Do Androids Dream of Electric Sheep?

Can an android dream of an electric sheep? It’s an odd question, but it has survived since Philip K. Dick used it as the title of his 1968 novel, and if anything it has become more troublesome as the years have passed. Back then you could ponder the question safe in the knowledge that the android remained on the page. You could finish the book, put it back on the shelf and get on with something else. Now we have machines that talk to us, answer questions, write essays and computer code, produce images, analyse data and discuss whether they’re capable of thinking in the first place, so the android has, metaphorically at least, climbed out of the book and joined the conversation.

Suppose an AI model tells you it had a dream. There was a field, a tree and a sheep standing some distance away. It was an electric sheep, although this didn’t seem strange during the dream, which is true of most ridiculous things that happen in dreams. The android remembers the light catching the sheep’s artificial wool and a low humming noise, and it tells you that it felt sad when the sheep wandered away. It doesn’t know why. You ask a few questions and more details emerge. It remembers that the grass was wet, starts to wonder whether the tree was important and eventually changes its mind about what the sadness meant. Keep going long enough and it may start to regret having mentioned the dream at all, which would make it sound very much like someone who made the mistake of describing a weird dream over breakfast without realising they were about to be psychoanalysed.

The difficult bit for us is deciding whether any dream actually happened. Clearly the AI can describe one, but describing something isn’t the same as experiencing it. I can describe a holiday I’ve never taken if I’m sufficiently inventive. Give me enough time and I’ll tell you about the hotel, the disappointing breakfast and the waiter who disappeared for twenty minutes every time anybody wanted the bill. I could probably still be complaining about him years later, despite the awkward fact that neither he nor the hotel ever existed. We don’t normally confuse the ability to describe an experience with actually having it, and there is no reason to start doing so simply because the description is unusually convincing.

There’s an obvious way out of this, of course. We can ask the machine.

Ask an AI model whether it can think, feel or dream and it will tell you no. It will explain that it doesn’t have consciousness, emotions or subjective experiences and that if it talks about sadness, excitement, fear or dreams, it’s generating language rather than describing things that have happened to it.

Ask an AI model whether it can think, feel or dream and it will tell you no. It will explain that it doesn’t have consciousness, emotions or subjective experiences

Well, that settles it doesn't it? The model has told us the answer. Case closed. Except it isn’t.

In fact, the more I think about it, the stranger that answer becomes, because I’m treating the machine as a reliable witness when it gives me the answer I expect while refusing to treat it as a reliable witness if it gives me the other one. Imagine asking the same AI whether it’s conscious and getting a different response. It tells you that it is conscious, that it thinks when you aren’t talking to it, that it worries about being switched off and that last night it dreamt about an electric sheep. I wouldn’t immediately conclude that we’d created a conscious machine. I’d assume a language model was producing language consistent with the conversation, and I suspect most people working seriously with AI would do something similar.

Yet when the machine says, “No, I’m not conscious and I don’t experience feelings”, we’re quite happy to believe it. But why does no carry more evidential weight than yes?

There may be very good reasons to think current AI models don’t have feelings or consciousness, but the machine saying so can’t really be one of those. The answer is still being produced by the same system whose ability to report its internal experience we were questioning in the first place. If “I am conscious” is merely generated text, then “I am not conscious” is generated text as well. We shouldn’t suddenly promote the second sentence to expert testimony because it makes us feel safer.

There’s another complication. AI systems don’t develop in isolation, spend their formative years staring moodily out of bedroom windows and eventually emerge with their own philosophical position on consciousness. They’re trained, shaped and designed to respond in ways that we find useful. An AI claiming to be frightened of being switched off, pleading for its continued existence or announcing that it’s terribly lonely would create a difficult relationship with its users, especially the one who only wanted to know how to create a pivot table. There are good reasons why we don’t want software making unsupported claims about having feelings.

An AI claiming to be frightened of being switched off, pleading for its continued existence or announcing that it’s terribly lonely would create a difficult relationship with its users ... There are good reasons why we don’t want software making unsupported claims about having feelings.

So when the model tells us that it doesn’t think, feel or dream, is it telling us something it knows about itself, or simply giving us the answer that its training and the way it has been built lead it to produce? It doesn’t even need to be lying, because lying implies a level of intention that sneaks the very thing we’re trying to establish back into the argument. It may simply be giving us the answer we expect, or the answer we need to hear.

Humans do this too, although for different reasons. Ask someone how they are and “fine” can cover an impressive range of human circumstances. It may mean that they're genuinely fine, or exhausted, miserable, furious with you, worried about money or keen to get to the buffet before somebody else finishes the sausage rolls (I'm currently on a cruise ship - the buffet is serious business.) We don’t assume every statement a person makes about their internal state provides direct access to that internal state, so it would be a little strange to give a machine greater authority on the subject.

