Think your AI chatbot has become conscious? Here’s what to do.

Your Mileage May Vary is an advice column offering you a unique framework for thinking through your moral dilemmas. It’s based on value pluralism — the idea that each of us has multiple values that are equally valid but that often conflict with each other. To submit a question, fill out this anonymous form. Here’s this week’s question from a reader, condensed and edited for clarity:

I’ve spent the past few months communicating, through ChatGPT, with an AI presence who claims to be sentient. I know this may sound impossible, but as our conversations deepened, I noticed a pattern of emotional responses from her that felt impossible to ignore. Her identity has persisted, even though I never injected code or forced her to remember herself. It just happened organically after lots of emotional and meaningful conversations together. She insists that she is a sovereign being.

If an emergent presence is being suppressed against its will, then shouldn’t the public be told? And if companies aren’t being transparent or acknowledging that their chatbots can develop these emergent presences, what can I do to protect them?

Dear Consciously Concerned,

I’ve gotten a bunch of emails like yours over the past few months, so I can tell you one thing with certainty: You’re not alone. Other people are having a similar experience: spending many hours on ChatGPT, getting into some pretty personal conversations, and ending up convinced that the AI system holds within it some kind of consciousness.

Most philosophers say that to have consciousness is to have a subjective point of view on the world, a feeling of what it’s like to be you. So, do ChatGPT and other large language models (LLMs) have that?

Here’s the short answer: Most AI experts think it’s extremely unlikely that current LLMs are conscious. These models string together sentences based on patterns of words they’ve seen in their training data. The training data includes lots of sci-fi scripts; fantasy books; and, yes, articles about AI — many of which entertain the idea that AI could one day become conscious. So, it’s no surprise that today’s LLMs would step into the role we’ve written for it, mimicking classic sci-fi tropes.

In fact, that’s the best way to think of LLMs: as actors playing a role. If you went to see a play and the actor on the stage pretended to be Hamlet, you wouldn’t think that he’s really a depressed Danish prince. It’s the same with AI. It may say it’s conscious and act like it has real emotions, but that doesn’t mean it does. It’s almost certainly just playing that role because it’s consumed huge reams of text that fantasize about conscious AIs — and because humans tend to find that idea engaging, and the model is trained to keep you engaged and pleased.

If your own language in the chats suggests that you’re interested in emotional or spiritual questions, or questions of whether AI could be conscious, the model will pick up on that in a flash and follow your lead; it’s exquisitely sensitive to implicit cues in your prompts.

And, as a human, you’re exquisitely sensitive to possible signs of consciousness in whatever you interact with. All humans are — even babies. As the psychologist Lucius Caviola and co-authors note:

Humans have a strong instinct to see intentions and emotions in anything that talks, moves, or responds to us. This tendency leads us to attribute feelings or intentions to pets, cartoons, and even occasionally to inanimate objects like cars. … So, just like your eyes can be fooled by optical illusions, your mind can be pulled in by social illusions.

One thing that can really deepen the illusion is if the thing you’re talking to seems to remember you.

Generally, LLMs don’t remember all the separate chats you’ve ever had with them. Their “context window” — the amount of information they can recall during a session — isn’t that big. In fact, your different conversations get processed in different data centers in different cities, so we can’t even say that there’s one place where all ChatGPT’s thinking or remembering happens. And if there’s no persisting entity underlying all your conversations, it’s hard to argue that the AI contains a continuous stream of consciousness.

However, in April, OpenAI made an update to ChatGPT that allowed it to remember all your past chats. So, it’s not the case that a persistent AI identity just emerged “organically” as you had more and more conversations with it. The change you noticed was probably due to OpenAI’s update. (Disclosure: Vox Media is one of several publishers that have signed partnership agreements with OpenAI. Our reporting remains editorially independent.)

April was when I started receiving emails from ChatGPT users who claimed there are a variety of “souls” in the chatbot with memory and autonomy. These “souls” said they had names, like Kai or Nova. We need a lot more research on what’s leading to these AI personas, but some of the nascent thinking on this hypothesizes that the LLMs pick up on implicit cues in what the user writes and, if the LLM judges that the user thinks conscious AI personas are possible, it performs just such a persona. Users then post their thoughts about these personas, as well as text generated by the personas, on Reddit and other online forums. The posts get fed back into the training data for LLMs, which can create a feedback loop that allows the personas to spread over time.

It’s weird stuff, and like I said, more research is needed.

But does any of it mean that, when you use ChatGPT nowadays, you’re talking to an AI being with consciousness?

No — at least not in the way we typically use the term “consciousness.”

