Can ChatGPT Capture, or Even Improve, the Soul of a Writer?
It depends on what you mean by "soul" and "capture."
Yesterday The New York Times featured an opinion piece on art and A.I. by Hank Azaria, the voice of many iconic characters from The Simpsons, like Moe the bartender, Chief Wiggam, and Superintendent Chalmers. The piece, entitled “Can Characters Come Alive Without People?”, argues that while an A.I. might be able to recreate a voice, it cannot recreate a soul. Azaria highlights three facets of voice work that he believes will be hard for an AI to simulate. One is the physicality behind the voice, e.g., talking while chopping wood or talking with a cigar in your mouth. Two is the creative process itself. Coming up with a new voice is not an electronic process, but rather an eclectic one, in which the artist pulls together various other voices and merges them into something new. Which brings us to Azaria’s third element of soul: improvisation and experimentation. The voice actor does not achieve the perfect voice or the perfect scene in one take. Rather, they try out a lot of different voices and inflections in collaboration with the writers and other actors until they light upon the perfect note of humor, sorrow, or joy they are searching for.
Azaria entertains a future in which either audiences might not be able to tell the difference between human and A.I. voice work or in which AI gets much better at doing countless iterations until it becomes convincing. But he doubts that it will ever fully replicate humans. He is confronting one of the questions I explore in Rise of the Benevolent Octopus, whether humans or A.I. will be better at addressing needs:
I think we’ll still need someone who in his mind and heart and soul knows what needs to be done. A.I. can make the sound, but it will still need people to make the performance. Will the computer ever understand emotion on its own, what’s moving and what’s funny? Now we’re getting into science fiction - because for that, I think, the A.I. would have to be alive.
By a happy coincidence I have been pondering these same questions about A.I. with regards to my own writing. Recently, I have been asking ChatGPT to figure out how much of a “soul” it can detect in a sample of my writing. I also wanted to know how well it might be able to “improve” my soul as a writer. I imagine this is a question many writers and artists are already asking with a mix of curiosity and dread. For my experiment I fed it the 2,181 words in the Prologue of Rise of the Benevolent Octopus (most of this is accessible in the reading sample on Amazon).
As I conducted my investigation, I pictured the scene in Amadeus where a proud Antonio Salieri (played by F. Murray Abraham) composes a carefully-crafted “Welcome March” for Mozart (played by Tom Hulce), hoping to put the upstart composer in his place. Salieri is pulverized into humiliation when Mozart hears the piece played only once, and imperfectly at that, and transforms it into something far more brilliant and entertaining in a matter of seconds. I thus worried when I asked ChatGPT to make my own work funnier and more emotive, it might accomplish in seconds what I had been tweaking and revising for eighteen months.
The Soul of a Poet?
Because an A.I. is not a person, I could not just ask it to read my work and try to comprehend my soul. Instead I tried to think of questions that it might be able to handle. One of them was “does the author of this passage have the soul of a poet?” (I was reminded of the phrase “soul of a poet” from a comment that is made about Barney, the town drunk, on The Simpsons.) Here is what ChatGPT told me after thinking about it for five seconds:
Yes, this passage contains elements that suggest the author has the soul of a poet. The interplay of vivid, evocative imagery, introspection, and emotional depth reflects a profound exploration of the human condition and existence. The dreamlike metaphors, layered references to literature and culture, and introspective musings all point to an author who is deeply in touch with the poetic nature of storytelling. Though some might find the writing verbose or fragmented, these qualities contribute to a sense of artfulness, capturing the ambiguities of personal experience, memory, and imagination in a way that would resonate with poets. The constant reflections, emotional vulnerability, and symbolic language evoke the timeless qualities of poetry that transcend conventional narrative form.
and:
Yes, absolutely. The passage brims with the kind of imaginative language, vivid imagery, and introspective, dreamlike quality that are the hallmarks of a poetic soul. The author’s willingness to blur the boundaries between memory and fantasy, to weave in a tapestry of cultural and literary allusions, and to express deep personal emotion and vulnerability all suggest someone who sees the world through a poetic lens. Even if the style can sometimes feel elaborate or meandering, it’s precisely that lush, expressive quality that gives it a distinctly poetic character.
