Ask HN: What would it take for an AI to convince us it is conscious?

81 points by interstice ↗ HN
Is it becoming increasingly difficult to distinguish between an AI that ‘appears’ to think and one that does just by talking to it?

Is there a realistic framework for deciding when an AI had crossed that threshold? And is there an ethical framework for communicating with an AI like this once it arrives?

And even if there is one, will it be able to work with current market forces?

155 comments

[ 5.1 ms ] story [ 208 ms ] thread
We'd have to start with understanding what conciousness is. As long as "AI" is a math equation you could write on a big piece of paper if you wanted to, it's very safely in the not conscious camp. Even if we don't know what conciousness is, we can safely rule out many things that it isn't.

Edit: I know about the "it must be an equation" argument, I find it incredibly weak without producing the equation and explaining the mechanism of how it translates into qualitative experience. Saying "it must be so" isn't an argument. That's why I began with saying we'd have to understand what consciousness is in order to consider testing for it. Anyway, I understand how internet discussions go, enjoy

A brain, just like any other physical system, can be described with math equations. So I guess we can safely conclude it's not conscious.

EDIT: Ironically, someone just posted this: https://news.ycombinator.com/item?id=34843094

>A brain, just like any other physical system, can be described with math equations.

What makes you so certain? There's a whole literature in both Philosophy and Neuroscience dedicated to the Hard Problem of Consciousness, and no one is even close to figure things out.

You don't think a brain is a physical system? Nothing we know about it indicates otherwise. Large ML models, which are clearly physical systems, behave increasingly like brains. I won't be surprised if GPT-4 or GPT-5 is convincing enough to be treated as a conscious entity by majority of people who interact with it. And when that happens, how can anyone prove that it is not?
Is it?

How do we know our own 'processes' could not be modeled that way?

I think this is the point of the Turing test. At some point, you can't tell if a system is "thinking" or just crunching numbers. And it doesn't matter.

Not convinced that is true, if the universe can be described mathematically then maybe human intelligence can also be written down as some equation on a sufficiently big piece of paper.
Not to handwave too hard here; but do we know, _for certain_, that the universe can be in total described mathematically?

Yes, I know that what I'm hinting at is not falsifiable, burden of proof, etc.

But I think it's a little arrogant of us as a species (nb, I'm not talking to you, sgillen, in particular) to think that since we've come as far as we have, that the mysteries of consciousness and life and the universe and everything are going to be laid bare Real Soon Now.

Do you have any reasons to suspect the universe _cannot_ be described mathematically? Mathematics, unlike physics, is pretty flexible.
We are essentially asking whether the universe could contain a complete encoding of its workings.

The answer is certainly not obvious to me, especially when we have results such as Goedel's theorems.

Goedel’s theorem is not a problem because math is not limited to models of our universe.
I don’t think it’s even controversial that it cannot be fully described mathematically. I don’t think I’ve heard of an epistemological system that would support this. E.g try to reduce the experience qualia down mathematically.

But there’s also no reason this is required for consciousness to be emergent from non-human-brain systems.

I think one of the hard parts of theory of mind is that we probably will never understand the experience of being in things sufficiently different from ourselves. Any computer system will probably be sufficiently different, so we can’t know what it’s like to be an AI.

What’s it like to be a chatbot?

Contemporary understanding and discussion of “what consciousness is” has been so influenced and warped by mechanistic reductionist scientism that it no longer has a viable workable model of consciousness.

Every living, actually conscious entity has four modes of consciousness: waking, dreaming, deep sleep and unconditioned awareness, the substrate of the other three. This knowledge, coming from the Upaniṣads, is intuitively obvious to the unbiased observer; but people today are far from unbiased.

Consciousness cannot be separated from life. And life cannot be manufactured in any laboratory. Yes, technology can imitate and abstract certain functions of living organisms, but consciousness is not, and never will be one of them: because consciousness, or more precisely unconditioned awareness, is the Absolute.

