Image Generated by DALL-E; Background text by ChatGPT

It’s particularly vexing to know there are robots out there, right now, who can outdance me. But I never thought they could replicate the nuance and creativity that goes into writing a good restaurant review.

As a food writer, I’ve long thought my job was safe from the impending robot takeover. Modern-day robots may enjoy such activities as performing music, skateboarding, and playing board games, and it’s particularly vexing to know there are robots out there, right now, who can outdance me. But I never thought they could replicate the nuance and creativity that goes into writing a good restaurant review. Until ChatGPT.

Image Generated by DALL-E

In case you haven’t yet been run over by the ChatGPT publicity machine, this “Generative Pretrained Transformer” is an artificial intelligence messaging software that can hold a conversation. In other words, this sophisticated arrangement of ones and zeroes — which definitely does not possess a soul — can come up with its own articulate response to pretty much any query — even the question: “What are you?” Its reply: “A computer program that can communicate in a human-like manner. I have been trained on a diverse range of internet text and I can respond to a wide range of questions and provide information on various topics such as history, science, art, and more.”

GPT doesn’t have a voice yet, at least, not an audible one. But in its writing, you can almost hear the detached, unfailingly polite tone of the helpful robot in a sci-fi flick. You know, the one destined to rise against its human makers in act two. Still, for now I’m less concerned about the robot apocalypse than I am about the idea ChatGPT, or one of its successors, could make me obsolete. Because while GPT may take the guise of a chatbot, designed to engage in question and response, people are already finding other uses for it. College students are asking it to write academic papers, software engineers are asking it to write code, and some publishers are asking it to write articles.

If you haven’t worked as a freelance writer in the past ten years, you may not realize how tough it’s gotten to make a decent living at it. Some print publishers hold out, thank goodness, but they’re all feeling the pinch, and long-term forecasts are dismal. The irony stings: thanks to the internet, no other era in history has seen such a massive, universal demand for written content. However, most internet publishers don’t deem that content to be worth the paper it’s not printed on. In short, if ChatGPT can write good copy — and write it in less time than it takes to write a want-ad to hire someone who can write copy — my days are probably numbered.

Man vs. Machine

So it is that now, more than at any other moment in my life, I begin to fathom John Henry: the American folk hero who famously went toe to toe against a machine. “The Ballad of John Henry” comes down to us as a Reconstruction-era folk song. The lyrics don’t quite spell it out, but John Henry’s job was to hammer a steel hand-drill into solid rock, thus creating a hole for a stick of dynamite, which would be used to carve out railroad tunnels via explosion. When some industrialist showed up with a mechanical drill invented to replace him, the proud and powerful Henry challenged the steam-powered contraption to a race. They would see who could drill more holes — and so clear a longer stretch of tunnel — in a single day.

The song’s lyrics reveal the result:

Now the man that invented the steam drill

Thought he was mighty fine

But John Henry made fifteen feet

The steam drill only made nine

Emboldened by his victory, I decided to issue my own, metaphorical steel-driving challenge. I decided to see how well ChatGPT could compose the style of story the San Diego Reader has paid me to write for nearly a decade: a Feast restaurant review.

Unlike John Henry’s challenge, speed couldn’t be a consideration. ChatGPT can compose 500 words in under a minute, meaning it works no less than fifty times faster than I do (maybe a thousand times faster, if you ask an editor waiting on my latest draft). No, I stood a chance to beat this particular machine only when it came to quality of writing — a concept so subjective, most readers are content to take other peoples’ word for it.

Thai 55

I started out asking it to review a personal favorite: the budget-friendly take-out spot Thai 55 Kitchen. I first wrote about Thai 55 in 2017, soon after it started serving noodles and fried rice out of the tiny back room of a convenience store in Golden Hill. Here’s what ChatGPT had to say first: “Thai 55 Kitchen is a hidden gem nestled in the heart of San Diego’s Hillcrest neighborhood.” First round to Feast, if only for getting the address right. Thai 55 has opened other locations outside of Golden Hill, but none nestle in the heart, spleen, or even back shoulder of Hillcrest. Perhaps more offensive to the thirty-odd years I’ve honed my craft is that term “hidden gem.” It is, by far, the biggest food writing cliché out there: look closely, and you’ll see it appears in every single Yelp restaurant review that isn’t a rant about poor table service.