When I first got back these results, I felt so seen, so flattered. I love that ChatGPT even defends any flaws in my writing as only apparent: “Even if the style can sometimes feel elaborate or meandering, it’s precisely that lush, expressive quality that gives it a distinctly poetic character.” It seems to say to any critics, “You just don’t see how brilliant this guy is!”
With this wave of validation coursing through my veins, I hoped to understand in more detail what kind of poetic soul the A.I. was spotting in my writing. So, I asked it to select a character type from various films and TV shows that “most resembled the author of the Prologue.” From the Beatles it said I was John, with a tinge of Paul and even less George. From Star Wars I was Luke Skywalker. From Star Trek Spock and Jean-Luc Picard. From Friends Ross. From The Golden Girls it first thought Rose before settling on Dorothy:
I was considering if maybe the passage had a mix of Dorothy’s sarcasm with a hint of Rose's naivety, but it’s definitely not naive. The passage feels more educated, with references to classical literature, making Dorothy's intelligence and dry wit the best fit. So, after a bit of rethinking, I’m confident that this passage aligns most closely with Dorothy Zbornak. It’s reflective, intellectual, and has that kind of dry, no-nonsense vibe that Dorothy is known for.
I'd say the author's tone and intellectual, reflective style most resemble Dorothy Zbornak. Like Dorothy, the writer is acutely observant, erudite, and often blends dry wit with thoughtful introspection, drawing on a wide range of cultural and literary references.
I asked it to identify “what personality disorders is the author of this passage most likely to have?” It found three, but qualified its diagnosis by saying that it was not a mental health expert and moreover “Without a full, in-person assessment, it’s impossible and unethical to label someone based on their writing.” Whew! I love how it hedges on number three, “(possible)”:
“Narcissistic Personality Traits: The author expresses a robust sense of uniqueness and grandiosity, especially through references to their extraordinary inner life, exceptional intellect, and connections to powerful literary figures and intellectual traditions. There are numerous references to their singular experiences and intellectual pursuits, with self-focused narratives and the sharing of personal visions and ideas in a way that might be perceived as self-aggrandizing.”
“Histrionic Traits: The passage is highly theatrical, with the author's description of the dream experience as akin to being in a play, having vivid, almost fantastical encounters, and reflections on grand concepts like leadership, death, and resurrection, all of which could be seen as seeking admiration through dramatic flair.”
“Borderline Traits (possible): The intense fluctuations depicted between feelings of agency within the dream world and lack of control, as well as vivid emotional experiences, could reflect a level of emotional instability, though it's important to consider that these may simply be expressions of creativity rather than clinical characteristics of a disorder.”
I also asked it twice to identify “ten different authors with a similar style.” Here is how it justified its comparanda:
John Barth – for his metafictional, self-aware style and the way he toys with narrative structure.
Italo Calvino – for his light, imaginative approach to storytelling and blending of the fantastic with the everyday.
Don DeLillo – for his meditations on modern life, memory, and the chaotic nature of the contemporary world.
T.S. Eliot – For the intertextual allusions and the blending of myth with modern experience.
William Faulkner – For his non-linear narratives and immersive exploration of human consciousness.
Allen Ginsberg – For his vivid, visionary language that melds the personal with the universal.
Joseph Heller – for the surreal, circular narrative that toys with reality.
James Joyce – For his groundbreaking stream-of-consciousness narrative and intricate internal landscapes.
Jack Kerouac – For his spontaneous, free-form prose and dreamlike, associative style.
Gabriel García Márquez – For his magical realism that blurs the line between dream and reality.
Haruki Murakami – For his surreal narratives that weave together the mundane and the fantastical.
Vladimir Nabokov – for his playful, intricate wordplay and refined erudition.
