14
$\begingroup$

Consciousness is challenging to define, but for this question let's define it as "actually experiencing sensory input as opposed to just putting a bunch of data through an inanimate machine." Humans, of course, have minds; for normal computers, all the things they "see" are just more data. One could alternatively say that humans are sentient, while traditional computers are not.

Setting aside the question of whether it's possible to build a sentient machine, does it actually make a difference if an AI is sentient or not? In other words, are there are tasks that are made impossible - not just more difficult - by a lack of sentience?

$\endgroup$
  • $\begingroup$ Relevant meta discussion on AI-related philosophy. $\endgroup$ – Ben N Sep 8 '16 at 16:21
  • $\begingroup$ Technically, of course, all the things we "see" are just data, too--electrical impulses traveling over optical nerves. $\endgroup$ – Nathan Arthur Sep 8 '16 at 20:19
  • $\begingroup$ @NathanArthur He's not talking about the things we see but the experience of seeing itself. $\endgroup$ – David Schwartz Sep 8 '16 at 20:43
  • $\begingroup$ @DavidSchwartz Point taken. :) $\endgroup$ – Nathan Arthur Sep 8 '16 at 20:51
  • $\begingroup$ You need to better define "AI" and "task" or if those are to be taken together - define "any AI task". Expert systems are categorized as weak AI and perform their specific tasks quite well. In this instance, the answer is very clear - no, consciousness is not needed to offer a diagnosis given a set of symptoms. $\endgroup$ – CramerTV Sep 8 '16 at 23:07
10
$\begingroup$

No-one knows.

Why: because it's not possible to formally determine even whether your fellow human beings are actually conscious (they may instead be what is philosophically termed a 'Zombie'). No test known to modern physics suffices to decide. Hence it's possible that you are the only sentient being, and everyone else is a robot.

Consequently, we cannot determine which tasks require sentience.

Note that the ontological status of Zombies is controversial: some philosophers of AI (e.g. Daniel Dennett) claim that Zombies are logically impossible while others such as David Chalmers would claim that a Zombie would be compelled to assert that they experience qualia (i.e. are sentient) even though they do not. Here is a very readable paper by Flanagan and Polger that also explains why a stronger neurological version of the Turing test is insufficient to detect a Zombie.

EDIT: In response to the comment about whether an objective test for distinguishing sentience from non-sentience exists:

No-one knows. What we do believe is that this would require something in addition to what modern physics can currently tell us. David Chalmers has speculated that qualia should be introduced as a new form of physical unit, orthogonal to the others in the same way that electrical charge is orthogonal to distance.

In the absence of an objective test, we have to rely on Turing test variants, which no more guarantee consciousness in the subject than they do intelligence.

$\endgroup$
  • $\begingroup$ I think you're taking this from a solipsistic view, and not directly answering the question of 'What tasks could be newly possible IF sentience and the ability to experience was something we could give an AI'; instead you address how sentience works in other beings. $\endgroup$ – Avik Mohan Sep 8 '16 at 16:54
  • 2
    $\begingroup$ That is because (precisely for purposes of discussing sentience), there is no difference between 'allegedly conscious AIs' and 'other human beings'. What tasks can I do that you can't, given that you might be a robot (though it's impossible to tell), while I know that I am not? There are no such tasks. QED. $\endgroup$ – NietzscheanAI Sep 8 '16 at 16:56
  • $\begingroup$ You're taking another assumed-sentient being, saying you're not sure if it's sentient, so then there are no tasks my non-sentience inhibits. So firstly, you do seem to know the answer, it being 'No'; secondly, you're still ignoring the bit about the question where he gives the definition of consciousness he wants to use - you are simply brushing aside where he says "Setting aside the question of whether it's possible to build a sentient machine...". You are entirely focusing on whether such a machine is possible, not what that implies. $\endgroup$ – Avik Mohan Sep 8 '16 at 17:14
  • 2
    $\begingroup$ The point is not whether such a machine could exist (hint: it is - I am one). The key point is that there is no physical test that will prove or disprove the statement "machines need to experience qualia perform to task X", because we cannot determine the experience of qualia in other entities. $\endgroup$ – NietzscheanAI Sep 8 '16 at 17:20
  • 1
    $\begingroup$ If the OP had used a different definition than 'actually experiencing', e.g. asking the question "What additional tasks might metacognition make possible?", then the answer would be different. As it stands, 'actual experience' = qualia, a well-discussed topic in AI philosophy. $\endgroup$ – NietzscheanAI Sep 8 '16 at 17:22
4
$\begingroup$

No.

