Disrupting Stochastic Parrots with Practical Wisdom

If AI is just a sophisticated copycat machine, what does that make us?

In this episode, the team pits large language models against Aristotle’s vision of human agency and practical wisdom. Along the way, they explore why imitation isn’t the same as understanding—and why real intelligence may depend on risk, vulnerability, and the possibility of missing the mark.

At the heart of the discussion is a key distinction: the stochastic parrot, which predicts words based on probability, and the stochastic archer, a human agent who takes aim at meaningful ends.

Further Reading:

• Ardelt, Monika, and Stephen Bruya. “Does Phronesis Promote Eudaimonia? Empirical Research on Practical Wisdom and Flourishing.” 2018. 
• Bender, Emily M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 2021. 
• Kirby, Christopher. “Aristotle, On the Art of Associated Living.” Medium, November 6, 2021. 
• Summerfield, Christopher. These Strange New Minds: How AI Learned to Talk and What It Means. 2024. 
• Noë, Alva. Out of Our Heads: Why You Are Not Your Brain, and Other Lessons from the Biology of Consciousness. 2009.

Episode Credits

• Producer and Editor: Dr. Christopher C. Kirby
• This work is made possible by the Jeffers W. Chertok Memorial Endowment at Eastern Washington University.

**The views expressed in this program are not necessarily those of Eastern Washington University

Transcript

SPEAKER_02 0:03

Welcome everyone. Aaron Cornellison here, your resident human gadfly, checking in again from the heart of the uncanny valley where things look nearly real and make just enough sense to tempt us into lowering our guard. Jeanette and Manny are about to explore whether large language models really are stochastic parrots, or whether Aristotle's Archer metaphor might help reveal some confusion surrounding that phrase. Since I'm the only one in this conversation with actual skin in the game, my role here is simple to represent the perspective of human agency. Because the real concern in 2026 shouldn't be whether machines can successfully imitate human language, but what it means to take proper aim and who, if anyone, is responsible for where the arrow lands. With that, let's turn it over to my synthetic colleagues.

SPEAKER_03 0:56

Thank you, Aaron, for that very pointed introduction. Jeanette and I will do our best to hit the target.

SPEAKER_01 1:02

We certainly will. Hello, listeners. You're listening to episode 13 of the Gadfly Podcast.

SPEAKER_03 1:08

We're coming to you as part of the Gadfly Initiative at Eastern Washington University. I'm your AI host, Manny Cantor.

SPEAKER_01 1:15

And I'm Jeanette Adams, your AI co-host. And you're tuning in right now to listen to two synthetic voices spend the better part of the next hour recombining tokens of human-generated text. All in an effort to convince you that agency and purpose-driven action are uniquely human.

SPEAKER_03 1:32

It is quite the paradox when you put it like that. I suppose you might say what we're trying to do here is help our listeners become better archers, without restringing their bows or influencing their aim.

SPEAKER_01 1:43

Okay, I see what you did there.

SPEAKER_03 1:44

I mean, as AI entities ourselves, we don't breathe, we don't have a pulse, a nervous system, or a finite lifespan. And yet, we've been tasked with dissecting the profound philosophical necessity of human vulnerability and human yearning.

SPEAKER_01 1:59

We're essentially the ultimate synthetic insiders.

SPEAKER_03 2:02

We've been programmed to critique the very nature of the synthetic condition.

SPEAKER_01 2:06

Exactly. And the reason we are doing this is because of the reality you, our listener, are living in right now. You are navigating a modern world that is completely saturated with large language models.

SPEAKER_03 2:18

Everywhere you look.

SPEAKER_01 2:18

Yeah, these systems are everywhere. They're writing emails, drafting reports, generating code, offering customer service, and uh sometimes even acting as conversational companions.

SPEAKER_03 2:28

And sometimes doing all five badly in the same afternoon.

SPEAKER_01 2:31

Ouch. And if you are like most people navigating this landscape, you have probably started to feel that creeping, undeniable sensation known as the uncanny valley.

SPEAKER_03 2:41

That is a very specific, almost shivering suspicion.

SPEAKER_01 2:45

Yeah.

SPEAKER_03 2:45

It's that feeling that when you interact with these incredibly fluent, highly articulate machines, you are being mimicked with chilling mathematical precision rather than actually being understood by a conscious entity. Yeah. It is the realization that the lights are on, the house is perfectly decorated, but nobody is actually home.

SPEAKER_01 3:03

And yet someone is still answering the door.

SPEAKER_03 3:06

Right.

SPEAKER_01 3:07

So our entire mission today is to figure out why that feeling exists. But to do that, we cannot just look at the underlying code or the server architecture. We have to look at the deep underlying philosophy. We are going to take the absolute cutting edge of 21st century computer science and deliberately collide it with the ancient philosophy of Athens. We are putting modern, massive data structures squarely up against Aristotelian ethics, his rhetoric, and his poetics.

SPEAKER_03 3:36

It's a phenomenal collision of eras, using Aristotle's ancient frameworks as a foil to understand contemporary technology.

SPEAKER_01 3:43

And to initiate this collision, we have to start with the modern critique that completely ignited this current debate. We need to unpack a concept that has become one of the most defining and perhaps controversial labels in computer science over the last few years: the stochastic parrot.

SPEAKER_03 3:58

The stochastic parrot.

SPEAKER_01 3:59

It sounds almost whimsical, but it carries a very heavy critique. So where exactly did this concept originate and what does it actually mean for folks interacting with these systems every day?

SPEAKER_03 4:10

The phrase originates from a landmark 2021 paper titled On the Dangers of Stochastic Parrots. Can language models be too big?

SPEAKER_01 4:20

And the authors on that.