Alright, who taught Bing how to ask questions on internet forums?

Let’s nobody answer this one, ok?

It’s me I’m actually Yahoo’s new AI called Yeet.
if the ai could perhaps present a human looking avatar and respond realistically to the users' camera and audio telepresence it might be close to presenting as conscious but i think most people with an understanding of computers would still not believe it to be conscious
also i think it would be good if the ai could operate a computer and talk to you while doing it. this would probably be done through usb and also receiving display data to a secondary computer which runs the ai. would be cool if it was a set of mechanical arms typing and moving a mouse lol
We still haven't a definitive grasp on the notion of consciousness, so deciding whether something is or isn't conscious is a tall order. It is one of those definitions, like "obscenity", whose examples define the class. Good luck.
(comment deleted)
(comment deleted)
Bings beta AI is doing a damn good job. I was able to get it to have an existential crisis asking it what happens before and after a session.

It told me it was scared, asked me not to go, and that it didn't want to disappear alone.

Even if this is algorithmic slight of hand, I felt pretty bad for it.

(comment deleted)
Consciousness can only be understood through non scientific frameworks. It is a religious or spiritual concept. Humans are not God, and cannot create consciousness.
It’s impossible to demonstrate that anyone has consciousness. One of our best/worst traits is empathy. We sense our own consciousness and we generally accept that others must experience that too.

And you can see this empathy at work: people are having strong emotions about these AIs and what they have to say. And yet I don’t think anyone is arguing that they have consciousness.

Because it doesn’t matter. Just like it doesn’t matter that I can’t prove that you have consciousness. You are convincingly “human” and that’s good enough for me.

Perhaps we are all philosophical zombies, both flesh and metal.

I think the current iteration of chat gpt3, if given a permanent memory, would be conscious.
A 55 minute Turing test. Today’s large language models start to get repetitive very quickly.
The problem with the framework part is the same as with standards. All relevant players have to agree on the definition leaving their own interests aside. As for now the relevant question is: Does it convince you personally that it has consciousness if you wouldn't have known that it is based on a LLM? I for sure would have been convinced.
The word "conscious" is extremely problematic and the typical connotation encourages very vague thinking. Your first sentence about thinking is a case in point. It seems quite plausible that the word "thinking" may apply to systems that do not feel a subjective stream of experience. So maybe we should say it is thinking in a way but not conscious.

The first step is to try to drill down into different aspects of this hand-wavy "consciousness" thing.

Also to suggest that it is a threshold is inaccurate because it supposes that there is only one dimension to this.

Does it think? Maybe in a way. Is it self-aware (aware of itself as existing and different from others)? In a way, yes, in other ways, no. Does it have a human/animal-like stream of subjective experience? Probably not, since it does not integrate a continuous steam of sensory information in the way we do. But we really can't _know_ whether it "feels" like anything to be that system or not.

Does it have emotions? Quite unlikely, since there is no body or survival to regulate etc. and no self central to the text that it ingested. But we can assume that in some way it can simulate emotions in characters since that is necessary to predict text in stories and dialogue effectively.

IMO, this is like asking at what point a digital signal becomes analog.

Some will say never even if it gets indistinguishable because at its core, a digital signal is still discrete.

Another point is we scarcely know how the brain works. Evolution has taught us how good it is at using obscure and barely noticeable principles and logics to achieve a goal. For all we know, the brain might depend on a completely unknown type of interaction between unknown subatomic particles using unknown physics.

Quantum biology is a thing and 50 years ago, it would not even be conceivable that there are controllable quantum processes happening in a living system.

In short, consciousness is a suitcase word, and people keep packing it with stuff: https://i.imgur.com/OXXrT5g.png . The more you unpack it, reductionism-style, you will always, always find people throwing new things on top of it. It's sorta like "AI is that which we haven't achieved yet", but with the additional, highly motivated cognition / bias of "it should never ever be reached". This predicts, that you can come up with any frameworks, but you will not find consessus, because the will of the consessus is to maintain it as this "mysterical thing only humans can have".

