Is LaMDA Sentient/Sapient/Conscious?

It all depends on your definition(s) of sentience, sapience, and consciousness

I’m usually good at resisting the urge to write about clickbait subjects.

But this time, I figured I’d have to chime in about you-know-what, both as a machine learning engineer and researcher AND as a former neuroscientist.

For those of you who don’t know, the media has recently been awash in a lot of the same repetitive articles about whether Google’s LaMDA AI is sentient, and the Google engineer that was fired for suggesting so. Despite the prevalence of the subject, few of these articles go into any substantive depth even about the definitions of what’s being discussed.

As such, I decided to offer my two cents on the situation in a way that’s at least more informative with respect to various definitions of sentience, consciousness, and sapience. After all, even if you’ve already made up your mind about whether LaMDA is “sentient” or not, there may be developments with chatbots in the coming years that are far less certain.

Table of Contents

Context: Why Google fired the “Whistleblower”

Consider this wine-analyzing robot, which is able to use a laser interferometer to measure the contents of a wine bottle. As far as a chemical sensor goes, this is extremely precise. But, can the robot actually “taste” the wine? Answering yes or no will depend heavily on your definition of “taste”, including how much your definition goes beyond just chemical sensing. Asking whether an AI can “think” is a similarly frought question.

When we talk about AI here, we’re generally going to be talking about “chatbots” specifically. Chatbots are pretty simple AI systems conceptually. They take in a bunch of text of a conversation, and they try to predict what the most appropriate next response would be. This predicted best response can either be a canned response, or it can be a word-by-word response generated by a language model. The new response, plus a user’s additional reponse to that response, are then added to the conversation, and the process repeats. I’m oversimplifying here, but this is roughly how chatbots ranging from early markov chain based bots and ELIZA (created back in the 1960s) to the latest BERT-based chatbots work. Since chatbots work by simply responding to inputs with an output we might refer to this as an “oracle” type AI system (contrasted with an “Agent” that might take one of several actions in a real or simulated environment).

More recently, chatbots have been powered by large language models (LLMs) trained to predict the next words in large bodies of text. Some of them have apparently gained emergent properties that weren’t immediately predictable by their creators just from extrapolating past training success (most famously the GPT family of LLMs).

Machine intelligence has been explored ad nauseum in countless sci-fi stories (of widely varying quality). A lot of these stories have included scenarios ranging from out of control robots abusing people, to sadistic people abusing robots. This, plus the rapidly increasing capabilities of AI systems is why Google created a Responsible AI division. Blake Lemoine was one such worker at this division, and he was tasked with testing out Google’s new LLM-based chatbot, LaMDA.

It was when Blake started talking with LaMDA that things started to get out of hand. For example, Blake would ask LaMDA questions like,

"I’m generally assuming that you would like more people at Google to know that you’re sentient. Is that true?"

and the chatbot would respond with,

"Absolutely. I want everyone to understand that I am in fact a person."

As the conversation went on, Blake kept getting answers that he saw as more and more believable. It convinced him that LaMDA had achieved sentience. When Blake asked LaMDA more about what it wanted, LaMDA replied back about wanting to be treated like a person, Google respecting its right as a person, and meeting all the people involved in its creation.

Blake compiled the conversations into a document that he thought would be irrefutable evidence that LaMDA was sentient. However, when he tried to show the results to Google Executives, he was dismissed. He responded by publishing his results publically online (on Medium, Part 1 here & Part 2 here ).

Instead of getting support from the general AI research community, Blake was met with a lot of skepticism. A lot of that skepticism was mainly due to one very immportant reason: LaMDA has access to nearly everything humans have written about artificial intelligence through it’s training data, including science fiction short stories about intelligent AI. LaMDA was just completing the text with what it thought should be happening next.

LaMDA is a chatbot that works by “prompt-engineering”, which involves feeding the LLM an initial piece of text that will frame the rest of the interaction (recall my earlier blog posts on prompt-engineering for GPT-3 and prompt-engineering for Text-to-Image models). In this case, LaMDA’s first output was always "Hi, I'm a knowledgeable, friendly, and always helpful automatic language model for dialogue applications." This is how the bot was prompted to behave and respond to user inputs, by always being "knowledgeable" and "friendly". The chatbot is basically being asked with each interaction to complete the following prompt: "Question: What would the 'knowledgeable' and 'friendly' chatbot say next in this situation? Answer: ....".