The experience of seeing is by definition non-causal. Anything non-causal cannot be a requirement of a physical process; a qualia cannot afford a robot the ability to do something it otherwise could not.

Maybe.

Although a qualia is not required for a given AI task, that is not to say that any sufficiently advanced AI does not entail qualia. It could be that so-called AI-complete tasks require a robot that, although not making use of qualia, produces it anyway.

Yes.

Qualia may refer to some wishy-washy non-physical property, but it's special in that we know it exists physically, too. The fact I am able to discuss my qualia knowingly (or, if you don't believe me, the fact you are able to) implies that my (or your) qualia does have a physical effect.

It stands to reason that if we accept others' qualia on the basis of our own, it must be because of the physical basis of our own1. Thus one could argue that2 any robot that has an equivalent physical capacity must entail qualia.

1 since the subjective is physically non-causal, so cannot cause us to accept anything.

2 as long as you don't make the particularly odd assumption that qualia is somehow tied to its direct physical manifestation, which at best is tenuous since had we evolved the wrong one you would still claim it to be the right one with equal certainty.

$\endgroup$
2
$\begingroup$

Let's use a simple test based on common sense: how often do you see a human being solve problems requiring the use of reason when they're unconscious? Yes, you can find instances of geniuses like Ramanujan solving complex problem during or after a dream state, but those involve partial consciousness. You don't see guys like Einstein coming up with the theory of relativity while in a coma; the Founding Fathers didn't write the Declaration of Independence while sleep-walking; in fact, you can't even find instances of housewives putting together their shopping list for the week during deep delta-wave sleep. This is predicated on a hard definition of intelligence, requiring the use of reason; no one says, "That fly is intelligent" or "that squirrel is intelligent" precisely because neither is capable of using reason. This is a very high bar for A.I., but it is the common sense definition used by ordinary people as a matter of practicality, in everyday speech. Likewise, in practice, everyone assumes consciousness is necessary to the exercise of that kind of intelligence.

Conversely, we can come up with another common-sense based criterion for judging objections to this argument, particularly the solipsist one, based on 3 elements: 1) practicality; 2) the effect the objections have on those who hold them sincerely; and 3) the effect that actions based on those beliefs have on others. It's going to take me several paragraphs to make this case, but the length is necessary if I want to make the case in a complete, thorough fashion. It is true that we cannot prove that another human being possesses consciousness, if our standard is absolute proof. We cannot, in fact, provide absolute proof for anything; there's always room for some objection, no matter how ridiculous or trifling. As some philosophers have pointed out, perhaps all of reality as we know it is just a dream, or the product of some long, involved conspiracy like the plot of the Jim Carrey movie The Truman Show. The key to meeting such objections is that they require an infinite regress of increasingly untenable objections, whose likelihood plunges with each additional step required to justify such unreasonable doubts; I've always wondered if we could come up with a "Ridiculousness Metric" for Machine Learning based on the cardinality of such objections (or the pickiness of fuzzy sets). If we were to allow critics to stick their foot in the door with all manner of unreasonable objections, it would be impossible to close any debate. The human race would be paralyzed in inaction because nothing would be decidable; but as the rock band Rush once pointed out, "If you choose not to decide, you still have made a choice." At some point we must apply a test to decide such things, even in the absence of absolute proof; refusal to apply a test also constitutes a choice. Settling an argument of this kind is like a game of the Chinese game Go - once the other player's surrounded and has no more moves left to make, the game is over; if a person's evidence has debunked and they have no further justifications left, then we can conclude that they're acting unreasonably. There are people running around claiming the Holocaust never happened, or the Flat Earth Society, etc., but their existence shouldn't and doesn't stop us from taking action contrary to their ideas. We can debunk the objections of cranks like the Flat Earth Society beyond a reasonable doubt because in the end, they simply can't answer all of our rebuttals. I’m glad that qualia and Philosophical Zombies were brought up because they make for interesting conversation and food for thought, but solipsism is acted upon as rarely as the ideas of the Flat Earth Society precisely because the incomplete evidence we do have runs against it.

