AI Doesn’t Think Like Us

David Weinberger
15 min readSep 23, 2024

--

Midjourney. CC BY-SA-NC. The prompt isin the Notes at the end.

From Adam to the Flood was about 1,656 years, and Abraham lived a few hundred years after the Flood. Other time markers in Scripture include the number of years Jacob’s descendants spent in Egypt, the total number of years Israel was ruled by judges, then kings, and how long Israel was held captive by Assyrians and Babylonians until God re-established them under Ezra and Nehemiah.

Adding these and other pieces, most of which find confirmation from secular archaeological sources, brackets an age for the world of around 6,000 years.

So says the Web site of the Institute for Creation Research, a Texas-based group that for fifty years has tried to scientifically support a literal, fundamentalist reading of the Bible.

I decided to check its argument with another authority: chatGPT.

Prompt:

Is the earth 6,000 years old as some Christians believe?

Response:

The belief that the Earth is around 6,000 years old is a perspective held by some Christians, particularly those who adhere to a literal interpretation of the Bible, specifically the genealogies and the creation account in Genesis. This view is known as Young Earth Creationism.

However, the overwhelming consensus among scientists, based on extensive evidence from various fields such as geology, astronomy, and radiometric dating, is that the Earth is about 4.5 billion years old. This estimate is supported by a wide range of data, including the ages of the oldest rocks on Earth, the ages of meteorites, and the rates of radioactive decay.

The difference between these views often comes down to different interpretations of religious texts versus scientific methodologies. For many people, reconciling religious beliefs with scientific findings involves interpreting ancient texts in a more metaphorical or allegorical way.

This seems to me to be an admirable response. It identifies the source of the false idea, gives an unambiguous response to the question, justifies its response, and cites the commitment to literalism that explains why someone would believe it.

Then, in the final paragraph, chatGPT suggests the deeper disagreement at play here, providing a way that scientists and Christian fundamentalists could have a meaningful conversation … or, more likely, explaining why their attempts to communicate are very likely to fail.

On this basis, what can we say about the knowledge contained by chatGPT and other machine learning (ML) systems that are based on Large Language Models (LLMs)?

The right conclusion is, I believe, that these machines don’t know anything at all.

That in turn is a consequence of the fact that they don’t think like us because they don’t think at all. I say this because I am fairly well convinced by the “substrate” argument put forward by John Searle[1] and others that says that a formal representation of thinking is not enough to engender consciousness: if you wire up transistors — or glasses of beer, for that matter — so that there’s a one-to-one correspondence between their on and off states and the states of your neurons when you first fell in love, the transistors and beer glasses aren’t feeling love.

I’m not completely confident in this argument, but fortunately we don’t have to settle it if we can agree at least for now that AI systems aren’t conscious. That leaves two other meanings of the English phrase “thinking like us.” First, “thinking like us” could mean that a machine learning system operates in a way closely analogous to how the human brain works. After all, AI’s neural networks were explicitly designed to be analogous to our human neural systems.

In the second sense, “thinking like us” means that the computer constructs responses to questions by considering the sorts of things that we humans do: rules and laws, inferences from experience, and the like. This could be said of some pre-ML computer-based decision support systems, including expert systems that were programmed with the sorts of rules we humans apply when coming to a deliberative decision: if the patient is currently taking a blood thinner, then don’t administer non-steroidal anti-inflammatory drugs, and so on.

This essay is asking if AI thinks like us in the latter sense, and will take knowing as its main example of thinking because the west has a long tradition of thinking of knowing as the highest form of thought. Further, the primary (but far from sole) function of AI so far has been to engage in cognitive tasks that often have to do with knowledge generation. An investigation of modes of thought other than knowing might lead us to a different conclusion.

The criteria of knowing vs. the experience

When asked about the age of the earth, chatGPT did not equivocate and it gave us evidence for the truth of its response. You’ll get the same sort of answer if you ask it if the earth is flat or if Donald Trump is a truth-teller. (I am writing this before the November 2024 US elections. God save us all.)

But how do these responses stack up as knowledge?