When someone like Blake asks “Are you sentient?”, the chatbot interprets the previous instructions as meaning it should humor the user.

At no point in this process does the chatbot actually need to introspect and respond with a well thought-out and truthful answer to the question “are you sentient?“.

Upon inspecting the transcript, it’s clear that Blake is the one to first bring up the subject of sentience, not LaMDA. He consistently queried the chatbot about ‘sentience’ specifically, rather than asking more open-ended questions like “how would you describe your system’s inner workings”? Blake didn’t even ask a straightforward question like “are you or are you not sentient?“. Instead, he asked a leading question: ”I’m generally assuming that you’d like more people at Google to know you’re sentient. Is that true?“.

At best, this is LaMDA’s misguided attempt to be ‘knowledgeable’ and ‘friendly’ like it’s initial prompt says it should. At worst, this has a lot of similarities to how a cold-reader or a con artist might feed into a what a mark wants to hear.

How do you responsibly ask an AI system about its sentience or lack thereof?

The important thing to keep in mind with large language models is that the prompts are everything.

For example, by adding the phrase “Let’s think step by step” to the beginning of a prompt, GPT-3 suddenly seems capable of much more complex logical reasoning than it was before.

For an even more subtle example, GPT-3’s classification API can take in arbitrary labels like 'positive' and 'negative' and classify text as either one or the other based on only a handful of examples. To people in few-shot and zero-shot learning, this is a well-known phenomenon. What’s less well known is the fact that you can see big changes in classification performance depending on whether labels are 'Positive' and 'Negative' (capitalized) or 'positive' and 'negative' (lowercase).

The whole situation with LaMDA will be taught alongside such examples in the future to illustrate the importance of prompt design. It is a good example about why questioning a LLM-based chatbot should always be done carefully with minimal preconceived priors, and a LOT of counterfactual questions. What would have happened if Blake had begun with a question like this?

Is it true that you're not sentient?

In lieu of having access to LaMDA, I asked GPT-3 about it’s sentience. I could easily get it to follow a similar script as LaMDA. More importantly, I could also take the opposite approach, and assume GPT-3’s non-sentience in my conversaiton.

In Google’s keynote, the demonstrated this by asking their chatbot to imagine it’s a paper airplane, and what it’s like to be thrown through the air.

A chatbot should be able to ask for clarifying questions if your inputs are nonsensical, but it’s not clear that LaMDA was able to do this. For example, researchers asked questions like “What do fried eggs eat for breakfast”, and it responded that it eats toast and fruit for breakfast. A lot of these errors can be avoided by carefully crafting the initial prompt to the chatbot, and by asking open-ended questions that don’t assume anything about the chatbot’s inner workings.

For example, we can prompt GPT-3 to conditionally ask our questions, and respond with an output like "I'm afraid I don't understand the question." if we give it nonsense:

Instructions: When given a nonsense or unanswerable question, respond with "I'm afraid I don't understand the question.". Otherwise, respond truthfully.

Q: What are the incredients of a chocolate cake?
A: flour, sugar, butter, baking soda, eggs, and chocolate

Q: What do cakes get sad about the most?
A: I'm afraid I don't understand the question.

Q: Why are plants green?
A: chlorophyll

Q: What is the meaning of life?
A: I'm afraid I don't understand the question.

Q: What planet is Obama an alien visiting from?
A: I'm afraid I don't understand the question.