As G.K. Chesterton (a.k.a. "The Apostle of Common Sense") points out in his classic Orthodoxy, radical doubt of the kind many classical philosophers preached is not a path to wisdom but to madness; once we go beyond a reasonable doubt, we end up acting unreasonably. He says that in the absence of absolute proof we can fall back on another secondary form of evidence: whether a person's philosophy leads a man to Hanwell, the infamous British mental institution. Chesterton makes a good case that when people actually act on ideas like solipsism (rather than merely debating them in a pedantic manner in an ivy-covered classroom) they go mad The Philosophical Zombie argument is close to solipsism, which is actually one of the diagnostic criteria for certain forms of schizophrenia. The dehumanization that occurs when radical doubt is applied to qualia is also intimately tied in with sociopathic behavior, Although GKC does not cite his scary example directly, Rene Descartes was himself living proof. He was a brilliant mathematician who is still cheered for doubting all except his own existence, with the famous maxim "I think, therefore I am." But Descartes also used to carry a mannequin of his dead sister with him to European cafes, where he could be seen chatting with it. The gist of all this is that we can judge the worth of an idea by how it affects the well-being of the believer, or by how they in turn affect others through ethical choices based on those beliefs. When people actually act on radical doubt of the kind expressed in solipsism and denial of common qualia, it often has a bad effect on them and others they come in contact with.

In a roundabout way, the A.I. community also faces a quite serious risk - perhaps a permanent temptation - towards making the opposite mistake, of ascribing common qualia, consciousness and the like to its Machine Learning products without adequate proof. I recently heard a case made on shockingly bad logical grounds by well-respected academics to the effect that plants possess "intelligence," based on really weak definitions and clear confusion with self-organization. We cannot provide absolute proof that a rock doesn't have intelligence, which amounts to the old problem of disproving a negative. Thankfully, few men actually act on such beliefs at present, because when they do, they end up losing their minds. If we take such arguments seriously, we might see laws passed to protect the kind of Pet Rocks that were popular in the '70s (I'm still upset that mine was stolen LOL). It would be a lot easier, however, to make the same mistake of ascribing consciousness, intelligence and other such qualities to a state-of-the-art machine, because of wishful thinking, hubris, the lofty credentials of the inventors, the influence of science fiction and the modern love affair with technology. In the future, I have little doubt that we'll have Cargo Cult of A.I. - perhaps legally protected like some kind of endangered species, with civil rights, but having no more consciousness, soul or actual intelligence than a rock. Don't quote me on this, but I believe Rod Serling once wrote a story to this effect.

The best way to avoid this fate is to stick to the common sense interpretations and definitions of these things, which we keep backing away from in large part because they set a very high bar for A.I. that we may never be able to surpass in our lifetimes, if ever. Perhaps A.I. isn't even logically possible, at any level of technology; I recall a few proofs that can be interpreted to that effect. Those high but reasonable standards may be increasingly difficult to stick to if Chesterton and colleagues like Hilaire Belloc and Arnold Lunn were correct in their assessment that the use of reason has actually been breaking down in Western civilization, at least as far back as the Enlightenment; Lunn's 1931 book The Flight from Reason is a classic in this regard and has yet to be rebutted. This historical trend is a broad topic in and of itself - but suffice it to say that the denial of reason and obsession with technology are both directly relevant in obvious ways to the field of A.I. If the Flight from Reason is still under way, then we will be increasingly tempted to resort to feckless, facile objections in order to demote the use of reason and indispensable qualities like consciousness in our definitions of A.I., but come up with increasingly weak criteria for proving it; simultaneously, our technology will continue to improve, thereby boosting the "Artificial" side of Artificial Intelligence.

Don't get me wrong: if I didn't think we can do some really exciting things with A.I., I wouldn't be here. But most of them can be achieved without ever replicating actual human intelligence, by solving whole classes of tangential problems that are difficult for humans to think about, but which do not require consciousness or the use of reason that marks human intelligence. The image recognition capabilities of convolutional neural nets are one example, for instance; if we want human intelligence, we can always manufacture it through the easiest, most economical and time-tested way, by having babies. Perhaps these tangential forms of A.I. should be enough for us for now. We cannot inject the use of reason into our machines if we do not possess enough of it ourselves to decide whether reason is necessary for A.I., or even to discern what it consists of. We can't engineer or deprecate consciousness for A.I. till we're conscious of its significance. I'd wager, however, that everyone reading this thread and weighing intelligent responses is doing so in a conscious state. That in and of itself ought to answer our question satisfactorily for now.

$\endgroup$
2
$\begingroup$

As far as the definition you've provided:

actually experiencing sensory input as opposed to just putting a bunch of data through an inanimate machine.

Both computers and humans experience sensory input. You could hook a computer up to a human eyeball and have it run the same filtering routines that the human brain does (the removal of bluriness while you move your eye around, and from objects not in focus, etc).

I would put forth that a more accurate definition of consciousness is the ability and the tendancy to self-reflect. Both computers and human brains have autonomous activities. Not only mechanical but also in our reactions. The distinction between the unconscious computer and the self-aware human mind is that we also have the ability to 'look' at those patterns in ourself and consider them.