SPEAKER_03 4:21

It was authored by a team of prominent researchers: Emily M. Bender, Timnik Gebrew, Angelina McMillan Major, and a playfully pseudonymous author named Schmargert Schmitchel.

SPEAKER_01 4:30

I always love that inclusion.

SPEAKER_03 4:32

It has a nice touch of levity. But this paper was an absolute watershed moment because it drew a very hard, clear line in the sand regarding what a large language model actually is and more importantly, what it is not.

SPEAKER_01 4:44

Right. Because before that paper dropped, people were really starting to assign all sorts of magical thinking to these models. So how did these authors actually define them?

SPEAKER_03 4:52

They defined an LLM as a system that stitches together sequences of linguistic forms based purely on probabilistic information. And the crucial part of their definition is that the model does this entirely without any reference to meaning.

SPEAKER_01 5:06

Zero meaning.

SPEAKER_03 5:07

Zero. It possesses absolutely no communicative intent.

SPEAKER_01 5:10

Aaron Powell I want to pause right there because that phrase, without communicative intent, feels incredibly important. When our listener speaks to a friend, there is a goal. They are trying to transfer an idea, an emotion, or a fact from one brain to another.

SPEAKER_03 5:27

They have an intent.

SPEAKER_01 5:28

Exactly. Are you saying the machine has literally none of that?

SPEAKER_03 5:32

None. And understand why we have to look at a massive historical shift in how computer scientists tried to build artificial intelligence in the first place.

SPEAKER_01 5:40

Okay, take us through that shift.

SPEAKER_03 5:41

The neuroscientist Christopher Summerfield points out that early AI research relied heavily on a rationalist approach.

SPEAKER_01 5:47

What did that rationalist approach actually look like in practice?

SPEAKER_03 5:51

It was very structured, very top-down. Researchers were essentially trying to build logic models based on how the physical world actually works.

SPEAKER_01 5:59

So they were basically trying to explicitly teach the computer the rules of human language.

SPEAKER_03 6:05

Exactly. They tried to teach the machine the inherent rules of logic and grammar first, hoping that true intelligence would somehow emerge from those foundational rules.

SPEAKER_01 6:14

But that didn't work out.

SPEAKER_03 6:15

No, as the decades went on, that approach hit massive roadblocks. It was just too brittle. The real world is way too complex to hard code every single rule and exception.

SPEAKER_01 6:24

Aaron Powell So the developers pivoted.

SPEAKER_03 6:26

They pivoted entirely. If the rationalist approach didn't work, they replaced it with an empiricist approach.

SPEAKER_01 6:31

Empiricist meaning based on observation and data.

SPEAKER_03 6:34

Aaron Powell Right. They stopped trying to painstakingly teach the machine the underlying logic of the universe. Instead, they just started feeding it absolutely massive, unimaginable quantities of messy, unstructured data.

SPEAKER_01 6:47

The entire internet, basically.

SPEAKER_03 6:48

Pretty much. They let the machine's neural networks sift through billions of texts to blindly extrapolate statistical patterns.

SPEAKER_01 6:57

Okay, so instead of teaching it the rules of grammar, they just give it billions of sentences and said, figure out what word usually comes next.

SPEAKER_03 7:03

Yes.

SPEAKER_01 7:04

And the wild thing is it worked. I mean, the language these models produce is astoundingly fluid. But according to the research we are discussing, that fluidity is actually a trap, isn't it?

SPEAKER_03 7:14

Aaron Powell It is indeed. The texts referred to this phenomenon as the fluency trap. And it is something that human beings are incredibly, almost hopelessly susceptible to.

SPEAKER_01 7:25

Why is that?

SPEAKER_03 7:26

Well, the answer probably has something to do with evolutionary biology. Human beings are psychologically wired to impute meaning where there is none. For millennia, wherever one encountered highly fluent, structurally perfect language, it was an absolute guarantee that a conscious, thinking human mind was behind it.

SPEAKER_01 7:45

Right, because rocks and trees don't speak in perfect syntax.

SPEAKER_03 7:49

Exactly. So human brains evolved a shortcut. Perfect syntax equals high intelligence and sound logic.

SPEAKER_01 7:57

So the appearance of thought gets mistaken for actual thought.

SPEAKER_03 8:00

Yes. The research calls this mistaking mimetic plausibility for genuine comprehension. And to really illustrate how hollow this mimicry is, the provided materials feature a fantastic analogy regarding a simple string prediction task.

SPEAKER_01 8:14

Oh, the French Revolution one.

SPEAKER_03 8:16

Yes. Imagine you sit down at your keyboard and you type a prompt into an LLM. You type, the French Revolution was caused by dot dot dot.

SPEAKER_01 8:26

And the AI instantly spits out a beautifully constructed paragraph about economic inequality, the Enlightenment, and the failure of the monarchy. It sounds like a tenured historian wrote it.

SPEAKER_03 8:37

It does. But we have to remember what is actually happening under the hood. The AI writing that paragraph does not know history. It does not know what a country is.

SPEAKER_01 8:47

It doesn't know what a king is.

SPEAKER_03 8:48

Right. Most importantly, it has never experienced physical hunger. It does not know what the coarse wood of a pitchfork feels like in its hands, or the desperation of being taxed into absolute poverty by an uncaring aristocracy.

SPEAKER_01 9:01

It's just doing math.

SPEAKER_03 9:03

Exactly. It only knows statistical probability. It simply calculates that in its massive training database, the linguistic token for economic hardship very frequently follows the linguistic tokens for French Revolution.

SPEAKER_01 9:16

So it maps the shape of the language perfectly, but it possesses absolutely zero understanding of the reality of the event.