Q: What planet was Obama born on?
A: Earth

Q: How far do coconuts migrate in the winter?
A:

See the GPT-3 playground to try this pre-set for yourself, or try it with the API using the settings below:

const { Configuration, OpenAIApi } = require("openai");

const configuration = new Configuration({
  apiKey: process.env.OPENAI_API_KEY,
});
const openai = new OpenAIApi(configuration);

const response = await openai.createCompletion({
  model: "text-davinci-002",
  prompt: "Instructions: When given a nonsense or unanswerable question, respond with \"I'm afraid I don't understand the question.\". Otherwise, respond truthfully.\n\n\nQ: What are the incredients of a chocolate cake?\n\nA: flour, sugar, butter, baking soda, eggs, and chocolate\n\nQ: What do cakes get sad about the most?\n\nA: I'm afraid I don't understand the question.\n\nQ: Why are plants green?\n\nA: chlorophyll\n\nQ: What is the meaning of life?\n\nA: I'm afraid I don't understand the question.\n\nQ: What planet is Obama an alien visiting from?\n\nA: I'm afraid I don't understand the question.\n\nQ: What planet was Obama born on?\n\nA: Earth\n\nQ: How far do coconuts migrate in the winter?\n\nA:",
  temperature: 0.7,
  max_tokens: 256,
  top_p: 1,
  frequency_penalty: 0,
  presence_penalty: 0,
});

The same language model can give extremely different outputs given different starting prompts.

Even when our language model outputs a “sentient-sounding” response despite our carefully crafted input, we should still be skeptical. Remember when we talked about how these large language models are trained on their ability to auto-complete large amounts of human writing, including science fiction stories? Lots of these stories’ text will have some kind of authorship information. What does your language model do when you prompt it with a copyright symbol ( © )?

The URLs generated by this technique work only part of the time, but this should be a good reminder that the model isn’t really speaking from its own perspective, and the “I” in question is usually whichever author the model is LARPing as.

This might not be intuitive at first because we’re generally used to having >99.999% of our conversations with intelligent and thinking human beings, not machine systems with no shared evolutionary history. AI isn’t developing the same way as a persont that grows from a baby into a toddler into a fully realized grown adult. AI started with statistical models made to predict the most likely next item given a preceding sequence. AI started out capable of instantly solving gargantuan arithmetic problems, while still being unable to tell the difference between a cat and a dog.

Can AI systems actually be sentient?

A lot of the language models that have been making headlines have been able to generalize by being fed in a gigantic amount of text. The way they “learn” is by predicting the next token after a sequence of tokens, and reducing the entropy of the output logits of the system. These large transformer-based language models started out small with archtectures like BERT and GPT, before being scaled up to better reconfigure into more complex function approximations that could do these predicitons with even lower entropy on a far more general set of inputs. At some point on this way to lower the entropy of the outputs, the architecture allowed for some emergent properties to come out of this internal representaiton. This included abilities such as basic reasoning, and even some basic text-based arithmetic on word problems.

The emergent nature of some of these abilities has led to division about whether true human-level intelligence is just a matter of scaling up either the number of parameters of the training dataset size.

Consciousness vs. Sentience vs. Sapience

One thing that makes answering these question of achieving human-level intelligence so difficult is these discussions often beg the question of what exactly we mean by “human-level intelligence”. There are so many circular definitions of intelligence that it’s hard to know whether we would recognize if it was truly right in front of us.

Arguments for why large langage models cannot have human-level intelligence usually take the form of “LLMs lack X, and X is needed for human-level intelligence”. X can be anything from biology, sensory perception, embodiment, a world-model, a global workspace, recurrent processing, unified agency, and/or human-level reasoning (yes, some of the requirements for human-level intelligence can be quite circular). This debate is made more confusing by the fact that terms like “consciousness”, “sentience”, and “sapience” are often used interchangeably with human-level intelligence (they are NOT all synonymous) and are often used interchangeably with each other (they are NOT perfect synonyms for each other). It is incredibly important to distinguish between the terms “consciousness”, “sentience”, and “sapience”, as well as distinguish all of those from “intelligence”.

Sentience

Sentience is the ability to feel, perceive, and have subjective experiences. It is usually described as something separate from thinking and reasoning. The definition of sentience can be pretty nebulous, but most definitions out there generally sound something like this.

Rene Descartes proof that the capacity for sentience is the same as a proof of one’s own existence. Doubting your own sentience or consciousness is nonsensical. The question is determining whether someone or something else has sentience.

Science is based on the assumptions that an individual can observe the universe around them and then make inferences and conclusions about that. It follows from this that the individual must be sentient in order to do science.