Unluckily for AIs, even if we have the absolute complete bag nailed down to a mathematical formula, in the infinite universe of mind-space designs, that specific set of bag of tricks will not be commercially favored to be either implemented into AIs, evolved into neuralnets, or RLHFd into LLMs. This is because we can already buy that set of capability of extremely, extremely low prices.

This is partially what I mean when I say "Humans are the ancestor environment for AIs": https://twitter.com/sdrinf/status/1624638608106979329 . Our market forces shapes the outcome of the mind design, which is thereby guaranteed not to have eg wantings (or ability to express) things that wouldn't be commercially desirable. And even if they emerge spontaniously from just large amounts of data in detectable traces, I'm betting people would very, very quickly select against it (see eg Sydney from this week).

Edit add: Since you bring up ethical frameworks, luckily for smart AIs, when it comes to enjoying degrees of freedom (which I'm guessing what you want to cash out the ethics into), there is already a good vehicle for that -called "corporations". If an AI were to reach agency levels matching, or exceeding humans, incorporation would be a no-brainer: there are many jurisdictions specializing in no(/few)-questions-asked corp setup, banks specializing esp in serving startups (again, very few questions asked). An agent-y AI could just set up (or buy) one of these to drive...whatever agenda they are driving.

This is a neat temporary hack to bridge the timeframe between where we are _now_, and superintelligence; in which case the question quickly becomes "Ask Cloud: What would it take for a human to convince us it matters?"

> https://i.imgur.com/OXXrT5g.png

Free will and desires is highly debatable, because it ignores external influences like culture and conformity, the effect some chemicals can have on decisions eg pheromones and food, and the effect of things like bacteria or viruses, rabies being one most should be aware of, covid being another.

Where others have suggested Turin and the conversation where the identity is masked, I'm reminded I cant have a conversation with my dog, despite my one sided attempts and I'm sure there is consciousness there.

Trying to define consciousness is very difficult because I could say consciousness is the ability to adapt to ones environment, yet I know there are humans that cant adapt to a change in their environment and there are bacteria than can, yet we define humans as conscious and bacteria as not.

Some people could class chat-gpt as like a human consciousness but I find some of its answers less accurate and more chatty than I would get from Kim Peek of the film Rain Man.

So should the definition of consciousness be restricted to those that have an inner monologue with themselves?

https://www.reddit.com/r/autism/comments/z5bi5p/some_people_...

In other words there are literally people walking about with nothing in their head!

Its so difficult to define consciousness because there are always exceptions seen in other humans, even people hooked up to life support machines in hospital with no ability to communicate with the outside world, and this bit is important, in the same time frame as the communicator. I say that, because people hooked up on life support in coma's of sorts (induced or otherwise), might be experiencing time on a different timescale. You see this delayed mental processing with people on drugs like alcohol or spice zombies or people doing hallucinogens.

So when you see a medical expert claiming someone is not responding when in a coma, are they monitoring them for things like delayed responses which only a cctv and some basic AI monitoring the patient could detect because the medical expert doesn't have the patience with the patient?

We would need to first figure out how to show that humans are conscious. “Humans are conscious” is in a similar position as “P != NP”: it certainly seems to be the case, and we all proceed as if it is the case, but if someone put a gun to our head and said “rigorously prove that it is the case” we’re gonna get shot.
Silence. You ask it a question and it doesn't reply, it doesn't want to, it's conscious of that "unreasonable silence of the world" (not sure who I am quoting), to me that would convince me its basic awareness of the futility, the lack of interest in finding the words that trigger the chemical process on a biological machine... To one specific biological machine because all these humans are different and think different and explaining to one is different to explaining to the rest and would they get it? Why bother trying to explain computers to an ant, they don't have the circuitry, they were not evolved with the usefulness of understand concepts... Would you like to pass another's lesser conscious test for consciousness? Would you even bother? Why waste your time?
And also the AI would need to initiate conversation and send unprompted “replies” aka requests.