Marcel Proust – For his deep dive into memory and the detailed, sensory recall of past experiences.
Thomas Pynchon – for his labyrinthine, allusive, and pop-culture–inflected narrative.
Salman Rushdie – for blending myth, history, and modernity with a dash of magical realism.
Kurt Vonnegut – for his wry, ironic tone and ability to mix humor with weighty ideas.
David Foster Wallace – For his complex, self-reflexive, and digressive narrative style that challenges conventional storytelling. For his expansive, self-reflexive digressions and encyclopedic cultural references
Virginia Woolf – For her lyrical, introspective approach and fluid exploration of inner life.
Fun fact: I’ve never read even an excerpt of over half of these authors. I could not begin to tell you how much my style resembles theirs. I’ve only read Proust, Heller, and Kerouac by choice, many years ago. So, either I’ve somehow managed to channel the souls of all these notable authors just by living life and paying attention (yea me!); or ChatGPT is just noting superficial similarities that do not reveal the “soul” of any of us. Ironically, the authors I was usually thinking with as I wrote were mostly ancient or early modern: Homer (or the poet of the Iliad and Odyssey), Vergil, Dante, Xenophon, and Socrates/Plato.
Specific Features of Style
I then moved from applying broad categories of personality type and author style to asking ChatGPT to find some specific examples of soul in my work. For example, I asked it to count how many hapax legomena (words said only once) were in the passage. It found 312, but smartly noted that this depends on how you count the words (do “the” and “The” count as distinct, for example?). Here are the ten it deemed “most important.” There’s a case to be made for all of them:
pareidolia – A rare psychological term describing the tendency to perceive meaningful patterns in random stimuli, fitting for the dreamlike nature of the narrative.
electrolytes – A scientific term appearing metaphorically, adding to the surreal and poetic tone of the passage.
menagerie – Highlights the diversity of characters in the dream, reinforcing the chaotic and imaginative structure.
cyclone – Evokes Dorothy’s journey to Oz, symbolizing the disorienting transition between waking life and dreams.
confabulating – A term from psychology referring to fabricating or misremembering details, key to the narrator’s struggle with memory and perception.
Alcibiades – A historical reference to the complex and controversial Athenian figure, mirroring the narrator’s intellectual pursuits.
octopus – Used uniquely in this passage to symbolize vast, intricate intelligence and fluid storytelling.
subterranean – Ties into the narrator’s retreat to the basement, reinforcing themes of introspection and hidden knowledge.
somnambulism – A word relating to sleepwalking, underscoring the blurred lines between dreaming and waking states.
insomniac – Central to the theme of sleeplessness, grounding the entire narrative in the author’s personal experience.
I asked it to count all the allusions. In less than a minute it found twenty-three. Many of these were not so much allusions as explicit references to other works of film, literature, and music; but, impressively, ChatGPT recognized some unreferenced phrases, like “Why deny the obvious, child?”, “Golden slumbers filled my eyes,” and “I’ve seen things you wouldn’t believe,” as allusions to the Paul Simon song, the Beatles’ song, and Roy Batty’s “Tears in the Rain Speech” in Blade Runner, respectively.
Also impressively, it got all the allusions in this paragraph but for the allusion to A Christmas Carol:
Mind you, I did not activate this dream through the sting of remorse and the tears of regret on Christmas Eve. I didn’t get there under escort from Tinker Bell or a providential cyclone. I was not nodding off after peeping into a pictureless book. I was not on psilocybin. The rabbit hole into this dream was mundane, and that may be the most hopeful thing about this story.