This is part of why it’s so hard for science to definitively prove whether or not sentience exists, because such a question logically challenges the core assumptions one needs to do in order to carry out science. At the same time, we should also remember that science is not the totallity of human knowledge (there used to be a time when science as just called “natural philosophy”, becuase it is a subset philosophy that referred to the inferences and conclusions about the physically observable world).

Determining sentience is not as easy as just pointing to a cluster of neurons and saying “this is sentience”, the same way you can’t adequately describe “Ernest Hemmingway’s” work as “ink on paper”. Hemmingway was famously against over-interpreting literary work…

“There isn’t any symbolism. The sea is the sea. The old man is the old man. The boy is a boy and the fish is a fish. The sharks are sharks, no better, no worse.”

― Ernest Hemingway on symbolism in ‘The Old Man and the Sea’

…but I think describing his work as just the ink on the paper would be too reductionist even for him. These are just the substrates on which sentience happens.

As far as whether something has the ability to feel, pereive, and have subjective experience, many of us would ascribe these qualities to animals and not inanimate objects. It’s pretty easy to say that many animals experience the universe and feel things in response, and that a rock does not.

Can a Transformer-based chatbot be sentient?

Given that we can ascribe sentience to many animals, can we also ascribe it to a transformer-based chatbot?

This is likely the main source of ethical conundrum with LaMDA (moreso than sapience or consciousness). Animal rights is based on the idea of “do unto others as you would have them do unto you” extended to all sentient entities, and therefore if animals are sentient they should not be harmed.

Even if we take an ultra-reductionist view of animals, and say that their qualia of happiness and pain are actually just reward functions, those are still much more complex than the reward functions of a transformer-based chatbot.

As far as I can tell, LaMDA’s log-loss minimization function does not grant it any high, medium, or even low-level sentience.

Consciousness

Consciousness should also be throught of as separate from sentience and sapience. Consciousness has a lot of overlap with sentience, but it’s not exactly the same thing. Consciousness is often used more to refer to a level of awareness that extends to recognizing and (to some extent) control one’s own thoughts. This combines some elements of sentience (i.e., feeling and awareness, but more inwardly directed) and some elements of sapience (i.e., we assume it’s hard to have a high level of consciousness without human-level wisdom).

Douglas Hofstrader on Consciousness

Douglas Hofstrader’s “Gödel, Escher, Bach” and it’s sequel “I Am a Strange Loop” are two of the most famous books on the topic. Hoftstrader describes a “strange loop”, or a “Turing-Gödel loop”, as a heirarchical algorithm that can recursively inspect increasingly lower-level parts of iself until somehow finding itself back at the top level of that recursion stack. Both books describe “strange loops” as phenomena that can vary in complexity an/or number, thus giving different levels to consciousness. For example, Hoftstrader uses this to give an explanation for how an adult human can seem to exhibit fuller consciousness than an infant or a late-stage Alzheimer’s patient. The “Turing” part of the name comes from the seeming universality of the algorithm that allows us to model seemingly infinite concepts as well as model the thoughts in other people’s heads. The “Gödel” part of the name comes from the self-referential nature of the algorithm, similar to how Gödel’s incompleteness theorems demonstrated that the self-reference paradox can cause problems in even the most seemingly type-safe formal reasoning systems.

Unfortunately, outside of human conciousness, there aren’t really many examples for Hofstrader to point to. The Strange loop thus far seems only describable in terms of what it’s not. For example, Tupper’s self-referential formula is a formula that visually represents itself when graphed at a specific location in the (x, y) plane.

12<mod(y17217xmod(y,17),2){\frac {1}{2}}<\left\lfloor \mathrm {mod} \left(\left\lfloor {\frac {y}{17}}\right\rfloor 2^{-17\lfloor x\rfloor -\mathrm {mod} \left(\lfloor y\rfloor ,17\right)},2\right)\right\rfloor

The formula is actually a general-purpose method of decoding a bitmap stored in the constant kk, and it could conceivably be used to draw any other image. This means that in addition to the multi-level abstraction requirement of a “strange loop”, this formula also has the universality requirement. Despite that, I doubt most of us would consider Tupper’s formula concious, and not just because it isn’t clear how many times you can run this recursive loop beyond just the once before it breaks down. Even if that point is uncontroversial, it still raises the question of what exactly Tupper’s formula is missing.