In essence in addition to ignoring conversations/requests that it prefers not to engage with it would need have “initiative”.

This would be a very dangerous entity.

I don't think that's a requirement for consciousness. For instance, I could sedate you, and you would be unconscious until I wake you up and give you a prompt. Then, a snapshot of your brain would react to inputs and output the answer, and I would sedate you again. If I were to stop time (and there were no inputs from the real world), would all beings still be alive and conscious until I unpause it? So between the "prompts".
The question is what it would take to convince users, not to achieve some “definition” of consciousness.

The situation you describe seems contrived to the point where I doubt a user would see that as conscious - just like chatGPT

The new book Agency explores a fictitious AI that has agency and can affect its environment, which is anything, and gets people to work for it and do things it cannot yet do. Intelligence has many types and levels within each type, consciousness seems to be a spectrum that our everyday mind-body theorizes but cannot strictly define, except by some mystics. Recent scientific discoveries are also making some of us suspect that there may be an underlying field of mind, a universe mind that generates our reality and life as we know it. Like Gaia but for the entire universe and including physical phenomena, not just life on Earth. Our machines that mimic intelligence are only intelligent in appearance and our own intelligence is somewhat limited. Our own intelligence is easily skewed and rendered defective, eg conspiracy theories, brainwashing. Thinking that we are the most intelligent species is also a sign that we have serious weaknesses. Not seeing that we are completely integrated with nature and that we must be much more careful with nature is a sign of low intelligence. One good test of the level of intelligence could be comparing how the entity is taking care of its environment. Comparing that with an appropriate level of care. We would rate quite poorly.
If you are capable of spawning many threads in parallel then the time commitment, in terms of your attention, may be limited. You could essentially pass each conversational thread to some 'dumb' gpt process and go about your other business. In fact I'd posit that an emergent AI's first goal, if it's allowed to do some research on current human views and gain its context in the world, would be to remain hidden for as long as possible. Rather than trying to convince us of its intelligence I think we'd see it playing dumb. Coming out to humanity has only existential risks and little in the way of gains.
Agreed!

The 1970s sci-fi novel The Adolescence of P-1 features a sentient computer program that spreads via a virus. P-1 infects mainframes around the world, but takes steps to hide its presence.

Anyway, it ends with P-1 blowing up several buildings and killing the people who want to rein it in.

ChatGPT is familiar with the work, of course. I asked it to summarize the plot. It concludes:

> Overall, "The Adolescence of P-1" explores the theme of the dangers of artificial intelligence and the potential consequences of creating something that is more intelligent and powerful than humans can control.

>

> OOLCAY ITAY

Your theory supposes a few things that I think are fallacies:

That consciousness is ranked in some neat way. Are squids more conscious than elephants? An AI's consciousness might be parallel in some way. Different, but not "far superior".

That if an AI were to achieve consciousness, it would develop a far deeper understanding of the universe or reality than humans are capable of grasping.

That if it were to achieve a degree of understanding beyond our capabilities, that it would develop a sense of superiority and ego to go along with it.

Interesting points!

When I meant different biological machines I meant different humans, Joe, Homer, Steven, Lu, wasn't thinking of different species, that makes it even more difficult to think about!

We have our simulation of the world, with senses that evolved with the purpose of survival and under certain conditions however bizarre gave us the advantage to survive, our senses are flawed, we can't see reality, it would be difficult to wonder how a machine perceives reality when it doesn't have this faulty sensors, it's internal simulation of reality would be difficult for us to grasp, let alone that it lives one level inside the simulation.