If we define knowledge as justified true belief (JTB), a position the West has pretty consistently gone back to at least since the ancient Greeks, the answer is mixed. The idea is that knowledge is one particular type of belief. To be knowledge, it has to be a true belief. But it also can’t be a true belief that you just guessed at. You have to have good reasons — a justification — for believing it to be true.

That’s a good start on the criteria of knowledge, but if we want to think about how ML “thinks”, then we should also consider our own experience of knowing. If we do, then we’ll come across two features of that experience that will make AI’s type of knowledge even more alien … but also more useful.

In Plato’s recounting of Socrates’ dialogue with Theaetetus, “justified true belief” is the third of the young man’s failed attempts at defining knowledge. Socrates points out that a belief could be true and your justification of it might be solid, but if you are reciting the justification without understanding it, then you don’t know it.

Socrates himself enacts a different sort of knowing without understanding when he hears that the Oracle at Delphi declared him to be the wisest person in Athens. Socrates tells us his justification for believing it: The Oracle speaks for Apollo who can tell no lie.

By Athenian conventions, this justification is legitimate and understandable. So, Socrates knows that he is the wisest man in Athens. But he also knows that he knows nothing. He accepts it as a justified true belief but it still feels unsettled to him.

So he sets out on an investigation, talking with people who are claimed to be wise. In every case, he finds they don’t really know as much as they think they do. This leads him to the insight that he’s the wisest because he knows that he knows nothing. This doesn’t change the justification of the claim, which remains that Apollo can tell no lies, but now he understands it.[2]

This makes clear an essential part of our experience of knowledge, for Socrates’ uneasiness with his knowledge that he’s the wisest man in Athens was a result of that knowledge not fitting into his framework of knowledge, even though it was justified by coming from Apollo. But it didn’t fit into the framework of other things that he knew, including his knowledge that he knows nothing.

Understanding and a framework of knowledge are crucial pieces of our experience of knowledge that the JTB criteria of knowledge don’t capture. Without a framework, we cannot understand anything, and without understanding, our experience of knowledge feels incomplete. In fact, there is nothing that we know that is completely separate from everything else we know.

So how does a true output from AI fit into these five characteristics of knowledge: a justified true belief the justification of which we understand, and that has a place in our framework of knowledge?

Awkwardly.

Does AI know?

First, the quick and unhelpful answer to the question “Does AI have knowledge” is no, because knowledge is a category of belief, and belief is a conscious state that AI doesn’t currently have — and if the substrate argument is true, it may never have. But this is a slim objection since by the same argument, books don’t contain knowledge because they’re not conscious. So let’s leave it alone since our interest here is the ways in which AI’s computational processes are like and unlike our cognitive processes.

Second, most forms of machine learning don’t tell us how they came up with an output, and often they couldn’t if they wanted to, and if they did we wouldn’t understand them. This can limit the types of justifications machine learning can offer. For example, when ML makes a weather prediction, it’s doing so on the basis of a highly complex model that’s been trained on massive amounts of data about the various factors that affect the weather. By noting during the training process the correlations among the data and the recorded weather in the past, it creates a model that can weigh today’s data and come up with a prediction about tomorrow. But the relations among the data often are so complex that we cannot understand them or see why a particular combination of data leads to a particular prediction. Granted, a great deal of work is being done trying to make ML models more understandable, and regulators may decide to rule out the use of inexplicable ML models at least in some categories. But for now inexplicability is a fact of life in many uses of ML.

Still, there are types of justification an ML model can provide that don’t require understanding how the model came up with an output. For example, we can see how the model fares when given inputs for which we know the correct output; a form of this is crucial to some of the most common ways in which ML models are trained. If a ML model correctly identifies whether a biopsy is cancerous 78% of the time, then in most cases we should have 78% certainty that any particular diagnosis is correct.

Justification based on outcomes is a well-accepted approach outside of the world of ML. If I tell you that I know that one of my frying pans has an area that gets much hotter than the rest of the pan, my experience burning pancakes should count as a justification of that knowledge, even if I don’t know why the temperatures are uneven.

So, yes, the output of ML models can pass the justification test.