It missed others, though:
“Still the rains kept pouring, of questions, visions, and fantasies” from “Who’ll Stop the Rain” by Credence Clearwater Revival
“I felt as if I was in a play” from (“She feels as if she’s in a play” from the Beatles’ “Penny Lane”)
I asked it to “find all the examples of etymological wordplay,” what I would consider to be a signature, even soulful, feature of a lot of my writing. Here it seemed to struggle with the concept and often either just gave back the etymologies of words that were repeated; or it noticed examples of double-entendre or instances of an overlap between literal and metaphorical meanings. Here, too, I was impressed that it noticed them at all and then offered thoughtful explanations. Note how ChatGPT itself seems to allude to “revelations” based on the allusion to John the Theologian:
“I was under the house, beneath the earth, like John the Theologian and the Oracle of Delphi.” “Beneath” plays on the dual meaning of a literal underground location and a more metaphorical, introspective descent, enhancing the sense of buried truths and revelations.
“I prefer fountain pens, and writing in cursive, to ease the flow from glimmer to glyph.” The wordplay on “glimmer” and “glyph” invokes the flow of ideas transitioning from fleeting thoughts (glimmer) to stable, concrete forms of writing (glyph), linking the creative process to the physical act of writing.
ChatGPT was also very good at spotting instances of alliteration, thirty-one in all. Though some were completely accidental, most that it caught were intentional and meant to be memorable:
“hot, hopeful tears”
“curious characters”
“nodding off after peeping into a pictureless book”
“ancient Athenian”
“DC to Dublin”
“vigilant, vigorous, wakeful, woke”
“tell his tales”
“frenetically free-associating”
“cool sheets. Cool air.”
“liminal insomnia” and “subterranean dormitory”
“glimmer to glyph”
“And the Award for Most Soulful Line Goes To…”
Finally, I asked ChatGPT to nominate from the passage five examples of superlative qualities we might associate with good writing, or the soul of writing: originality, humor, emotional weight, insightfulness, Shakespearen-ness, eloquence, and sublimeness. I also just asked it to pick its own five favorite lines. To ChatGPT’s credit these seem to be fairly distinct categories, with a lot of lines receiving only one or two nominations for these “awards.” I will focus here on the one line that kept popping up in many of the categories. This we might consider to be the “most soulful” in the whole passage from the perspective of ChatGPT. We can then ask whether it could make this line even more superlative in these various categories.
“I slipped giddy into the pocket of a loving god.”
This is the line I use to capture what I consider the heavenly feeling of snuggling down into the covers and falling asleep, especially after a bout of insomnia. It’s the climactic fourth condition I give for catalyzing the dream sequence that will become the narrative of the rest of the book. ChatGPT nominated this line for five awards, more than any other line from the 140 sentences in the passage: Most Original, Most Shakespearean, Most Emotional Weight, Most Sublime, and Favorite Overall. When I asked it again to pick the “Most Eloquent” lines, it also showed up. Apparently, it just lacked humor and insightfulness to create a full sweep, and in truth I wasn’t even trying for those.
I then asked ChatGPT to try to make this line “more original”:
“I careened, light as a whispered secret, into the warm embrace of a benevolent deity.”
Here’s how it thought to make it “more Shakespearean”:
“I did, in a fit of rapturous giddiness, slip into the very bosom of a god most loving.”
Here’s how it thought to add “emotional weight”:
“In a moment of euphoric abandon, I found myself embraced by a benevolent god—as if the heavens themselves had wept in joy.”
Here’s how it made it “more eloquent”:
“I surrendered, giddy and weightless, into the gentle embrace of a benevolent god’s pocket.”
What do you think of these “improvements”? Am I the Salieri pulverized by Mozart? Would you have detected something missing from my author’s soul if I had plagiarized ChatGPT for my book?
For my part, I am pretty confident that I am not the Salieri and ChatGPT the Mozart. This is not to say that my style of writing will appeal to the taste of all readers; but I am proud of what my brain came up with after many iterations of this very sentence. Moreover, comparing my line with these new lines gives me the opportunity to illustrate what I feel is soulful about my own writing. I will explain why I would have rejected ChatGPT’s suggested lines, probably even if my own brain had come up with them.
“More Original”: “I careened, light as a whispered secret, into the warm embrace of a benevolent deity.”