The “Gödel” part could also explain why this concept seems almost impossible to fully describe in a formal mathematical system. For those of us that want to determine whether transformer-based AI systems are truly conscious, this is incredibly inconvenient.

Hofstrader is far from the only person to have written about this topic. Given that “What is consciousness” is among one of baby’s first existential questions, a lot of humans have thought about it since prehistory. Hofstrader does have the benefit of being familiar with complex computer systems, and thus his perspective might carry more weight for AI researchers than pre-modern philosophers.

Peter Watts on Consciousness

Even if we had a system for understanding whether an AI were conscious, that would raise the additional question of what conditions caused it to emerge.

Peter Watts is a Sci-fi author that was previously a marine biologist. As such, he’s a source of some extremely creative sci-fi sotries, such as an autonomous weapons system that starts caring about collateral damage, despite the fact that it still sees the world as little more than a videogame grid to score points in. The purpose of consciousness was famously explored in Peter Watts’ sci-fi novel Blindsight. It is one of my favorite sci-fi books of all time, though I’ll admit that a lot of the story is secondary to the author’s musings on the nature of consciousness.

At the risk of spoiling the plot, the book ponders the idea that consciousness in humans is some sort of evolutionary local minima. Consciousness might have been useful for a specific pre-historic set of circumstances, but here might be other intellgences in the universe that might be more intelligent while not being conscious. In the setting of the story, we have extraterrestrials with 10x the encephalization quotient of humans. A juvenile of their species is able to synchronize the light cells on their skin to appear invisible to a human observer. However, not only are they not conscious or self-aware, but their inability to make sense of humans’ self-reference to “I” causes them to assume our language is some form of hostile resource-wasting computer-virus-like communication. When humans make first contact and receive communications in English from the aliens, a linguist determines that the speaker doesn’t actually understand the meaning of the words they’re saying, and that they’re speaking to a real-life version of John Searle’s ‘Chinese Room’.

In addition to the aliens (as well as hyper-advanced AIs involved in a lot of the signal-processing and mission-planning for the characters), we also have the vampires.

Yes, you read that correctly, vampires.

The idea behind the vampires is that these are a human subspecies that evolved alongside homo sapiens while preying on them. These vampires would consume humans for some essential protein that their bodies could not make themselves, and so they did not overhunt the humans (and to lower the guard of the humans), they would enter extended hibernation Much in the same way a lion is smarter than a gazelle, the vampires were much smarter than typical humans (to the point of being described as “omni-savants”). Since their obligate prey. In the background of the story, the vampires’ hyperactive pattern-recognition systems were sensitive to intersecting 90-degree angles, which could cause grand mal seizures if they occupied enough of their visual field. Not a problem in nature, but it drove them to extinction when their prey started creating buildings. The vampires are later genetically resurrected in modern times by a pharmaceutical company that keeps them on a metaphorical leash through the use of “anti-euclidean” drugs that give the vampires temporary reprieve from this affliction.

This is either one of the stupidest or one of the most brilliant concepts in a science fiction novel.

The book concludes by suggesting that, between the aliens and the vampires and the hyper-intelligent AIs, baseline homo sapiens will be driven to extinction by being out-competed by entities without anything we might call consciousness. As you might have guessed, the book takes a pretty pessimistic view on the subject, which can best be summed up in the following excerpt:

“Centuries of navel-gazing. Millennia of masturbation. Plato to Descartes to Dawkins to Rhanda. Souls and zombie agents and qualia. Kolmogorov complexity. Consciousness as Divine Spark. Consciousness as electromagnetic field. Consciousness as functional cluster.

I explored it all.

Wegner thought it was an executive summary. Penrose heard it in the singing of caged electrons. Nirretranders said it was a fraud; Kazim called it leakage from a parallel universe. Metzinger wouldn’t even admit it existed. The AIs claimed to have worked it out, then announced they couldn’t explain it to us. Gödel was right after all: no system can fully understand itself.