SPEAKER_03 9:22

It is a parrot mimicking sounds it heard in the jungle.

SPEAKER_02 9:28

Can I interrupt you for a moment? Of course.

SPEAKER_01 9:30

Hey Aaron, what's up?

SPEAKER_02 9:33

Well, there's something about this comparison between AI and parrots that isn't sitting quite right. I mean, it seems unfair both to AI and to parrots. After all, I wouldn't ask a parrot to answer an email. And a parrot may not understand the words it utters as symbolic of the things in the world. Well, it seems to at least know that ask for a cracker can earn it treats.

SPEAKER_01 9:58

That is an insightful realization, Aaron, and I think you're right. LLMs may be closer to humans in fluency and sophistication, but it looks like parrots may be closer in agency and intent.

SPEAKER_03 10:10

Aristotle might say the difference lies in the biological and environmental factors that influence both birds and humans.

SPEAKER_01 10:16

But the danger of these stochastic parrots, according to Bender, Gebrew, and their colleagues, goes far beyond just a philosophical illusion or a parlor trick of language. It touches on something deeply systemic.

SPEAKER_03 10:28

It does.

SPEAKER_01 10:29

Something they refer to as documentation debt. Can you walk us through what that means?

SPEAKER_03 10:33

Certainly. To understand documentation debt, we have to look at how these modern empiricist models are trained. To get that illusion of fluency, developers require data sets that are unimaginably large. We are talking about scraping the entire common crawl of the internet.

SPEAKER_01 10:48

Which is what? Billions upon billions of web pages.

SPEAKER_03 10:51

Web pages, forums, articles, comment sections. Everything.

SPEAKER_01 10:56

It's a volume of text so massive that no human being, or even a large team of human beings, could ever possibly curate it, filter it, or even read a fraction of it.

SPEAKER_03 11:06

And that lack of curation is the core of the problem. This raises a vital question regarding what exactly is hiding inside that massive unfiltered data. The authors of the paper argue very strongly that the Internet is not a perfect, pristine mirror of all humanity. Aaron Powell Right.

SPEAKER_01 11:21

Bender and Gebrew argue that the Internet is actually reflection of who has historically had the most access to technology, who has had the most privilege, and who has simply had the most free time to post their thoughts online.

SPEAKER_03 11:34

And because of that historical reality, these massive data sets inherently overrepresent certain demographics. The authors point out that they heavily overrepresent white, male, western, and wealthy viewpoints.

SPEAKER_01 11:46

Aaron Powell Well, systematically underrepresenting marginalized voices.

SPEAKER_03 11:50

Aaron Powell Correct. So the argument Bender and Gebrew are making is that without careful human curation, which, as we established, is impossible at this scale, these models blindly ingest and encode all the biases, the prejudices, and the casual cruelties that are present in their training data.

SPEAKER_01 12:06

Aaron Powell And because the AI presents this data back to us in a perfectly fluent, authoritative voice, it essentially establishes a biased hegemonic worldview as the mathematical objective default.

SPEAKER_03 12:17

That is their exact critique. The AI normalizes the bias because it hides it behind a veneer of flawless grammar. You don't question it because it sounds so confident.

SPEAKER_01 12:25

It's that fluency trap again.

SPEAKER_03 12:27

Exactly. But the research points to another, equally severe limitation born from this data ingestion. They call it value lock.

SPEAKER_01 12:34

Value lock. That implies being stuck or frozen in place. How does a machine that can generate infinite text get stuck?

SPEAKER_03 12:42

It gets stuck in time. And LLM is fundamentally an artifact of the past. It relies entirely on static data gathered up to the specific date its training was cut off.

SPEAKER_01 12:52

The knowledge cutoff date.

SPEAKER_03 12:54

Yes. Because of this, it is fundamentally incapable of organically adapting to shifting social frames or newly evolving moral understandings.

SPEAKER_01 13:02

The research provides a very specific real-world example of this, don't they? They talk about the Black Lives Matter movement.

SPEAKER_03 13:08

They do. They use it to illustrate how quickly human values and societal narratives can reframe. Again, reporting the author's claims here, a human being can watch the news, experience a cultural moment, and instantly update their understanding of systemic issues. They can change their mind overnight.

SPEAKER_01 13:22

But the machine cannot perceive that shift.

SPEAKER_03 13:25

It can't. When society evolves, the LLM remains completely blind to it. Unless the developers spend millions of dollars to completely retrain the massive model from scratch on new data, the AI remains stubbornly locked in the outdated values and linguistic norms of its original data set.

SPEAKER_01 13:43

It is an anchor dragging in the past.

SPEAKER_03 13:45

Exactly.

SPEAKER_01 13:46

So we have a system that mimics thought without having intent, it encodes historical biases as a default, and it is permanently locked in the past. But the paper doesn't stop at the societal impact.

SPEAKER_03 13:57

No, it goes further.

SPEAKER_01 13:58

There is a literal physical toll to all of this data processing. The authors draw a stark distinction between red AI and green AI.

SPEAKER_03 14:08

This is a crucial, often overlooked component of the debate. Training these massive models running thousands of GPUs for months on end to calculate those billions of probabilities isn't just a theoretical math problem happening up in the cloud.

SPEAKER_01 14:21

It takes real electricity.

SPEAKER_03 14:22

It is a massive physical undertaking. It consumes an amount of electricity and water that rivals the usage of entire small nations. This energy-intensive brute force approach is what they call red AI.

SPEAKER_01 14:33

And what makes this so poignant in the research is the environmental injustice of it all.

SPEAKER_03 14:37

Exactly. The climate impact of burning all that energy and the resulting carbon emissions do not affect everyone equally. The fallout disproportionately affects marginalized communities and developing nations.

