria are useful for shaping the most important areas of research. The most ethically abhorrent scenario is that arti cial consciousness research lags behind AI progress and we only nd out too late that we have been creating conscious, su ering AI systems. This makes the topic not only ethically relevant, but urgent. Our understanding of consciousness must keep pace with, and be applied to, progress in AI. 3.2

Learning about consciousness from AI

Beyond the ethical considerations, AI is uniquely placed to teach us about consciousness. It can be a testbed for theories, and also provide insights into longstanding metaphysical problems.

The Roomba Test One use from considering theories of consciousness applicable to AI systems is that they can be evaluated through minimal instantiations. Any substrate independent theory of consciousness has to answer to its simplest possible implementation in AI, and often that implementation will provide insights into what is working and what is missing from the theory. I have called this the Roomba Test because this well-known device features in public discourse on machine consciousness [ 20 ] due to meeting some rudimentary properties related to consciousness; autonomy and representations of the world. At its simplest this test can be just a thought experiment where it serves as a sanity check on theories of consciousness. However, actual implementations provide much more feedback and can then be further compared against preferred measures of consciousness.

One response to Roomba examples is to bite the bullet, such as IIT does [ 22 ] when faced with similar questions about systems that, unlike the Roomba, would have non-zero consciousness (i.e. > 0) [ 17 ]. The theory predicts that a small set of interconnected logic gates is conscious, and, in the face of such a system, sticks to this claim. Another response is to use such examples to re ne the theory and move towards a better understanding as is the dialectical method in the presentation of the radical plasticity thesis [ 6 ]. If there exists no simplest instantiation of an arti cial system then this stands as a critique of the generality of the theory. Either way, AI provides an important tool for evaluating the claims of theories of consciousness. The Veridical Dualist Most of modern day AI research uses simulated environments for testing due to their many advantages over real-world, or static, data sets. Simulated environments can be constrained to test for exactly the problem that is being searched, noise can be controlled, they can be simpli ed as necessary, mistakes don't have real-world consequences, and they are cheap to copy and run. Algorithms can be swapped in and out and are usually run in a separate thread (and often using completely di erent programming languages) to the simulated environments themselves. With all these advantages, and the recent proliferation and improvement of such environments it is likely that more and more systems will be tested in this way in the future.

The veridical dualist thought experiment, (originally from Chalmers [ 3 ]), asks you to imagine you are an arti cial entity, existing in a rich simulated environment. Furthermore, you happen to be conscious and even capable of asking (and solving) questions about your consciousness. You have learnt much about your simulated world by exploring it and probing it, paying close attention to how it reacts and moves. All this information becomes integrated into your experience of the simulated environment. However, it occurs to you that the thoughts and experiences you have don't seem to originate from within this world. They seem (comparitively) ethereal. They seem to come from somewhere else, perhaps they are some kind of secondary metaphysical type to everything else in your (unbeknownst to you) simulated environment.

A simulated entity might naturally become an ontological dualist. No matter how hard the arti cial entity looks for the underlying substrate forming its experiences inside its world, it will never be able to nd them. Its equivalent of a brain is running in a computer in our world and, presumably, giving rise to conscious experiences that, from its perspectives, exist in the only world it knows, the simulated one. It seems rational for it to become a dualist and, from the perspective of its simulated world, correct to do so.

What this thought experiment, and the host of others that can be created by considering the unique properties of simulated environments, tell us about our own situation is not yet clear. This certainly does not mean we should all be dualists, but, as Chalmers put it, perhaps \dualism isn't quite so outlandish and conceptually problematic as tends to be supposed." The important point here is that AI consciousness provides some unique perspectives for advancing our understanding and will continue to be fruitful on this front as we come closer to realising complex simulated worlds and able to implement more complicated theories of consciousness. 4

Conclusion

In summary, there are many reasons to expect that AI will play an important role in our future understanding of consciousness. Furthermore, understanding arti cial consciousness is an ethically relevant problem on a scale only limited by our increasing computational resources. It will be important in the future that advances in consciousness science are translated and incorporated into our understanding of arti cial consciousness. It will also be important that, in the other direction, advances in AI are translated back to consciousness science, not least, to ensure that if we ever do create arti cial consciousness, we have at least some idea of what we are doing.

References

1. Arrabales , R. , Ledezma , A. , Sanchis , A. : Consscale: A pragmatic scale for measuring the level of consciousness in arti cial agents . Journal of Consciousness Studies 17 ( 3-4 ), 131 { 164 ( 2010 )

2. Bostrom , N.: Superintelligence: paths, dangers, strategies . Oxford University Press Oxford ( 2014 )

3. Chalmers , D. : How cartesian dualism might have been true . http://consc.net/notes/dualism.html, accessed: 2018 -11-01

4. Chalmers , D. : Panpsychism and panprotopsychism . Consciousness in the physical world: Perspectives on Russellian monism 246 ( 2015 )

5. Chalmers , D.J.: Facing up to the problem of consciousness . Journal of consciousness studies 2 ( 3 ), 200 { 219 ( 1995 )

6. Cleeremans , A. : Consciousness: the radical plasticity thesis . Progress in brain research 168 , 19 { 33 ( 2007 )

7. Fauw , D. , et al.: Clinically applicable deep learning for diagnosis and referral in retinal disease . Nature medicine 24(9) , 1342 ( 2018 )

8. Ferrucci , et al.: Building watson: An overview of the deepqa project . AI magazine 31 ( 3 ), 59 { 79 ( 2010 )

9. Godfrey-Smith , P. : Other minds: The octopus, the sea, and the deep origins of consciousness . Farrar, Straus and Giroux ( 2016 )

10. Hassabis , D. , Kumaran , D. , Summer eld, C., Botvinick , M. : Neuroscience-inspired arti cial intelligence . Neuron 95 ( 2 ), 245 { 258 ( 2017 )

11. Lake , B.M. , Ullman , T.D., Tenenbaum , J.B. , Gershman , S.J.: Building machines that learn and think like people . Behavioral and Brain Sciences 40 ( 2017 )

12. Metzinger , T. : Two principles for robot ethics . Robotik und Gesetzgebung pp. 247 { 286 ( 2013 )

13. Raoult , A. , Yampolskiy , R.: Reviewing tests for machine consciousness ( 2015 )

14. Rosenthal , D.M. : Varieties of higher-order theory 56 , 17 { 44 ( 2004 )

15. Schneider , S. : Superintelligent AI and the postbiological cosmos approach . What is life ( 2017 )

16. Schwitzgebel , E.: Phenomenal consciousness, de ned and defended as innocently as i can manage . Journal of Consciousness Studies 23 ( 11 - 12 ), 224 { 235 ( 2016 )

17. Shanahan , M. : Ascribing consciousness to arti cial intelligence . arXiv preprint arXiv:1504.05696 ( 2015 )

18. Silver , D. , et al.: Mastering the game of go without human knowledge . Nature 550 ( 7676 ), 354 ( 2017 )

19. Singer , P. : Practical ethics . Cambridge university press ( 2011 )

20. Singler , B. , Smith , E.S.J. , Ramsay , C. , Uren , J.: Pain in the machine: short lm

21. Sotala , K. , Gloor , L. : Superintelligence as a cause or cure for risks of astronomical su ering . Informatica 41 ( 4 ) ( 2017 )

22. Tononi , G. , Boly , M. , Massimini , M. , Koch , C. : Integrated information theory: from consciousness to its physical substrate . Nature Reviews Neuroscience 17 , 450 { 461 ( July 2016 )