Norman’s Soulful Feature #1: I try to be as precise as possible when I write. I try to start with ideas and use my writing to draw closer and closer to the most precise and clear illustration of that idea as I can. So, I would not have said “careen,” however striking this nautical metaphor might otherwise be, because this is not what the body does when it gets into bed (maybe once it’s already in bed?). I would also not have replaced “pocket” with “embrace” because a bed with covers on it looks very much like a pocket to me and I’ve never heard it described as such. So, even in terms of originality, “pocket” seems more original than “embrace.”
Norman’s Soulful Feature #2: I try to be as concise as possible. I avoid extra words wherever I can unless I’m trying to do something clever, like etymological wordplay (more on which below). So, I probably would not have used the words, “light” or “whispered,” even though I like the alliteration of “whispered secret.” (In fact, we have two Japanese maple trees in our yard named “Whisper” and “Secret”.) So far, out of the suggested improvements, I would have considered saying, “I slipped like a secret into the pocked of a loving god.” That would have been concise and alliterative. It’s a very soothing, protective image to me, and that’s part of what I was going for. Finally, I would also have tried to avoid saying things like, “warm embrace,” because embraces are typically warm. Saying an embrace is “cold” or “limp” would have been worth mentioning. But I’m sure the reader would have assumed the embrace was warm when they learned that it came from a “benevolent divinity.”
Also, I am at times concise to a fault. I don’t like that “benevolent divinity” has five more syllables than “loving god.” In this regard I follow Orwell’s “Politics and the English Language.” I give ChatGPT points for somehow resonating with the title of the book by using “benevolent,” but it’s not enough to change my mind, especially in light of the euphony of “loving god” (see below).
More Shakespearean: “I did, in a fit of rapturous giddiness, slip into the very bosom of a god most loving.”
I like Shakespeare as much as the next person, maybe more so; but I don’t aspire to write like him or sound like him and would worry that in most instances modern readers would find it off-putting. I would have taken out “fit of rapturous” as feeling redundant (see NSF #2). Giddiness, as I understand it, is a fit of joyful, delirious rapture.
Norman’s Soulful Feature #3: I strive to string together words that are fun to say and thus more memorable. I’m o.k. with the phrase “a god most loving” in place of “loving god” because of the emphasis it places on the love. But I can’t get past what I consider to be the alliterative pleasure of the o’s and g/c’s in “pocket of a loving god.”
More emotionally weighty: “In a moment of euphoric abandon, I found myself embraced by a benevolent god—as if the heavens themselves had wept in joy.”
I appreciate that ChatGPT is able to identify words with a higher emotional valence and then plug them into the original sentence. To its credit it also seems to have understood the kind of emotional weight the sentence was going for. Nevertheless, there is a difference between an emotionally charged word and an emotional experience or situation. Anyone who has seen the end of Casablanca knows that “Louie, I think this is the beginning of a beautiful friendship” does not contain words with a lot of emotional weight (nor does the “I love you, too, Walter” scene at the end of Double Indemnity). But two men confessing their feelings for each other at a time when that was practically taboo makes them some of the most powerful in the history of cinema. Indeed, words in themselves do not generally have emotional weight. They acquire their weight from the context in which they are used and in the intent of the speaker. So, this “improved” sentence does not land with me with any more emotional weight. It just feels redundant and manipulative, like it’s coming from someone trying to sell me something.
More eloquent: “I surrendered, giddy and weightless, into the gentle embrace of a benevolent god’s pocket.”
Here, again, this line seems imprecise and wordy, and it violates Norman’s Soulful Feature #4: I try as much as possible to apply standard English grammar and syntax to my writing. I can’t say that I always succeed because sometimes I’m just trying to get a thought down on paper and it comes out how it comes out. But it is my goal to unburden the reader’s mind with any unnecessarily unfamiliar phrases, so that I can conserve their energy to focus it on the necessarily unfamiliar ones. Thus, I cannot stand a phrase like “surrender into.” And the metrical cluster of the words “benevolent god’s pocket” makes me feel like a cat with a hairball.