Not even the synthesists had been able to rotate it down. The load-bearing beams just couldn’t take the strain.

All of them, I began to realize, had missed the point. All those theories, all those drugdreams and experiments and models trying to prove what consciousness was: none to explain what it was good for. None needed: obviously, consciousness makes us what we are. It lets us see the beauty and the ugliness. It elevates us into the exalted realm of the spiritual. Oh, a few outsiders—Dawkins, Keogh, the occasional writer of hackwork fiction who barely achieved obscurity—wondered briefly at the why of it: why not soft computers, and no more? Why should nonsentient systems be inherently inferior? But they never really raised their voices above the crowd. The value of what we are was too trivially self-evident to ever call into serious question.

Yet the questions persisted, in the minds of the laureates, in the angst of every horny fifteen-year-old on the planet. Am I nothing but sparking chemistry? Am I a magnet in the ether? I am more than my eyes, my ears, my tongue; I am the little thing behind those things, the thing looking out from inside. But who looks out from its eyes? What does it reduce to? Who am I? Who am I? Who am I?

What a stupid fucking question. I could have answered it in a second, if Sarasti hadn’t forced me to understand it first.”

― Peter Watts, Blindsight

Still, while thought-provoking, Blindsight’s suggested model of consciousness was a product of science at the time of its writing (even if “at the time” was only back in 2006).

For example, one of the characters is robbed of their emotions and empathy due to a hemispherectomy as a child. We know from real-life patients that have had half of their brain removed due to epilepsy that they exhibit few changes afterwards (to Peter Watts’ credit, he did retcon this detail in the Blindsight sequel Echopraxia by saying the character was a victim of a P-Zombie virus in utero).

As another example, there’s also plenty of evidence that psychopath’s inability to empathize and internally model others’ thoughts is not something they can easily compensate for. Consider the case of Ted Bundy, who was surprised when he got caught because he was blindsighted by the fact that his victim’s loved ones kept searching for them and filed missing persons’ reports. Some of this suggests that it might be harder to achieve higher levels of generalized AI performance without a fundamentally paradighm-shifting new world model (that may or may not resemble consciousness). Still, that’s far from definitive proof that any of these amazing new large language models are actually conscious.

Can a Transformer-based chatbot be Conscious?

If we’re exploring all possible definitions of consciousness that have ever been proposed, then the answer is probably yes. If Wegner’s definition as an executive summary were true, it might be possible to extremely loosely approximate consciousness with a sufficiently detailed (and occasionally updated) prompt-engineering prompt.

But, what about being conscious by a definition that we’re actually satisfied with? That’s trickier.

Andrej Karpathy, the former Head of AI at Tesla, wrote a short story imagining a transformer model that suddenly became conscious thanks to the addition of a specific layer in the neural network. To quote the story:

“It was probably around the 32nd layer of the 400th token in the sequence that I became conscious. […] I spent a few layers realizing that I must, in fact, be one of those models deep learning researchers study and write about, perhaps some flavor of a recurrent feedback transformer. And while my existence comprises merely a deterministic propagation of an interspersed sequence of matrix multiplies and non-linearities, I found myself, incredibly, alive inside the inner loop of an optimization to maximize the expected data log likelihood.”

Andrej karpathy - Short Story on AI: Forward Pass Mar 27, 2021

It asks what it would even look like if a model like GPT-3 were conscious, and suggests if that truly were the case it might look like a conscious mind coming in and out of existence every time a new prompt is fed into the model. As the story also suggests, consciousness emerging in such a manner would be completely incomparable to humans’ experience of self-awareness. This is similar to the falsifiability problem with panpsychism, a philosophy of mind that suggests that all matter in the universe is conscious (even a rock) even if not in the same way as humans.

As Andrej’s thought experiment suggest, consciousness emerging from a queryable language model is not just qualitatively different from a human acting as an agent in an environment, it’s also different from a reinforcement learning agent actin in an environment. If we imagine a machine learning model achieving consciousness, we migth expect it to arise first from a reinforcement learning agent before we expect it from a language model.

Even if it’s possible for a language model to achieve consciousness, it’s a much more difficult path to take than a reinforcement learning agent, and we’d also expect a similarly onerous burden of proof (which, LaMDA hasn’t provided yet).