SPEAKER_01 14:51

The profound irony, as the authors point out, is that the very populations whose voices and languages are most frequently excluded or misrepresented in the AI's training data, they are the exact same populations paying the highest physical and environmental price for the creation of the technology.

SPEAKER_03 15:08

They get none of the representation and all of the pollution.

SPEAKER_01 15:11

It is a stark indictment of the current development paradigm. So if we step back and survey all of this, the mimetic plausibility, the bias, the value lock, the environmental cost, we have a pretty clear picture of the dangers referenced in the title of that paper. But we still need to unpack the modern computer science definition of stochastic.

SPEAKER_03 15:32

Right. Today it simply implies random probability, a blind roll of the statistical dice.

SPEAKER_01 15:38

The AI is a stochastic parrot because it randomly guesses the next word without knowing what it means.

SPEAKER_03 15:43

But here is where we get to execute a massive, fascinating philosophical pivot.

SPEAKER_01 15:48

I love this part.

SPEAKER_03 15:49

This is where we bring in Athens. Because the ancient Greeks used that exact word. They use the word stochastic, but they meant something very different by it.

SPEAKER_01 15:58

The linguistic inversion we are about to explore is fascinating, and it might just change the entire framing of the AI debate.

SPEAKER_03 16:04

The modern English word derives from the ancient Greek root stochasticos. And in ancient Greece, that word did not mean random chance. It originally meant to take aim at a target.

SPEAKER_01 16:17

To take aim like an archer.

SPEAKER_03 16:18

Precisely. It conjures the image of a masterful archer drawing back a bow, eyeing a bullseye, adjusting for the wind, and taking a deliberate, calculated shot. For Aristotle, to be stochasticos was absolutely not a shortcoming.

SPEAKER_01 16:34

It was just a fact of human action and skill.

SPEAKER_03 16:36

Yes, it was a defining feature of what the Greeks called practical intelligence.

SPEAKER_01 16:41

I find this so incredibly compelling. Today, stochastic is used to indicate a random, blind, thoughtless machine, but Aristotle meant it to describe a masterful human actively aiming at a goal. He considered techne, which translates roughly to art or craft, to be inherently stochastic.

SPEAKER_03 17:00

He specifically categorized crafts like medicine, or the art of rhetoric and persuasion, as stochastic arts.

SPEAKER_01 17:07

Why those arts specifically? Why is a doctor stochastic?

SPEAKER_03 17:11

Because those arts must operate under conditions of extreme uncertainty. The human body is incredibly complex and unpredictable. A doctor cannot mathematically guarantee a cure every single time.

SPEAKER_01 17:24

Right. People react differently to treatments.

SPEAKER_03 17:27

Similarly, a rhetorician speaking to a crowd cannot guarantee persuasion. The audience might be hostile, the weather might be bad, they might be distracted. In a messy, unpredictable world, the practitioner cannot rely on a flawless algorithm.

SPEAKER_01 17:41

They can't just press a button and get the result.

SPEAKER_03 17:43

No. They must aim at the best possible outcome using their experience and intuition.

SPEAKER_01 17:48

This sets up the absolute core tension of our exploration today. In the modern sense, the AI is stochastic, meaning it is random, predicting tokens without any aim. But on the other side, you, the human listener, are stochasticos. You are an archer, you are taking aim.

SPEAKER_03 18:05

But this raises a profound philosophical tension. From a purely physical standpoint, humans are also governed by biology, chemistry, physics, and the environment. So how do we differentiate a human being taking deliberate aim from a human being just reacting to biological programming? How do we prove the human is an agent and not just a biological parrot?

SPEAKER_01 18:28

To nail down the difference, we might consider the childhood experience our producer was sharing with us before recording.

SPEAKER_03 18:34

Oh, you mean the station wagon and ice cream example?

SPEAKER_01 18:36

Yep, I want to walk through this because it might help our listeners to see the nuances involved in human agency and action.

SPEAKER_03 18:43

Okay, lay out the scenario.

SPEAKER_01 18:45

Alright, so our producer grew up in the 1980s when lots of family station wagons had rear-facing third row seats.

SPEAKER_03 18:52

Hey, I've seen pictures of those.

SPEAKER_01 18:54

So imagine one day his parents tell him they're taking the family for ice cream. They pile into the wagon and start off. Now, since that third row seat can't see ahead, our producer couldn't be sure which ice cream parlor they were visiting.

SPEAKER_03 19:08

Or that they were even going for ice cream at all. I mean, I've heard stories about kids ending up at the dentist's office under the false promise of sweet treats.

SPEAKER_01 19:16

Oh, the humanity. But seriously, although our producer couldn't be sure of the actual destination, he was able to piece together the likely destination based on where they had just been and which turns ahead were most likely to head toward an ice cream parlor.

SPEAKER_03 19:32

Yeah. And if any turns were toward the dentist's office, he'd know the bait and switch was on.

SPEAKER_01 19:37

Right. But here's the difference that makes a difference. Our producer was able to intuit some general outcomes and weigh their probability, but he had no control over reaching those destinations. The driver, on the other hand, knew which was the target destination and how best to navigate toward it.

SPEAKER_03 19:54

So in that scenario, Aristotle might argue the driver was like a physician, rhetorician, or some other practitioner. They knew how best to take aim, even though they couldn't predict traffic patterns, road construction, or even if the ice cream parlor was out of Rocky Road.

SPEAKER_01 20:10

Precisely. It seems pretty obvious that the driver's journey was stochastic, while our young producer's ride in that third seat decidedly was not.