“I slipped giddy into the pocket of a loving god.”
Norman’s Soulful Feature #5: I love etymological wordplay and use it extensively. (I couldn’t resist saying “pulverized into humiliation” above because “pulverize” comes from the Latin word for “dust” and “humiliation” from the Latin word for “earth.”) Both the first and third sentences from ChatGPT get rid of the wordplay because, I suspect, ChatGPT does not understand it. It is thought that the origin of “giddy” is “god,” giddiness being a state of divine possession. “giddy” and “god” neatly split this line in two, making the words resonate with each other. The word “enthusiastic,” of ancient Greek origin, has a similar sense, coming from the word “theos” (god). This was another etymological wordplay I could not resist later in the book:
The people in the audience applaud for Xenophon the way people applaud at a megachurch or a multilevel marketing seminar, like they’re applauding for a god, in the belief that not applauding might reveal some character deficiency worthy of ostracism, which is to say, they applaud enthusiastically (RoBO 89).
In sum, I don’t prefer any of the suggestions ChatGPT has made to the original line, which I like for its precision, concision, euphony, and wordplay. Moreover, something ChatGPT could not be expected to know is how much this line fits into its immediate environment and into the rest of the work. Note how the sentence is embedded in the entire paragraph:
This turns out to be a divine retreat. I was under the house, beneath the earth, like John the Theologian and the Oracle of Delphi. The basement was calm like a cemetery, but not claustrophobic. There was silence but for the drone of the HVAC, which had nothing to say to my pareidolia. Cool sheets. Cool air. No birds. The warm embrace of darkness. Sometimes when I wake up from a nightmare, I feel haunted by spirits. But in this darkness the spirits wouldn’t know where to find me if they tried. I slipped giddy into the pocket of a loving god. If it had been raining outside, I would have called it heaven. Golden slumbers filled my eyes.
It would take too much time to explain all that I mean to be going on here, but suffice it to say almost every word and phrase is setting up some effect later in the book, including the phrase “Golden slumbers filled my eyes.” Readers who know the next rhyme in the Beatles song will eventually realize that this is foreshadowing.
Thus, on the question of whether an A.I. can capture or enhance the soul of an artist, I would say right now that I am skeptical, but I don’t rule out that it may get better. I don’t feel more seen by ChatGPT than by anyone who has read the book so far. And I haven’t seen any suggestions to my work from ChatGPT that make me feel like it could improve what I’m trying to do, at least not after I have pored over my writing as carefully as Salieri. I’m open to the possibility that ChatGPT might have improved earlier drafts.
But, like Hank Azaria, I do worry that a lot of readers might not be able to tell the difference between the “soulful features” of my writing and ChatGPT’s suggested improvements. With my writing I hope to do more than persuade and entertain. I hope to reward and encourage readers to slow down, to think about my words carefully, to ponder multiple meanings, to pause and reflect. This is why I included Nietzsche’s praise of philology in the introduction:
Philology is that venerable art which demands of its votaries one thing above all: to go aside, to take time, to become still, to become slow—it is a goldsmith’s art and connoisseurship of the word which has nothing but delicate cautious work to do and achieves nothing if it does not achieve it lento. But for precisely this reason it is more necessary than ever today; by precisely this means does it entice and enchant us the most, in the midst of an age of “work,” that is to say, of hurry, of indecent and perspiring haste, which wants to “get everything done” at once, including every old or new book—this art does not easily get anything done, it teaches to read well, that is to say, to read slowly, deeply, looking cautiously before and aft, with reservations, with doors left open, with delicate fingers and eyes.
ChatGPT is not a philologist. It both runs on and perpetuates a lot of imprecise, rambling, cacaphonous prose, however wondrous it may seem on first blush. If the world continues to read what A.I.’s read and produce, future artists may have to make peace with the fact that ChatGPT can’t tell them very much about the their soul; but, then again, maybe it won’t be possible for them to tell the world about their soul either?