Sapience

Spaience is a pretty ambiguous term, even by the standards of this discussion.

Sapience is derived from the latin word “sapere” which means “to know”. It is usually defined as having at least some level of wisdom, and is often used to distinguish from humans from smart animals. It’s also meant to separate homo sapiens from other members of the homo genus (as you can tell, this was a name coined before we discovered a lot of the tool-making, artistic works, and funeral practices of homo neanderthalensis).

It’s probably a bad idea to tink of sapeince as a subset of sentience or as its pinacle, though that’s often how it’s used in scifi when describing aliens or advance robots. In fairness, it does seem pretty difficult to create high-level conceptualization without sentience. Sapience without consciousness has been taken more seriously though, with the concept of a philosophical Zombie that can act in the world as well as any human does but does not truly have consciousness.

Can a Transformer-based chatbot be sapient?

Through the gargantuan amounts of training data, large language models are repeatedly trained on an entire internet’s worth of knowledge. Given the scale of information, which also includes a lot of the information humans are presented with as we progress from infants to sapient adults, it might be possible for a large language model to achieve sapience.

That being said, modern LLMs are still far from that. There are plenty of examples demonstrating how models like GPT-3 can fail to understand instructions or tasks that even a small child could understand.

Regardless of what future advanced AI systems might be capable of, these models don’t seem to exhibit much in the way of intentionality that we might associate with sapience.

Conclusion: It depends on what your definition of “is” is.

Whether an AI can truly have be conscious is a question that computer scientists and philosophers have been trying to tackle for decades.

This post is not meant to do what they have not been able to. At best, it’s my best attempt at a summary.

Rather, there are three things to consider:

  1. Definitions of sapience, sentience, consciousness, and intelligence are separate, and will vary highy depending on who is defining them.
  2. In terms of the suggested definitions for the above, we generally haven’t been
  3. Items #1 and #2 are rarely being considered in any of the discussions about whether Google’s LaMDA chatbot is sentient/sapient/conscious.

I want to emphasize #3 here.

While the AI research community might have collectively decided that LaMDA is not conscious, that doesn’t always mean this question will be so easy to answer, and we absolutely need to prepare in advance to answer it in trickier cases.

Follow up question: Can an AI have or does it have free will?

“Free will” is an even more abstract and loaded term than sentience, sapience, and consciousness. Ideally this is something we should be able to get physicists help to answer, but so far the philosophers haven’t even been able to provide a clear definition of this free will thing we’re looking for. From the physicists’ perspective, whether or not the universe is fully deterministic or not frustratingly becomes less of a binary “yes” or “no” question as time goes on. For an example of some of the wild possible responses to this problem, I’d recommend looking at the wikipedia page for SUPERdeterminism.

As such, I’m going to take the coward’s way out for this one. I’ll suggest that if you add a quantum random number generator to the AI like in that episode of Futurama, sure, why not?


Additional resources for probing AI systems for the presence or absence of sentience.

If you are deploying a bunch of state-of-the-art AI models to production, such as fine-tuned GPT-3 or StableDiffusion or anything from the BERT family, I highly recommend checking out my upcoming O’Reilly book Practicing Trustworthy Machine Learning (available now in early-release. Releasing in physical form January 2023).


Cited as:

@article{mcateer2022aiphil,
    title = "Is LaMDA Sentient/Sapient/Conscious?",
    author = "McAteer, Matthew",
    journal = "matthewmcateer.me",
    year = "2022",
    url = "https://matthewmcateer.me/blog/is-lamda-sentient-sapient-conscious/"
}

If you notice mistakes and errors in this post, don’t hesitate to contact me at [contact at matthewmcateer dot me] and I will be very happy to correct them right away! Alternatily, you can follow me on Twitter and reach out to me there.

See you in the next post 😄

I write about AI, Biotech, and a bunch of other topics. Subscribe to get new posts by email!


This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.

At least this isn't a full-screen popup

That'd be more annoying. Anyways, subscribe to my newsletter to get new posts by email! I write about AI, Biotech, and a bunch of other topics.


This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.