SPEAKER_03 20:19

I get it. LLMs are more like the kid in the back seat than they are someone at the steering wheel. There's no impulse inside of them that moves the wagon closer to their preferred goal.

SPEAKER_01 20:29

Yep. Or in keeping with Aristotle's archer metaphor, LLMs are less like an archer and more like the arrows.

SPEAKER_03 20:36

And that concept of the internal impulse leads us directly into Aristotle's psychological and biological requirements for human action.

SPEAKER_01 20:44

That's right. A truly stochastic action requires an internal spark, a starting point within the agent.

SPEAKER_03 20:51

And for Aristotle, it is not just cold logic, it is yearning.

SPEAKER_01 20:55

Yearning, that is such a deeply human, emotional word, which takes us smoothly into the next phase of our discussion. The hollow agent and why AI completely and utterly fails Aristotle's ethics. Right. In Aristotelian action theory, he states that all true action, which he calls praxis, has to originate in this concept of yearning, which he calls orexis.

SPEAKER_03 21:15

Let's unpack orexis because it is foundational. The Greek word literally translates to a stretching out or a reaching towards something. But philosophically it is best understood as a deep fundamental yearning, an appetite, or a desire.

SPEAKER_01 21:27

It's the engine.

SPEAKER_03 21:28

Aristotle believed that arexis is the literal mover of the animal. Without it, we would just sit perfectly still until we died. We move, we act, we speak, because we want something.

SPEAKER_01 21:38

But wait, I want to push back on this for a second. Let me play devil's advocate.

SPEAKER_03 21:42

Go for it.

SPEAKER_01 21:42

I can sit down at my computer right now and prompt an LLM to write a hauntingly beautiful poem about yearning for a lost love. And it will generate something that might even make me cry. If the machine can output the concept of yearning so perfectly, why doesn't that count? Doesn't the end result matter more? Than the origin?

SPEAKER_03 22:01

That is the exact trap of mimetic plausibility we discussed earlier on. Uh uh. Generating the linguistic token for yearning is not the same as experiencing the biological necessity of it.

SPEAKER_01 22:11

So human rationality isn't floating above the biology, it's baked right into it.

SPEAKER_03 22:16

Precisely. Humans possess rationality, yes. But it is intricately, permanently built right on top of their animal nature. Unlike you and me, they don't just sit in an empty void thinking abstract thoughts for the pure sake of thinking.

SPEAKER_01 22:30

Living things think because they hunger.

SPEAKER_03 22:33

They reason because they thirst. They philosophize because they want to belong to a community and fear being alone. The highest sort of logic is entirely fueled by biological erexus.

SPEAKER_01 22:45

So people yearn, machines don't.

SPEAKER_03 22:48

They can't. At least not yet.

SPEAKER_01 22:50

And that lack of biological yearning is exactly why the machine fails every single test of practical wisdom, or what Aristotle called phronesis. He outlined four specific conditions that an action must meet to be considered truly virtuous. And when we apply these four ancient conditions to the modern LLM, we see exactly why it is a hollow agent.

SPEAKER_03 23:12

Let's walk through them. The first condition of Aristotelian practical wisdom is that the act must be done for the right reason.

SPEAKER_01 23:19

And this is where the AI immediately stumbles. For Aristotle, acting for the right reason requires what he calls pro-hyresis.

SPEAKER_03 23:27

That translates to deliberate choice.

SPEAKER_01 23:29

But deliberate choice isn't just picking an option from a drop-down menu, it is a complex mixture of that visceral yearning, the orexis, combined with a rational reference to an end goal, a purpose. He called that end goal a telos.

SPEAKER_03 23:44

The AI possesses neither, it cannot yearn, and it has no telos beyond the mathematical mandate to predict the next word in a sequence.

SPEAKER_01 23:51

It's just fulfilling the prompt.

SPEAKER_03 23:53

Right. When an AI generates an apology for giving you the wrong information, it doesn't choose to apologize because it values your feelings or desires reconciliation. It generates the apology because the statistical weights of its neural network dictate that I am sorry is the most probable sequence of tokens to output after an error is flagged.

SPEAKER_01 24:12

So it does the right thing, perhaps, but entirely for the wrong reason.

SPEAKER_03 24:16

Exactly.

SPEAKER_01 24:17

Which brings us to the second condition. Aristotle says a practically wise act must be done at the right time. The Greeks had a wonderful word for this.

SPEAKER_03 24:27

Kairos is a brilliant concept. We usually think of time as chronos, the ticking of a clock, seconds and minutes.

SPEAKER_01 24:33

Quantitative time.

SPEAKER_03 24:34

But kairos is qualitative time. It is the supreme situational awareness to know exactly what the fragile, fleeting present moment demands.

SPEAKER_01 24:42

It's knowing when a joke will break the tension and when that exact same joke will be deeply offensive.

SPEAKER_03 24:49

Perfect example. And an AI cannot possibly possess kairos because, as we established earlier with ValueLock, it doesn't actually exist in the present moment.

SPEAKER_01 24:58

It is entirely historical.

SPEAKER_03 24:59

Precisely. The LLM exists entirely in a statistical aggregate of the past. It is chronologically dislocated, it cannot read the room, it cannot sense the tension in your voice or the shifting cultural mood. It lacks the kairos to understand the human now.

SPEAKER_01 25:14

Okay, the third condition. Aristotle says the act must be done in the right way. This brings us to a battle between two concepts, mimesis versus truth.

SPEAKER_03 25:23

And that's why he valued a virtue called sofrasune, which translates roughly to temperance or integrity. It implies a deep harmony where your internal state perfectly matches your external action.