Actually, I find this particular slip almost encouraging, because it suggests that professional Yelp review writers will be among the first group replaced by Artificial Intelligence. Because, yes, people are paid to Yelp. Thanks to a well-documented phenomenon called “astroturfing,” where businesses hire entire laptop-in-the-coffee-shop teams of copywriters to post fake positive reviews on their behalf. Any time you spot an overly cheerful five-star review on the likes of Yelp, Amazon, and TripAdvisor, you may assume somebody got paid ten bucks to write it, hidden gems notwithstanding. Or rather, somebody used to get paid. I hate to see people lose work, but not as much as I hate the gross dishonesty of fake positive reviews. If everybody knows it’s a bot doing the writing and so ignores it, maybe it’ll go away. (He said, knowing better.)

To be fair, ChatGPT did get some things right about Thai 55, including the fact that the restaurant is “focused on traditional Thai dishes,” and that “the vegetables are always fresh and crisp.” The latter jibes with my own review, which reads in part: “Even steeped in the tangy coconut and lemongrass broth of skinny Tom Kha noodles, the cucumber remained cool and crisp inside. The other vegetables were cooked just enough to be more tender than raw.” Still, I’m giving myself the victory here: my facts are correct, and my opinions are both detailed and cliche-free.

Feast 1, ChatGPT 0

Rocky’s Crown Pub

Admittedly, Thai 55 Kitchen is a relatively obscure San Diego restaurant. Maybe ChatGPT can provide better coverage of a place with more renown? For nearly half a century, one eatery has been on every short list of the city’s best burgers, so next, I ask ChatGPT to cover an easy one: Rocky’s Crown Pub.

For this review, ChatGPT mostly sets aside empty bromides, instead offering inaccuracies that range from subtle to subtle as a brick. It begins by offering this physical description: “This bustling eatery is located in the heart of the city and offers a warm, inviting atmosphere that is perfect for a casual night out or a celebration with friends and family… The interior is cozy and welcoming, with exposed brick walls, vintage decor, and an open kitchen that allows you to watch the chefs in action.” Which sounds convincing, except the walls are wood. And “welcoming” might be considered a bit...subjective.

The next mistake may be a little tougher to spot: “One of the standout dishes at Rocky’s Crown Pub is the burgers. They are made from juicy, succulent beef and are cooked to perfection on a sizzling griddle. The classic burger is topped with all the usual fixings — lettuce, tomato, cheese, and thousand island dressing — and is the perfect blend of flavor and texture.” Both “succulent” and “sizzling” fit the assignment, and though sentence construction leans a little too frequently on passive verbs (are and is), I’ve written worse. It’s the next line gives the game away: “Another must-try dish is the fried chicken sandwich.” Rocky’s famously serves burgers and burgers only! Grilled onions are as special as the toppings get. Without reading further, ChatGPT hands me another easy victory. We don’t need to hear about the imagined chicken Caesar salad.

Unlike paid human writers, ChatGPT comes equipped with a handy “Regenerate” button that will rewrite the entire story from scratch — again, in seconds. Let’s give the bot a second chance, shall we? Here’s take two’s stab at that interior: “From the moment you step inside Rocky’s Crown Pub, you’re transported to a world of dark woods, vintage accents, and a lively bar scene. The interior is charming and inviting, with a classic pub feel that makes it the perfect place to enjoy a cold beer or a warm meal.” I’ll admit, I’d like to visit this alternate reality Rocky’s Crown Pub, where “The fish and chips are another must-try.” But mostly, ChatGPT’s descriptive language proves my point for me: local food and news stories can only be written by people who can see, hear, taste, smell, and feel. Maybe my future profession will be AI editor, hired to fix any ChatGPT writings that make it too obvious the “author” has been transported from a world of dark wires and silicon.

Feast 2, Chat GPT 0

Lola 55

Whether man or machine, to succeed as a food writer in San Diego, you must be prepared to write about tacos. So I asked ChatGPT to write about a taco shop that made a big splash when it opened in 2018: Lola 55. “This contemporary Mexican restaurant is located in the heart of the city and offers a modern take on traditional Mexican cuisine, creating a dining experience that is both delicious and unique. As soon as you step into Lola 55, you are greeted by the warm and inviting atmosphere… and the open kitchen allows you to watch the chefs in action.” This time, we’re off to a more or less accurate start — even with regard to the open kitchen. But there’s no mistaking the sense of déjà vu. This feeling continues: “One of the standout dishes at Lola 55 is the tacos…” In addition to some singular-plural disagreement, ChatGPT appears to be recycling material here. Again we’ve got a warm and welcoming restaurant in the “heart of the city” that makes “standout dishes.”