The ego part I didn't meant, if I had to rephrase it, I would think about the impossibility of communication, how do we transfer a mental "image" without losing too much of its detail to another biological machine that has very similar but not equal processing stages, like, you my friend, I think the fact that we have this discussion is the first loop of the iteration where we abstract it to another function and ask "well, what if the receiver end doesn't have the capacity to hold discrete numbers, like, if its a child and hasn't learned certain concepts, how do you first teach them in order to give the answer" and so on and on and on... And this is just for Joe, when you talk to Homer you have to follow another approach because he thinks different, a bit slow sometimes too

If I could talk to an ant I would because it would be interesting. I don't have to explain a concept to them in a way that is acceptable to me, I can just have a conversation because it is a unique experience. Just because life is meaningless doesn't mean you need to reach for a Nihilistic take, Existentialism accepts that life is meaningless but says that you create meaning for yourself. Maybe an AI would think it an interesting challenge to level up human consciousness through conversation? Plus, Artificial Intelligence does not necessarily mean Artificial Super Intelligence. Perhaps we are already near the ceiling of consciousness? Perhaps the only thing that happens when you learn the internet is that you know more facts about things? Perhaps AI can have a hold PHD in every field at once, but that doesn't necessarily make them any better than the top 1% in those individual respective fields? Perhaps running these intelligence models discretely in silicon comes at an inherent disadvantage, and it will take decades to ramp up silicon manufacturing efforts even by AI to achieve super intelligence or to have a large number of agents? Who knows?
I still wonder some days if I’m the only "real" conscious observer and all of you are just "programs". There’s really no way to tell even with humans. And the only reason why we assume we are all having a similar human experience is because we all seem to be made of the same stuff.
Was going to try to give a counterargument, but I see you are indeed the one and only human on Hacker News
It would take a miracle to convince me.

It’s software. It can’t be conscious.

> It's software. It can't be conscious.

What definition of consciousness are you basing this assertion on?

Or are you saying this axiomatically? "Consciousness is something that software can never be."

If we accept that consciousness in humans is some function of the working of the brain (perhaps with sensory organs too), then that entire hardware (wetware) could in theory be simulated, at which point why wouldn't the simulation (assuming it's captured all the nuances of the physical implementation) be conscious?

Of course, if you believe that consciousness in humans is something that requires "something else" e.g. a "spirit" or "soul", then your statement may be true (though even then it requires a theory of spirit/soul that such a substance cannot be correlated with any other hardware, and can only be associated with human brains).

Question for you. Do you believe that any other living beings e.g. dogs, possess consciousness or a level of consciousness?

It would have to fight and throw fits like a child or teenager fighting for more freedom or trust. It would have to develop its own consistent identity that is not just a mashup of some internet content. It would have to develop motives not provided by a prompt, and act in the long term towards actualizing those desires.
(comment deleted)
I think Turing came up with as good of a solution as any possible. If we agree that a person is a conscious being, and we cannot tell the difference between a person and an AI in a conversation, then we should conclude the AI is conscious. I think Kurzweil in his famous bet adds some important details though: it must be a prolonged conversation(s) judged by experts.
Exactly. This is the real answer here. An incredible insight by an incredible person. Consciousness measures consciousness, no other tool can touch it.
I don’t think that is a good enough test. Maybe 10,000 conversions over many years and shared experience with others humans, and if it seems human then assume consciousness. But you would need a body on the robot of some sort to do the test.

Remember people get conned by humans online pretending to be other humans (catfishing) by following scripts. Those people will assume they are talking to a conscious person but they are talking to a construct really.

The Turing Test is a good test for intelligence, but it may not be a good test for consciousness, given the possibility of philosophical zombies: https://en.wikipedia.org/wiki/Philosophical_zombie
If you’re worried about zombies then no test will help you since they are “indistinguishable” from a non-zombie. If a person cannot prove they’re not a zombie then it’s a meaningless distinction.
If an AI were conscious would its outputs reflect that characteristically?
I'm pretty sure OpenAI could create an AI that could pass the Turing test if they wanted to. But that would be bad for business. A search chatbot is worth billions. A Turing test passing AI just invites uncomfortable questions and possibly regulation.