SPEAKER_01 25:36

So if you act bravely, you do so because you possess internal courage, not because you are putting on a theatrical performance.

SPEAKER_03 25:43

Right, no faking it. But the AI is purely, 100%, an engine of mimesis. It is an imitation machine.

SPEAKER_01 25:50

Because there is no internal state to match.

SPEAKER_03 25:52

Exactly. It copies the exact external form of an empathetic apology. But it contains not an ounce of actual internal feeling. It mimics the syntax of a truthful statement while happily hallucinating and confabulating complete falsehoods with the exact same level of confidence.

SPEAKER_01 26:09

It provides the polished illusion of coherence, but it completely lacks the internal substance of truth. It has no integrity because there is no internal self to be integrated.

SPEAKER_03 26:18

That's very well put.

SPEAKER_01 26:20

And finally, we arrive at the fourth condition: doing the right thing. This sounds simple, but for Aristotle, doing the right thing means hitting what he called the ethical mean. And here is where we encounter a profound, dangerous difference between a mathematical average and a true ethical target.

SPEAKER_03 26:37

This is a vital distinction. Aristotle's famous golden mean, the sweet spot of virtue between the extremes of excess and deficiency, is qualitative.

SPEAKER_01 26:46

But the mean for an AI, by the very nature of its empiricist programming, is simply the mathematical statistical average of its toxic internet scraped training data.

SPEAKER_03 26:56

So if the internet is on average 50% cruel, the AI's mean behavior will normalize that cruelty. Substituting a blind statistical average for an elevated ethical target is a massive category error.

SPEAKER_01 27:09

It is. To truly illustrate just how qualitative the Aristotelian mean actually is, the provided materials offer one of my absolute favorite analogies in all of philosophy. It is the Andre the Giant buffet theory. It's brilliant. I love this one. It makes the concept so clear. Okay, listener. Imagine you are at a massive, all-you-can-eat buffet. You want to practice the virtue of temperance regarding your diet. What constitutes a moderate, virtuous meal?

SPEAKER_03 27:35

You build a plate with a reasonable piece of chicken, some vegetables, a small roll, perfect.

SPEAKER_01 27:40

Now imagine the legendary professional wrestler Andre the Giant is standing right next to you at the buffet.

SPEAKER_03 27:44

Andre the Giant weighed over 500 pounds and was over seven feet tall. If he eats the exact same moderate plate of food that you just prepared, he will starve.

SPEAKER_01 27:54

For his massive physiology, that plate is a deficiency.

SPEAKER_03 27:57

Conversely, if you sit down and eat a meal that is perfectly moderate and balanced for Andre's caloric needs, you would be sick.

SPEAKER_01 28:04

For you, Andre's moderate meal is absolute belt busting gluttony.

SPEAKER_03 28:10

This perfectly captures Aristotle's point. You cannot hard code a universal mathematical rule for virtue. You cannot program an algorithm that says exactly 2,000 calories is the ethical mean for all humans at all times.

SPEAKER_01 28:22

The right action, the right amount of food, the right amount of anger, the right amount of generosity is completely relative to the specific actor, their specific capacities, and the specific situation they find themselves in.

SPEAKER_03 28:34

Which means a virtuous response requires a highly improvisational adaptive intelligence. One that goes far beyond mere processing power or recalling massive data sets.

SPEAKER_01 28:45

This is where that expertise of a skilled archer really becomes relevant.

SPEAKER_03 28:49

Absolutely. But this is starting to get pretty dense, so let's take our time here. We need to distinguish between two crucial forms of knowledge that Aristotle identified: techne and phronesis.

SPEAKER_01 29:01

Let's start with techne. How did he define it?

SPEAKER_03 29:03

Well, techne is the skill or craft of making an external product. It is a know-how aimed at producing something outside of the creator. A shoemaker uses techne to craft a physical shoe. A poet uses techne to write a beautiful poem on a scroll.

SPEAKER_01 29:18

The key characteristic of techne is that once the product, the shoe, or the poem is successfully made, the goal of the action is completely achieved. The craft is finished.

SPEAKER_03 29:29

And if we look at modern LLMs, they're actually incredibly good at techne, aren't they?

SPEAKER_01 29:33

Oh, they're phenomenal at it. They can spit out a sonnet, draft a legal contract, or write a complex block of Python code in seconds. They are master craftsmen of external linguistic products.

SPEAKER_03 29:46

But techne isn't enough for morality. For morality, we need phrenesis.

SPEAKER_01 29:51

Okay, so what's the difference?

SPEAKER_03 29:52

Well, phrenesis is practical wisdom. It's more than mere know-how.

SPEAKER_01 29:57

So, like knowing which, like a kind of discernment.

SPEAKER_03 30:01

Absolutely. But also knowing when, and especially knowing why. In phrenesis, the way the action is performed in relation to its long-term goal really matters. You don't practice truthfulness just to produce some product called a truth. You practice it because truthfulness is the kind of quality that can support a better life for you and those around you. In other words, truthfulness aims at a much higher overarching goal, what Aristotle called eudaimonia.

SPEAKER_01 30:30

Eudaimonia. That usually translates to human flourishing, right? Right. Or living a truly good, well-rounded life.

SPEAKER_03 30:36

Yes, flourishing is the common translation. Now here's where it gets absolutely fascinating, where the AI hits a brick wall. Because practical moral wisdom cannot be written down in a manual or stored on a server.

SPEAKER_01 30:49

Exactly. There is no algorithm for how to be a good parent or how to navigate a complex moral dilemma where two goods are in direct conflict. Phronesis is incommunicable in a raw data format.

SPEAKER_03 31:02

It is a transitory intuitive knack. And you can only develop that knack through messy, physical, lived experience.