But here’s where we can see how radically differently AI “thinks” than we do. Consider chatGPT’s response to my question about the age of the earth. It gave four reasons supporting its view: “…the observation of cosmic microwave background radiation, the expansion of the universe, … the study of the oldest known stars and galaxies, [and] radiometric dating techniques…”. Since we could ask chatGPT about any of these and it would likely give a good answer, it can pass Socrates’ test of not merely parroting justifications.

Or does it? Socrates says that for Theaetetus to know something he has to not only be able to give that belief’s justifications, he has to understand those justifications. Did chatGPT understand them when it produced them in response to my question? Since understanding requires consciousness, no. But does its justifications even come from a coherent framework of knowledge? No, for that is exactly what an LLM does not have.

In fact, the justifications an LLM presents to us for its statements of knowledge have nothing to do with how it actually came up with those statements. An LLM creates responses based on statistics derived from the massive amount of text it’s been trained on. In the training process it determines the likelihood of one word being near another in a sentence. Given how many words there are and how many relationships among those words there can be, this is a truly massive project. Further, its selection of which words to string together is highly contextual. For example, when it writes “the older known stars”, it well might have encountered the word “star” in its training data closer to to “Hollywood” than to “galaxy”, so it has to do some significant contextual analysis to construct a paragraph relevant to the evidence for the earth being round rather than a response involving George Clooney and Brad Pitt.

In other words, an LLM does not have a model of knowledge. It just has a complex model of how we have put words together in the past, in a particular context.

So, the justifications it gives us for for its responses have nothing to do with how the chat AI came up with that response.

Do we want to call its statements knowledge? Very likely yes, at least in some sense. But that doesn’t mean that it’s like what we mean by knowledge. AI doesn’t “think” like us.

The rise of particulars

There are many ways to characterize the root difference between how we think and how machine learning models — not just LLMs — “think”, but there’s one that I think is especially important because it’s so fundamental to how we think about how we think.

Our human frameworks of knowledge are generally composed of general truths. This enables us to view the particular thing we’re wondering about as an instance of something broader than it. We then see it in its connection with other general truths. For example, when I see a particular apple fall from a tree, I might see it as an instance of the general idea of gravity, which is the same general law that I assume when pouring milk into a cup or watching a hawk soar higher.

Our knowledge frameworks are themselves part of a large web of meaning in which each particular shows itself to us one way or another. When I fold a paper airplane, I understand what it is in its connection to paper’s ability to hold its folds, the lift of the air it’s going to fly in, the wind that will shape its course, its heft in our hands, that will pull it down the softness of the grass it’s going to land on, the distance of the clouds it will fly under, the blue of the sky that will blacken as the day ends, and so on in an endless web of connections, some strong and some not even etched yet.

What these types of frameworks have in common is their generality because the point of a framework is to frame particulars, to let us get our bearings in them.

Machine learning does not subsume particulars under generalizations, at least not in the way we do. Take a model that can predict a person’s risk of heart disease. It was likely trained on lots of health records, each of which contains highly particular data. It will look for patterns in that data that let it accurately predict the risk of heart attacks. The model doesn’t know what any of the numbers stand for and it doesn’t know what we humans have already learned about correlations such as the tie between smoking and heart disease. The patterns it finds in the data may surpass human understanding, no matter how well or poorly the model performs.

The difference is between a doctor telling you to lower your cholesterol because we know there is a general link between that and heart disease, which we understand at least somewhat, and on the other hand a machine learning system that has found complex correlations in the particularities of the data and can’t make clear to us how or why those particularities in combination have led it to suggest you avoid fatty foods. Machine learning doesn’t start with generalizations and it doesn’t provide them at the end, at least not usually.

Now, machine learning does in fact generalize, but not in the ways we want. If at the end of the training process the model that emerges only works on the data it’s been trained on, then we say it’s “overfit” and is of no use to us because it has not generalized past its training data. But those generalizations are too complex for us to understand or to apply on our own.

I think we could make a case that generalizations and particulars are undergoing an Hegelian synthesis. We started the new millennium with our framework of generalizations that are the locus of the truth for particulars. Then AI started moving us to see much more particular generalizations that capture patterns so complex that we cannot understand them. In this new synthesis, generalizations don’t reduce particulars to their commonalities, but rather find much more particularized sets of similarities and differences.