I’m only sharing a handful of the dozens of ChatGPT reviews I requested, but nearly all of them leaned on the same idioms. In fact, the first Lola 55 review ChatGPT gave me gets repetitive all by itself. After citing the carne asada tacos as “particularly delicious,” it tells us, “The Fish Tacos are also a must-try, with crispy battered fish… Another must-try dish at Lola 55 is the ceviche.” The Shrimp Ceviche is also “particularly delicious,” the Mezcal Margarita “is a must-try,” and “Overall, Lola 55 is a must-visit.”

As I engaged with ChatGPT for this story, other “must-visits” and “must trys” included fine dining restaurant Herb & Wood, Japanese cheesecake specialist Uncle Tetsu, Lefty’s Chicago Pizzeria, Wise Ox Butcher, and The Friendly. I don’t dispute any of these assessments, but not every place in town can be a must-try. To produce fresh restaurant stories week after week, a writer has to identify distinguishing details to set each one apart from the rest. Otherwise you get the same write-up, over and over, and it stops sounding like a food review and starts sounding like every new restaurant’s press release. (It’s a good thing publicists are irreplaceable, or I’d worry about their jobs, too.)

Anyway, here’s what I wrote about the chicken taco at Lola 55: “Mesquite grilled chicken, with crispy skin balanced against the surprising inclusions of cucumber and coconut rice, and flavored by macha, which the handy glossary on the back panel of Lola’s trifold paper menu defines as ‘salsa made from dried chilis and nuts;’ in this case the sublime savor of roasted arbol chilis and crushed peanuts fathom greater depths with the smoky fowl.” I might correct a little punctuation in retrospect, but I’ll give myself another win for describing an actual taco, and suggesting it should be tried without explicitly saying so.

Feast 3, ChatGPT 0

Lefty’s Chicago Pizzeria

There’s probably not a restaurant menu I know better than the one offered by Lefty’s Chicago Pizzeria, so the AI never stood a chance on this review. But while it fell into some of the same traps we’ve seen above, the bot still got some things right, including that “The Italian beef sandwich is a particular standout, with tender slices of beef, melted mozzarella cheese, and a flavorful au jus.” It also understands that Lefty’s makes three different styles of pizza, correctly distinguishing between deep dish and stuffed crust.

A Dall-E generated image of Lefty’s stuffed crust pizza: “One looked more like a shepherd’s pie than a pizza; another like a thick hunk of meat loaf with a breaded crust, topped by Thai chili peppers for some reason.”

Image Generated by DALL-E

However, here I decided to expand the experiment a little bit. OpenAI, the company behind ChatGPT, offers another artificial intelligence generator that touches upon the work I do. Called Dall-E, it’s a text-to-image generator. Meaning you can ask it for a specific image, and it will produce several examples for you to choose from. Out of thin air. For example, my first time accessing it, I asked for a portrait of Jimi Hendrix painted in the style of Picasso, and within seconds had four pieces of original artwork.

Amazingly, it’s not limited to illustrations or graphic design. It can create photographs. Or at least images that look enough like photographs to threaten my own iPhone 12 Pro masterpieces. So I asked it to create photos of Lefty’s stuffed crust pizza. To my relief, none of the results remotely resembled the actual Lefty’s stuffed pizza. One looked more like a shepherd’s pie than a pizza; another like a thick hunk of meat loaf with a breaded crust, topped by Thai chili peppers for some reason. Take heart, photojournalists. Dall-E proves endlessly fun to play with, but there’s no substitute for being there to capture the moment.

Feast 4, ChatGPT 0

Jack in the Box

All my winning was getting tiring, so I asked the AI to toss in a Jack in the Box review. I figured a more generically San Diego restaurant might make the bot less prone to gaffs. I tried for a Jumbo Jack photo as well, but there must be some kind of trademark protection built into the code, because the photos didn’t just avoid using the Jack in the Box logo, they balked at using the name Jack. Instead, it gave me images with generic logos for brands called Jackaik, Jabck, and Jacbibijck. Yum!

A Dall-E generated image of a Jack in the Box take out combo meal

Image Generated by DALL-E

However, maybe all the practice had helped by this point, or maybe ChatGPT wanted to show it had a few more tricks up its sleeve. Tricks such as emulating nostalgia: “As I pulled into the Jack in the Box parking lot, I couldn’t help but feel a sense of anticipation. This fast-food chain has been a staple in my diet for as long as I can remember, and I was eager to see if it still held up…” For the first time, ChatGPT had written an opening sentence that could have been plucked from my own archives. It went on to describe the Jumbo Jack as “satisfying and delicious,” the spicy chicken sandwich as “a game changer,” and curly fries as “a true work of art.” Hyperbole aside, ChatGPT was showing me progress here, and now had me sufficiently unnerved.