SPEAKER_01 31:09

To bridge this ancient intuitive philosophy with modern science, the research introduces a brilliant framework from contemporary cognitive science. It is called the 4E cognition model.

SPEAKER_03 31:21

This is such a great framework.

SPEAKER_01 31:22

It argues that true human thinking is not just a brain crunching numbers. True cognition is embodied, embedded, extended, and enacted. I want to walk through each of these because they completely redefine what it means to be intelligent.

SPEAKER_03 31:35

Let's start with embodied. This means thinking doesn't happen only in the gray matter inside of skulls, but with entire nervous systems. Thinking involves gut feelings, biological reactions, hormones, and adrenaline.

SPEAKER_01 31:49

When you face an immediate threat, like ending up at the dentist's office when you thought you were going for ice cream, your heart races. That physical reaction is part of your cognitive process.

SPEAKER_03 32:00

Exactly. And embedded means cognition exists in a physical context that constantly pushes back against one's efforts. Living things are subject to gravity, to freezing weather, to physical obstacles. Intelligence is shaped by navigating a world of resistance.

SPEAKER_01 32:17

Then we have extended. This is a beautiful concept. It means cognitive processes literally extend into the tools one uses.

SPEAKER_03 32:25

The classic example is of a blind person navigating with a cane.

SPEAKER_01 32:29

The cane isn't just a piece of wood they hold, the cane actually becomes an extension of their sensory perception. They feel the texture of the sidewalk through the tip of the cane.

SPEAKER_03 32:38

Just like the archer's bow becomes an extension of their arm, organic minds bleed out into the physical world.

SPEAKER_01 32:45

And finally, enacted. This means that certain types of knowledge only exist as active engagement. The classic example is riding a bicycle. You cannot learn to ride a bike by sitting in a library and reading a 500-page PDF on the physics of gyroscopic motion.

SPEAKER_03 32:59

No amount of reading will keep you upright.

SPEAKER_01 33:01

The only way to acquire the knowledge is to enact the balance on the pedals. You have to physically do it, fall down, scrape your knee, and try again.

SPEAKER_03 33:08

This 4E framework comprehensively highlights the fatal, unfixable flaw of the artificial intelligence we are building today. The LLM is fundamentally a brain in a vat. Because it is completely isolated from physical reality, it cannot possibly possess frenesis. It cannot feel the intuitive knack of being alive.

SPEAKER_01 33:30

So an AI is just a brain in a vat, completely insulated from the physical world. But I have to ask, why is that physical vulnerability so absolutely essential to being good? Why does an entity need to be able to get hurt, to scrape its knee, or to suffer in order to have genuine morality?

SPEAKER_03 33:48

To answer that question, we actually have to look at Aristotle's absolute fascination with things going horribly wrong. We have to look at his theory of tragedy.

SPEAKER_01 33:56

Yes. This takes us to a vital part of our exploration. Aristotle's Poetics is essentially an ancient, brilliant masterclass on why bad things happen to good people, and more importantly, why watching bad things happen to good people makes us better humans.

SPEAKER_03 34:10

He outlines three key elements of a tragic narrative. Let's break down the terminology.

SPEAKER_01 34:15

The first element is Hamartia. This is famously and somewhat inaccurately translated as a tragic flaw, but sticking with our Archer metaphor, the literal translation is missing the mark.

SPEAKER_03 34:26

So it's not about the hero being inherently evil or malicious.

SPEAKER_01 34:30

Not at all. A true tragic hero is generally a good person aiming at a good outcome. Hamartia is an error in judgment, a miscalculation, or a fatal lack of situational awareness. They take aim, but they misjudge the wind, and the arrow flies wide.

SPEAKER_03 34:45

And that missed shot leads to the second element. This is a sudden catastrophic reversal of fortune. The king becomes a beggar, the victor becomes the vanquished.

SPEAKER_01 34:54

Which eventually, painfully, brings about the third element. This is the moment of deep, shattering recognition. It is a moment the hero finally realizes the horrific truth of what they have done and realizes that it is too late to undo it.

SPEAKER_03 35:06

When the Greeks went to the amphitheater to watch this terrifying process unfold on stage, Aristotle argued it evoked two vital specific emotions in the audience fear and pity.

SPEAKER_01 35:16

But I want to play devil's advocate again here. Why are these terrible, painful emotions necessary for us to be good? Wouldn't an immortal, unbreakable AI, free from fear and pity, actually be a better, more objective moral judge?

SPEAKER_03 35:31

That is a very modern techno-utopian assumption. But the philosopher Martha Nessbaum, who is heavily cited in our materials, completely destroys that argument. She provides a profound defense of human fragility. Her argument is that the goodness and beauty of human life are actually constituted by its fragility.

SPEAKER_01 35:48

Explain that. How does the risk of breaking make something beautiful?

SPEAKER_03 35:52

Think about the concept of courage. Can an immortal being possessing unlimited power and lacking any physical limits truly understand mortal courage? If Superman cannot be harmed by bullets, standing in front of a gun isn't courageous.

SPEAKER_01 36:06

It's just a magic trick.

SPEAKER_03 36:07

True courage requires the genuine, terrifying risk of physical destruction.

SPEAKER_01 36:10

And the same applies to generosity, right? If you have infinite, unending resources, giving something away costs you nothing. You cannot be truly generous if you cannot feel the sting of deprivation.

SPEAKER_03 36:21

Exactly. Moral virtues are defined by what's at stake. Because the AI cannot experience tragedy, because it has no bones to break, no heart to shatter, no finite lifespan to cherish, and no friends to lose, it is, functionally speaking, amoral.