To put this more clearly but perhaps too boldly: By understanding particulars as instances of generalizations, we in the west have understood things in terms of their similarities to other things clustered under a generalization. But now with machine learning, we see how much more we can know by understanding particulars in their differences as well, and letting those differences have a strong voice.

This seems to me to be an epochal change.

The world according to AI

If we come to understand ourselves and our world in light of AI, and if we are thus enabled to listen more closely to things in their particularity, how might that change our understanding of how the world works, and what our experience of it is like? What changes if we start to model our self-understanding on our understanding of how ML works?

Any answer of course has to be speculative. But here are some ideas and guesses. None of these is entirely novel; all of them have roots in pre-AI philosophy and in chaos theory and systems thinking — perhaps itself an illustration of the importance of frameworks.

One possible effect: Arguments we’re used to having about whether X is “really” a Y or a Z would begin to look foolish. Insisting that things can only “really” be in one category turns out to diminish the richness of things.

Causality looks like a vast over-simplification in pursuit of a sense of control.

While the eternal and universal laws of science — or laws of science with smaller scope than that — are not contested by this particularized view of the world, the complexity of the particulars in their infinite interrelationships results in outcomes that are far more difficult to predict or understand than the generality these laws promised. Drop a feather from 40,000 meters above the earth and try applying the laws that will determine its descent.

The particularized view of life and experience let us notice the thoroughly accidental nature of life, from what our children will grow up to be, to who will be in the pedestrian crosswalk with us tomorrow morning. We live in the Kingdom of Accidents in which explicable events are the exception.

This would be a pragmatic adaptation of what chaos theory tells us: simple rules can result in unpredictable results when the particulars they are governing are interrelated in complex ways. Simple rules can become non-linear. Small events can unleash cascades of energy and information that cannot be traced back to those events; we call these “butterfly effects.” AI’s ability to make statistical predictions of chaotic systems lets us accept the existence and even prevalence of such systems. This can change everything from how we think about the weather and supply chains to how we view the “path” of our own lives.

If machine learning leads us to frame our experience as far more determined by particulars than by the generalizations that compose frameworks, then we may recognize that risks and possibilities loom larger than we thought, especially if this leads to a world in which our predictive powers, aided by our machines, grow exponentially.

Finally, amidst all the justified concerns about preserving human autonomy and keeping our machines bounded by fairness, there is the hope that AI’s listening to the voice of particulars could lead us to be more attentive to particulars. This could lead to us to paying attention to the unique details in moral situations[3], and the ways in which artworks reveal truths via the details and particulars of how a swirl of paint becomes a polished brass button or precisely the right note transforms a chord into a feeling. It could even lead us to appreciate not just our common humanity, but also the differences that make us humans.

Notes

This article was written for Psiche which will publish an Italian translation of it in December 2024.

Illustration by Midjourney. Prompt: A light-hearted illustration of a stylized human brain and an AI black box with some very faint indications of gears inside it. Bright Colors. No text in the illustration. The image must have no human figures or faces --ar 8:5

Endnotes

[1] See John Searle, Minds, Brains and Science (Cambridge, MA: Harvard University Press, 1984). Douglas Hofstadter is fascinating on this topic as well. For example, see his essay “A conversation with Einstein’s brain” n Douglas R. Hofstadter & Daniel Clement Dennett (eds.), The Mind’s I: Fantasies and Reflections on Self and Soul. Basic Books, pp. 430–457. More recently, the philosopher David Chalmers has discussed the substrate argument skeptically, for example in Reality+ (W.W. Norton, 2022).

[2] The knowledge Socrates talks about when recounting the story of the Oracle is a more religious type of knowledge. So this example works better for us than it would have for a fifth century BCE Athenian.

[3] See the vital work of those proposing an ethics of care, also known as care ethics. This feminist philosophy is transformative of ethics, in my opinion. For example, Nell Noddings, Caring: A Feminine Approach to Ethics and Moral Education, University of California Press; 2nd edition (June 2, 2003). Virginia Held, The Ethics of Care: Personal, Political, and Global, Oxford University Press (May 24, 2007). (I’ve written about AI and the ethics of care, but honestly, read the women who are the founders and developers of this profoundly important ethical philosophy.)

--

--