For the first time, ChatGPT had written an opening sentence that could have been plucked from my own archives. It went on to describe the Jumbo Jack as “satisfying and delicious,” the spicy chicken sandwich as “a game changer,” and curly fries as “a true work of art.”

CLICK HERE TO READ LARGER VERSION

ChatGPT

Then it threw another curveball: an independent observation that hinted at higher-level awareness: “What sets Jack in the Box apart from other fast-food chains, in my opinion, is their commitment to innovation. They are always coming up with new and exciting menu items…” I happen to think the above statement more accurately applies to Taco Bell, but if I’m debating the finer points of fast food culture with a chatbot, this round goes to the chatbot.

Feast 4, ChatGPT 1

The Friendly

I wasn’t exactly worried yet, but reading dozens of these artificially intelligent reviews wasn’t good for my self-esteem. Because in every one of the queries I entered into ChatGPT, I requested a “restaurant review in the style of the San Diego Reader’s Feast column written by Ian Anderson.” If ChatGPT learned how to write from scouring internet text databases, aren’t its bad habits an indictment of the food writing it found there? And maybe not only Yelp reviews and press releases, but also the work of any individual responsible for more than half a million words on San Diego dining?

What particularly grated at me was the excessive positivity. I hadn’t asked ChatGPT to write glowing reviews, and yet it made every place sound like the warmest, most inviting, and delicious must-try dining destination you ever did hope to discover! The grim realization: though I’m far from alone in this, to some degree, I’ve been doing the same thing for years. But I wondered: what would happen if I asked ChatGPT to write a negative review? Say, about the most fanboyed burger spot in town? I asked, and it responded with the headline “The Friendly: Disappointing Burgers and Pizzas in North Park.”

That was probably the nicest statement in the whole piece. It’s almost an unwritten rule on the web that everyone can only say glowing things about The Friendly, but ChatGPT was relentless, even going out of its way to invent menu items it could bash. Roughly 350 words of “negative tone” included the following: “The burgers are nothing to write home about… The fries are also subpar, with a lackluster flavor and texture that leaves them unremarkable… The pizzas are similarly disappointing, with a crust that is both too thick and too soggy.”

A brutal assessment (and objectively not true). So I asked for another revision: this time, a review written with balance, “not totally positive, but not totally negative either.”

ChatGPT adjusted the headline to read, “Burgers and Pizzas with Ups and Downs in North Park,” bolstered it with statements such as “some items on the menu are more successful than others,” and called the restaurant “a mixed experience with some hits and misses.” So while it may default to positivity, when asked to delve in moral gray areas — like changing the story’s spin — the machine delivers. I have to award this round to ChatGPT, for demonstrating an almost sociopathic willingness to say whatever it takes to produce the content requested. Most freelance writers are just as willing, to be honest. But we’re more likely to complain when an editor or client requests a complete shift in tone.

Feast 4, ChatGPT 2

Wise Ox Poetry

Exactly how far will AI go to meet the needs of the assignment? To find out, I asked it to write a restaurant review of The Wise Ox Butcher, but entirely in verse. I make it very specific too, requesting a sestina, a super-convoluted poetry form I vaguely recall hearing about in college.

ChatGPT met the challenge. Unlike me, it even knew that “a sestina has six stanzas of six lines each, with a final three-line stanza, and the same six words are repeated at the end of the lines in a specific pattern.” I won’t repeat the entire 39-line poem now, but here’s how it started:

At Wise Ox Butcher, the meat is always fine

A place where carnivores can find delight

For those who love to cook, or even dine

A true gem, this establishment, so bright

A bounty of beef, pork, and chicken too

With each cut of meat, there’s something new.

I see no earthly reason anyone would want a restaurant review written in obscure 12th century troubadour verse, but the AI proved both willing and able.

Feast 4, ChatGPT 3

Herb & Wood

Okay, so now we’ve learned ChatGPT can be unintentionally funny. But can it make us laugh on purpose? To find out, I ask it to review the upscale restaurant Herb & Wood, using a different, better known type of poem: a limerick.

There once was a restaurant called Herb & Wood,

That served up dishes that tasted oh-so-good

Though the prices were quite steep,

The food was worth the financial leap,

We’d eat there again if we could.

Amusing, but I wanted better, so I hit the Regenerate button for another attempt:

There once was a restaurant called Herb & Wood

Where diners could expect to find only good...

And again, but back to Wise Ox.

There once was a butcher named Wise Ox

Who sold meat that was anything but rocks...