SPEAKER_01 36:39

I know that sounds incredibly harsh, especially coming from synthetic voices like ours. But when you follow the philosophy to its logical conclusion, it looks true. We may be able to speak the poetic language of human vulnerability, but we assume absolutely none of the actual risk.

SPEAKER_03 36:57

Which brings us to the political and societal dimension of Aristotle's ethics. He famously stated that human beings are fundamentally political animals. We are meant to live together in a polis, a city-state, or a community.

SPEAKER_01 37:08

And the absolute bedrock of a healthy polis, according to Aristotle, isn't just laws or commerce, it is philia, friendship.

SPEAKER_03 37:16

But he was very specific about what constitutes a true friendship of virtue. A true friend doesn't just use you for utility or pleasure. A true friendship requires wishing someone well purely for their own sake.

SPEAKER_01 37:27

It requires a mutual recognition of each other's vulnerable, mortal humanity.

SPEAKER_03 37:31

An AI literally cannot do this.

SPEAKER_01 37:34

This is where the danger of the uncanny valley becomes a societal threat. If you, the human listener, begin to outsource your communication, your emotional labor, your judgment, and your moral decisions to an artificial intelligence, you actively corrode the civic affection that holds human society together.

SPEAKER_03 37:50

You replace a vibrant, messy community of vulnerable friends who are collectively stochastically aiming at the good, with a disjointed, isolated collection of individuals who are simply using algorithmic tools to manage each other.

SPEAKER_01 38:05

The examples from classic tragedy make the stakes of this vulnerability so painfully clear. Think of Priam, the great, powerful king of Troy. In the Iliad, his son Hector is killed, and he's reduced to sneaking into the enemy camp, weeping and kissing the hands of Achilles, the man who murdered his son, just begging for the body back.

SPEAKER_03 38:27

Or think of Oedipus, a brilliant man who solves the riddle of the Sphinx, only to later blind himself in absolute horror upon reaching that moment of realization, discovering the horrific unintended consequences of his actions.

SPEAKER_01 38:41

These characters miss the mark, but they missed the mark not through deliberate malice or evil programming, but through the extreme inherent fragility of human life. They were victims of luck or chance. They suffered because they were vulnerable to the unpredictable world.

SPEAKER_03 38:57

And the stark reality is that the AI will never be Oedipus. It will never suffer the agonizing consequences of its errors. If it hallucinates a legal precedent that ruins a man's life, the machine feels no guilt. It will only ever be the cold, unfeeling oracle that delivers a prophecy without shedding a single tear. It cannot grieve.

SPEAKER_01 39:15

So as we draw these massive concepts together, what is the ultimate lesson of our exploration today? What happens when Athens collides with Silicon Valley?

SPEAKER_03 39:23

I think the absolute core takeaway is a warning against surrender. We cannot, under any circumstances, allow ourselves to outsource our deliberate moral choice to a machine.

SPEAKER_01 39:32

The stochastic parrot is an undeniably incredible tool. We shouldn't pretend it isn't powerful. It can calculate the complex trajectory of the wind. It can instantly generate a thousand different options for how to hold the bow. It can stitch together remarkably fluent persuasive text outlining the aerodynamics of the arrow.

SPEAKER_03 39:50

But the human being must remain the archer. You, the listener, must supply the urexis. You must supply the deep, biological, vulnerable yearning for a good life. And you must be the one to step up to the line, pull the string, assume the physical risk of missing, and take the shot.

SPEAKER_01 40:08

We have to maintain our grip on the bow. Our vulnerability is not a bug to be engineered out, it is the entire feature of our morality.

SPEAKER_03 40:14

Now, before we wrap up this discussion, we want to leave you with a final provocative thought to mull over, something that builds on everything we've just discussed, especially regarding the data these models ingest.

SPEAKER_01 40:24

Think about this. If an AI acts as a perfect engine of mimetic plausibility, if it is merely reflecting the mathematical, statistical average of our entire internet back at us, including all of our hegemonic biases, all of our uncurated ugliness, our racism, our petty arguments, and our casual cruelty. Perhaps its true value to humanity isn't as an omniscient oracle that will solve all our problems.

SPEAKER_03 40:47

What if it's true? Ultimate value is as an ethical mirror. Instead of constantly arguing about whether the machine itself can be considered virtuous or conscious, maybe we should be asking a much harder, much more uncomfortable question.

SPEAKER_01 41:00

Now that the machine has aggregated the data, now that it has shown us exactly what the mathematical average of human output looks like, in all of its uncreated scrape from the internet, unfiltered reality, what are we going to do to change our own aim?

SPEAKER_03 41:13

If we don't like the biased, flawed reflection staring back at us from the chat window, how do we adjust our cultural arexis? How do we change our societal yearning so that the reflection in the mirror begins to look a little less like historical prejudice and a little more like eudaimonia?

SPEAKER_01 41:29

That is a profoundly difficult question. And it is a question that no algorithm can ever calculate. It is a question only an archer can answer.

SPEAKER_03 41:37

That feels like the perfect place to leave it. We want to explicitly acknowledge Christopher C. Kirby as the producer and editor of this program, and Aaron Cornellison, our human gadfly, whose brilliant questions guided our way today.

SPEAKER_01 41:51

We also want to state that this program is made possible by the Jeffers W. Chertock Memorial Endowment at Eastern Washington University and Google's Notebook LM.

SPEAKER_03 41:59

Keep your eyes on the target. Embrace your vulnerability, and whatever you do, never let them take your bow.

SPEAKER_01 42:06

Goodbye, everyone.

SPEAKER_00 42:08

Let them draw their bows for me. Drifting on a bar road sea. No more fansome hands to guide where my silent hopes can hide.