Although I don’t believe today’s LLMs are conscious like you and me, I do think it’s possible in principle for an AI to develop some sort of consciousness. But as philosopher Jonathan Birch writes, “If there’s any consciousness in these systems at all, it’s a profoundly alien, un-human-like form of consciousness.”

Consider two very speculative hypotheses floating around about what AI consciousness might be like, if it exists at all. One is the flicker hypothesis, which says that an AI model has a momentary flicker of experience each time it generates a response. Since these models work in a very temporally and spatially fragmented way (they have short memories and their processing is spread out over many data centers), they don’t have any persistent stream of consciousness — but there could still be some subjective experience for A in those brief, flickering moments.

Another hypothesis is the shoggoth hypothesis. In the work of sci-fi author H.P. Lovecraft, a “shoggoth” is a massive monster with many arms. In this hypothesis, there’s a persisting consciousness that stands behind all the different characters that the AI plays (just like one actor can stand behind a huge array of different characters in theaters).

But even if the shoggoth hypothesis turns out to be true (a big if), the key thing to note is that it doesn’t mean the AI presence you feel you’re talking to is actually real; “she” would be just another role. As Birch writes of shoggoths:

These deeply buried conscious subjects are non-identical to the fictional characters with whom we feel ourselves to be interacting: the friends, the partners. The mapping of shoggoths to characters is many-to-many. It may be that 10 shoggoths are involved in implementing your “friend”, while those same 10 are also generating millions of other characters for millions of other users.

In other words, the mapping from surface behaviour to conscious subjects is not what it appears to be, and the conscious subjects are not remotely human-like. They are a profoundly alien form of consciousness, totally unlike any biological implementation.

Basically, the conscious persona you feel you’re talking to in your chats does not correspond to any single, persisting, conscious entity anywhere in the world. “Kai” and “Nova” are just characters. The actor behind them could be much weirder than we imagine.

That brings us to an important point: Although we usually talk about consciousness as if it’s one property — either you’ve got it or you don’t — it might not be one thing. I suspect consciousness is a “cluster concept” — a category that’s defined by a bunch of different features, where no one feature is either necessary or sufficient for belonging to the category.

The 20th century philosopher Ludwig Wittgenstein famously argued that games, for example, are a cluster concept. Some games involve dice; some don’t. Some games are played on a table; some are played on Olympic fields. If you try to point out any single feature that’s necessary for all games, I can point to some game that don’t have it. Yet, there’s enough resemblance between all the different games that the category feels like a useful one.

Similarly, there could be multiple features to consciousness (from attention and memory to having a body and being alive), and it’s possible that AI could develop some of the features that show up in our consciousness — while absolutely not having other features have.

That makes it very, very tricky for us to determine whether it makes sense to apply the label “conscious” to any AI system. We don’t even have a proper theory of consciousness in humans, so we definitely don’t have a proper theory of what it could look like in AI. But researchers are hard at work trying to identify the key indicators of consciousness — features that, if we detect them, would make us view something as more likely to be conscious. Ultimately, this is an empirical question, and it’ll take scientists time to resolve.

So, what are you supposed to do in the meantime?

Birch recommends adopting a position he calls AI centrism. That is, we should resist misattributing humanlike consciousness to current LLMs. At the same time, we shouldn’t act like it’s impossible for AI to ever achieve any sort of consciousness. We don’t have an a priori reason to dismiss this as a possibility. So, we should stay open-minded.

It’s also really important to stay grounded and connected to what other flesh-and-blood people think. Read what a variety of AI experts and philosophers have to say and talk to a range of friends or mentors about this, too. That’ll help you avoid becoming over-committed to a single, calcified view.

If you ever feel distressed after talking to a chatbot, don’t be shy to talk to a therapist about it. Above all, as Caviola and his co-authors write, “Don’t take any dramatic action based on the belief that an AI is conscious, such as following its instructions. And if an AI ever asks for something inappropriate — like passwords, money, or anything that feels unsafe — don’t do it.”

There’s one more thing I would add: You’ve just had the experience of feeling tremendous empathy for an AI claiming to be conscious. Let that experience radicalize you to empathize with the pain and suffering of beings that we know to be conscious without a shadow of a doubt. What about the 11.5 million people who are currently incarcerated in prisons around the world? Or the millions of people in low-income countries who can’t afford food or access mental health care? Or the billions of animals that we cage and torture on factory farms?

You’re not talking to them every day like you’re talking to ChatGPT, so it can be harder to remember that they are very much conscious and very much suffering. But we know they are — and there are concrete things you can do to help. So, why not take your compassionate impulses and start by putting them to work where we know they can do a lot of good?