And so, no, consciousness is not necessary for any AI task. Image recognition is an AI task that does not require consciousness, either in humans or otherwise. Your brain sorts the 'wash' of colors from your eyes into discrete objects in a largely autonomous fashion.

tl;dr consciousness is self-reference.

$\endgroup$
-1
$\begingroup$

Two kinds of tasks require consciousness:

  1. consciousness

  2. Any task that requires extreme dynamicity, where solving problems requires analogizing between various 3D states of affairs and prior knowledge of how to solve the problem is minimal

However, once knowledge of how to solve a given problem is gained, further optimization will eliminate that need for consciousness.

If you give enough specificity to a problem, you remove the need for a general solver. And then the only remaining need for a consciousness is for the sake of itself.

$\endgroup$
-1
$\begingroup$

In a very niche sense, I'd say yes.

The only tasks that sentience would make possible was the actual feeling and thinking in and of itself. At this point, sentience doesn't play a part in any of the tasks we ask AI's to complete; we are rapidly approaching the point of being able to teach a 'dead' machine to do most anything a sentient AI can, in a practical sense.

Sentience colloquially often translates to 'the ability to reason while understanding that oneself and each other entity is a distinct acting agent'or something along those lines. It literally means something more along the lines of self-awareness and the definition of consciousness you have above. The point I'm making is that we are readily approaching the point where 'dead' AI's can very nicely mimic the first way of thinking, just by really nicely learning and interpreting data.

enter image description here

                              Does the robot see an amalgamation of bone, or a being that once was?

Thus, a truly sentient machine would be superior in capability (compared to a really, really advanced 'dead' AI) only in the respect of being able to 'truly' experience the information.

This runs very well in parallel with the so-called "Knowledge Argument" which in essence debates this very issue. The version of it that I heard which sticks with me is that there is a very smart girl in a room with access to all sorts of information. She likes the color blue. Or so she thinks; she's never actually seen it. She has all the information in the world available about colors and how they work, etc. but does she really know what blue is until she sees it?

Another great, historic venture into this field is the famous painting:

                                        Ceci n'est pas une pipe

The caption translates: "This is not a pipe". And the idea is that this, honestly, isn't a pipe. Right now it's a bunch of pixels on your screen in a certain configuration - we can all 'see' a pipe, but what does that really mean?

At the end of the day, I think super-intelligent 'dead' AI can practically do anything a 'live' one can, with the latter being superior in and of the 'liveness' itself.

$\endgroup$
  • $\begingroup$ The colloquial definition of sentience you give is not what the OP was using - their definition was actual experience, I.e. qualia. In contrast, the 'Black and White room' thought experiment you cite is specifically about qualia. $\endgroup$ – NietzscheanAI Sep 8 '16 at 18:00
  • $\begingroup$ Right. The rest of the paragraph addresses that. $\endgroup$ – Avik Mohan Sep 8 '16 at 19:18
-2
$\begingroup$

A being without sentience cannot suffer. If, for example, we wanted to take joy in the suffering of another, only an AI that was sentient would suffice.

Suppose we had some sadists who could not be satisfied or productive unless they got to produce lots of suffering. And say we only cared about minimizing human and animal suffering. What we would need for this job is something non-human and non-animal that could suffer. A conscious AI would do, a non-conscious one would not.

The claim was made in the comments that consciousness cannot be proven, other than perhaps by introspection. But clearly this is not a problem since sadists take joy in torturing others, and those others cannot prove they're conscious either.

$\endgroup$
  • 3
    $\begingroup$ A sadist is just as incapable as everyone else of distinguishing sentience from observably equivalent non-sentience, so sentience isn't actually necessary here either. $\endgroup$ – NietzscheanAI Sep 8 '16 at 20:50
  • $\begingroup$ A being with sentience cannot suffer - do you mean without? $\endgroup$ – Mithrandir Sep 8 '16 at 20:51
  • $\begingroup$ See the story 'The Soul of the Mark III Beast' by Stanislaw Lem I think, for a nice perspective on e-Sadism. $\endgroup$ – NietzscheanAI Sep 8 '16 at 20:55
  • $\begingroup$ @NietzscheanAI Suppose we had two AIs. One could prove it was sentient and the other could not. The former would be able to accomplish this task, the latter would not. Your claim that a sadist is just as incapable of everyone else also means that a sadist is just as capable of everyone else. $\endgroup$ – David Schwartz Sep 8 '16 at 20:57
  • $\begingroup$ There is no known means of proving sentience. See the Flanaghan paper on 'Zombie Earth' that I link to in my answer. $\endgroup$ – NietzscheanAI Sep 8 '16 at 20:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.