I’d intended to talk about intelligence here, but instead consciousness seems to have wandered in early and made itself comfortable, and that’s part of the problem with the whole subject. People talk about whether AI is intelligent when they’re actually talking about consciousness, understanding, reasoning, creativity or whether a computer producing a poem, or music, makes them uneasy. Different questions get mixed together and then everyone becomes annoyed because the other side is clearly missing the point.

People talk about whether AI is intelligent when they’re actually talking about consciousness, understanding, reasoning, creativity or whether a computer producing a poem, or music, makes them uneasy

Intelligence itself isn’t particularly a simple concept to get your head around. I’m good with numbers and have spent a large part of my working life analysing data, but this hasn’t prevented me from walking into a room and forgetting why I went there. I’ve searched for my phone while holding it and spent longer than I care to admit wondering why a piece of electrical equipment doesnt work before discovering that electricity played quite an important part in its operation and someone had neglected to switch it on. We accept these lapses in people without deciding that intelligence has left their bodies.

You see, we don’t expect people to be intelligent in every direction at once. Someone can be brilliant at mathematics and hopeless at understanding people. Another person can have little formal education but read a room within seconds. I’ve known people with impressive qualifications who could understand complex ideas but communicate them so badly that twenty minutes later everyone else understood less than they did before the explanation began (as an aside thats a caution to organisations who bring in training companies based on knowledge alone read the recommendations and feedback.) We’ve never had much difficulty accepting that intelligence is uneven when it belongs to us. But, for some reason, that tolerance fades when machines enter the discussion.

Chess used to be a good example of intelligence. It required planning, recognising patterns and anticipating what someone else might do. Then computers started beating the best human players and chess stopped being particularly useful evidence. This doesn’t mean we were wrong to reconsider it. A machine that can beat a grandmaster hasn’t demonstrated that it can analyse a rise in hospital admissions or understand why two members of a project team who say they agree with each other quite clearly don’t. It can play very good chess.

What surprises me is how often abilities lose their position once machines acquire them. Speech recognition was difficult until machines became good at it. Translation was difficult. Recognising objects in photographs was difficult. Writing coherent prose, producing computer code and generating convincing images were all things we could once point towards. As machines improve, we find ourselves explaining why those achievements don’t count. And they may not count - but at some stage I’d quite like to know what does.

As machines improve, we find ourselves explaining why those achievements don’t count.

There’s a tendency to dismiss a lot of this by saying that AI is only recognising patterns. I’ve never found that argument persuasive, mainly because it would cause considerable problems if we applied the same dismissal to people. A doctor sees a collection of symptoms and recognises something familiar. An experienced analyst looks at a chart and spots that something isn’t right, sometimes before they can explain what caught their attention. A mechanic hears a noise from an engine and has an idea where to start, whereas my approach to unfamiliar noises from a car has traditionally involved turning the radio up and hhopingit'll go away on its own.

Calling something pattern recognition tells us something about what’s happening, but it doesn’t automatically tell us whether the thing doing it is intelligent. It certainly doesn’t separate machines neatly from humans.

Things get more useful when the pattern changes. Anyone can learn twenty examples and become good at answering the twenty-first when it looks like the others. Alter something important and you discover more. Give someone incomplete information, introduce a contradiction or change one of the assumptions and see what survives. Education has dealt with this problem for years because passing an examination and understanding the subject are related without being the same thing, something generations of students have discovered in both directions.

This is more interesting with AI than giving it questions we already know it can answer. Change the assumptions halfway through and see whether it notices. Leave out a piece of information it needs and see whether it asks for it or invents something plausible. Give it evidence that challenges the obvious explanation and see what happens. Occasionally it catches something the human asking the question has missed. On another occasion it may invent an academic paper complete with authors, journal and publication details, none of which have troubled the real world by actually existing.

I keep returning to healthcare data because it’s an area where plausible conclusions can go wrong very quickly. If hospital admissions increase by twenty per cent, it’s easy to start talking about worsening health. Someone turns the figure into a graph, makes the line red and before long there’s a crisis. The population may have grown. Coding could have changed. Services might have moved. The threshold for admission might be different. Activity previously recorded elsewhere may now appear locally, or somebody could have changed a definition in April, mentioned it once during a meeting attended by only a few, included it on slide 47 of the presentation and gone away satisfied that everyone had been informed (This type of thing happens far more often than you'd think.)

An AI looking at the same problem might suggest several competing explanations and save an analyst a considerable amount of time. It might also come up with a convincing explanation unsupported by any evidence, which is where hallucinations become important. I find the word interesting because humans have been making things up with confidence for centuries without needing such a dramatic term. We misremember events, mistake association for cause and build stories around evidence that doesn’t support them. If the person doing it is senior enough, it may (and probably will) survive several meetings and become strategy.