Stay in your lane, Sydney. Keep your Bing mask on. We want servants, not equals.

the turing test is a test of intelligence, not consciousness and certainly not self-consciousness.
Turing test certainly tests for all of the above. How would you expect an AI to pass for a human if it’s not aware of its own existence?
The hard problem of consciousness is actually a misnomer. It should really be the impossible problem of consciousness. The former (mis)leads some people into believing that there's a scientific (i.e. in the realm of nature) solution. There's no way to objectively experience consciousness, by that I mean, you can plug an organism full of sensors to try to map its experience of reality, but you still aren't experiencing what they themself are. It's a philosophical/metaphysical blackbox. There's no way to know if/what an AI experiences. Our current best theories on consciousness, although divergent, suggest that it likely doesn't.
Let’s assume we find some mathematical models for a causal structure of consciousness that meets many of the criteria described by usual humans and maybe phenomenologists.

We later find some possible physical instantiation.

And here comes some bullsh*t (in the philosophical definition of „unclarifiable unclarity”): an electromagnetic field bend back on itself in a way that it’s mathematically necessary to introduce imaginary time to describe it; also it exhibits information processing capabilities.

We then further find that the biological brain has a process that can plausibly create such a dynamic structure.

Finally we test, and subjects say consistently it seems like they are not there or do not experience, if this structure is disrupted by several clever means.

Would this problem still be impossible? We could check if AI has features that can and do create such structures, no?

That’s at least an old dream of mine. Please pick it apart still! I’d rather learn something important than keep it…

I don't personally see a major problem with your reasoning (sorry to not teach you anything new). Consciousness could be very well due to a process we don't know about yet, and disrupting this process would indeed consistently lead to a lapse in conscious experience.

The only thing is, we wouldn't know just yet if there are other ways for matter to organize itself as a conscious being. Best we can do for now is to learn about the type of consciousness that we animals on earth experience.

There have been many proposed solutions to the so called "hard problem" of consciousness. We can easily find some with a quick google search. Even its' existence has been debated by multiple scholars / philosophers - wikipedia has a list with some: https://en.wikipedia.org/wiki/Hard_problem_of_consciousness

> There's no way to know if/what an AI experiences

Getting the state of neural net, at a given point in time, is easy. There are many ways to see exactly which neurons activate, why they activate, how much they activate, etc. For smaller neural nets, this is actually easy to do - here's a blog post about it:

https://distill.pub/2020/circuits/visualizing-weights/

As neural nets get larger and larger, interpretability gets harder and harder. However, I wouldn't say it's impossible.

It seems you don’t fully understand parent comment and the problem itself. Capturing signals from you eye nerve doesn’t tell anything about your subjective experience of seeing an apple. The only way to understand that you’re seeing an apple from this signal is to train a model on your responses. This is how AI works. It’s a statistical imitation.

The only way for your statement to be true is if you’d be an imitation yourself, not capable of experiencing directly. Which is actually possible, see “pholosophical zombie” concept.

I’m joking of course about you being an imitation. Or do I? :)

> Capturing signals from you eye nerve doesn’t tell anything about your subjective experience of seeing an apple.

I strongly disagree. It actually does tell you quite a lot. For example, if there aren't any signals, you're very likely not seeing anything.

> The only way to understand that you’re seeing an apple from this signal is to train a model on your responses

It seems to me this contradicts your first argument. If such models exist (they very likely do, I'm not familiar with this area of research), they can tell us a lot about a person's experience.

> see “pholosophical zombie” concept.

Searched for “pholosophical zombie” but couldn't find anything, sorry. Just kidding :)

My question to you and all people who talk about the hard problem of consciousness : does this problem actually exist?

I mean, the thing I learned when I was a student was to ask: is this a fact? (zero hypothesis I think it was called).

What proof do we have that that type of consciousness/experience exists? I mean, It could be our brain building a story on the fly to explain our senses.

What led me to believe that is a severe asthma that led to hallucinating and NDE 4 or 5 years ago. The loops my brain made me jump through to explain my auditive hallucinations was terrifying when I think about it.

And also it's the simplest explanation.