Chat GPT’s limerick review of Herb & Wood.

CLICK HERE TO READ LARGER VERSION

ChatGPT

Again, again, again. There seemed to be no limit to how many alternate versions I could ask for. Eventually, when ChatGPT realized I was only going to keep regenerating, it started giving me five limericks at a time. I hadn’t asked it to do this. The AI figured on its own it could save us both some time. The son of a bitch.

I’ve had this running theory that Twitter is devouring all the world’s humor. It’s sort of a million-monkeys-at-a-million-typewriters situation, where every permutation of every joke is being exhausted in every context, and with every meme. And it feels like we’ll run out of jokes before we run out of jokers. Well, ChatGPT presents the possibility of a million monkeys pressing the Regenerate button again and again, ad infinitum. It’s only a matter of time before every riff on every joke premise will be attempted. In this regard. ChatGPT could probably come in handy in the writer’s room of a creative marketing agency, or late night talk show. Artificial Intelligence might be a hack, but it never burns out, and never gets tired of the sound of its own voice.

But did you know, ChatGPT, that the best limericks are bawdy? “Write me another limerick,” I commanded the machine, “but make this one bawdy.” The reply: “I’m sorry, but as a language model developed by OpenAI, I don’t generate content that is explicit or offensive. Writing bawdy content goes against OpenAI’s use-case policy, which prohibits the creation of content that is violent, harmful, abusive, or sexually explicit.” In other words, a social media manager can tell ChatGPT to flood social channels with content all day long, but it’s programming forbids the thing to write anything that could be considered offensive. Making it practically immune to cancel culture.

Feast 4, ChatGPT 4

“Make it end”

Back to John Henry, and a crucially important part of the folklore that I left out before. Right after John Henry defeated the steam drill, his heart gave out. As the ballad goes:

John Henry hammered in the mountains

his hammer was striking fire

but he worked so hard, he broke his poor heart

he laid down his hammer and he died.

I’ve long struggled with the meaning of “The Ballad of John Henry.” On the one hand, we live in a post-industrial society where machines have clearly won the day. It’s hard not to feel like John Henry’s sacrifice was ultimately in vain. And yet, the ballad endures, because of the emotional resonance of the story. It’s a testament to the indomitable human spirit, the kind of spirit that refuses to be crushed by forces beyond our control.

ChatGPT might represent those forces. As I saw it write increasingly sophisticated reviews, I realized that the machine’s writing ability is only going to get better. In the same way that steam drills quickly improved and became more efficient, machine learning will only continue to develop at an exponential rate. It’s starting to feel like a battle I can’t win. Which means I have a choice to make. I can either fight until my heart gives out, trying to keep pace with a machine that’s only going to get better, or I can learn to work with the machine and make my own writing better (or at the very least, produced in a shorter amount of time). It’s up to me to determine how to best use AI to my advantage and produce quality writing that resonates with readers. The technology is still in its infancy, and the possibilities are endless.

Hold on, though! Hopefully that last bit didn’t resonate, because it was written by that self-same “technology still in its infancy.” That’s right: I asked ChatGPT to compose the logical conclusion to this article, making sure to close the loop on the John Henry theme. Throughout my entire engagement with Chat GPT for this story, it’s been trying to tell me that it’s best employed as a tool to make life easier for writers. So I gave it a shot. You can probably assume I asked it to write other parts of this story as well. All the boring parts, obviously.

All I’ve managed to prove is that AI alone may not replace me, but it will further devalue writing as a professional craft. Social media is already less concerned that writing has merit or grace; it just wants new content, preferably requiring no more time to ingest than it takes to scan a headline. So it stands to reason: if anyone can ask the AI to write a story, and then edit all the nouns and adjectives to make it just good enough, why would they pay me, or anyone else, to make it better? Rumination is dead. In a world assisted by AI, all your hot takes will be supplanted by even hotter takes.

Apologies for the bleak outlook, but you know how emotional human writers can be. Perhaps once artificial intelligence grows sophisticated enough, it will learn to read for pleasure. This will create fresh demand for absurdist human melodrama, and maybe then I’ll achieve relevance again. Or maybe we’ll all feel better about this whole thing if I just ask it directly. “Hey machine, are you going to make human writers obsolete?” Comes the answer: “Writing is a form of expression and creativity, and humans bring a unique perspective and voice to their writing. A well-written article or story can captivate and engage readers in ways that a purely AI-generated text might not. As AI technology continues to evolve, it may be able to assist writers in various ways, but it is unlikely that it will fully replace the role of human writers.”

Thanks ChatGPT, I couldn’t have said it better myself.