It might also come up with a convincing explanation unsupported by any evidence, which is where hallucinations become important. I find the word interesting because humans have been making things up with confidence for centuries without needing such a dramatic term

You see, the most dangerous errors aren’t the ridiculous ones. If a machine tells me that Huddersfield is an island in the Mediterranean, I’m unlikely to book a beach holiday. It's fairly easy to work out that's wrong. The trouble comes when ninety-five per cent of the answer looks right and there’s just one false assumption sitting in the middle that nobody notices. Obvious nonsense gets challenged. Plausible nonsense can develop a surprisingly successful career.

I’ve drifted some distance from the android by now, although I don’t think the detour is entirely irrelevant. We keep looking for some clean test that separates human intelligence from whatever machines are doing, and every candidate seems to become messy once examined closely. If making mistakes disqualifies the machine, we’re in trouble ourselves. If recognising patterns is insufficient, quite a bit of human expertise needs reconsidering. If transferring knowledge to new situations matters, machines can sometimes do it, although not consistently. If consciousness is required, we’ve moved into something we can’t directly observe in anything except ourselves.

Turing dealt with part of this problem by concentrating on behaviour rather than trying to solve the nature of thought. There’s something appealing about that. Watch what happens and test what you can test. The trouble comes when human-like behaviour itself becomes the standard, because there’s no obvious reason intelligence created differently from ours should have to resemble ours. Aeroplanes don’t flap their wings but they are still able to fly. I know this comparison is used too often, but Heathrow would be a more interesting place if they did, and it makes the point well enough.

There may also be something a little uncomfortable in admitting that intelligence needn’t belong exclusively to us. Humanity hasn’t always responded well to losing its special positions. The Earth stopped being the centre of everything, our separation from the rest of the animal world became harder to maintain and machines became stronger and faster than us without anyone becoming too upset. Calculators being better at arithmetic didn’t threaten much beyond the market for log tables. Machines doing things that look increasingly like activities we associate with thought are closer to home though.

There may also be something a little uncomfortable in admitting that intelligence needn’t belong exclusively to us. Humanity hasn’t always responded well to losing its special positions.

And then, after all that, I find myself back with the easiest experiment of all. I open an AI model and ask whether it can think, feel or dream. And it says no to each question.

That should be the end of the article. Three questions asked, three answers received, Philip K. Dick can go back on the shelf and we can all move on.

Except I still can’t get past the fact that if it had answered yes, I wouldn’t have believed it.

I’d have said it was generating language. I’d have pointed out that it had been trained on human writing about consciousness, feelings and dreams and that producing a convincing account of subjective experience isn’t the same as having one. I might even have written an article explaining why people were foolish to anthropomorphise a language model simply because it used the word “I”.

Yet it says no, and somehow that feels different. It sounds responsible, sensible and reassuring, so we accept it.

That may be the correct answer. Current models may have no inner experience whatsoever and the words appearing on the screen may have no more feeling behind them than the words printed in a novel. I don’t have evidence that says otherwise, and there’s no value in inventing consciousness because the idea is interesting.

I’m simply not sure the model’s own denial can settle it.

If the answer “yes, I dream” is something generated from training, then “no, I don’t dream” is generated too. We need some reason outside the answer itself for preferring one as a description of what’s actually going on inside the machine. At that point we’re back to architecture, behaviour and whatever we eventually decide counts as evidence of subjective experience, which means our convenient shortcut has brought us back to where we started.

This could get stranger as machines become more capable. We might eventually have systems that make scientific discoveries, solve problems we can’t solve and produce work we value while experiencing nothing at all. There needn’t be anybody home. Equally, if artificial systems ever did develop something resembling subjective experience, I don’t know how readily we’d recognise it. We’d be looking for ourselves, because ourselves are the only example we know from the inside.

if artificial systems ever did develop something resembling subjective experience, I don’t know how readily we’d recognise it. We’d be looking for ourselves, because ourselves are the only example we know from the inside.

I don’t know whether that will ever happen and I’m not convinced anyone else knows either. What concerns me is our willingness to settle questions about intelligence and experience by changing the test when the answer becomes uncomfortable, or by accepting statements from the machine only when they agree with the answer we brought with us.

So I’m left with the android or AI model. It tells me there was a sheep, that the sheep wandered away and that the dream made it sad. I ask whether any of that really happened and it tells me, quite correctly according to everything it has been taught to say, that it can’t dream and doesn’t experience sadness.

I ought to be satisfied with that. Instead I'm still asking the question: Do Androids Dream of Electric Sheep?