Let's break down some AI-generated text
And don't you dare use AI generation detectors
Over the past couple of months, I’ve witnessed a dramatic surge in well-liked, widely-shared AI-generated content across the Internet. If you don’t believe me off the bat that 1) this content is everywhere and 2) everybody loves it, I’d recommend checking out my last post before reading any further.
Since I wrote that post, the situation has only gotten worse. Nearly all of my favorite business’s Instagram captions are AI generated. Every recent Wikipedia page contribution is AI generated. The narrations in the Facebook reels my dad watches out loud in the living room (love you!) are AI generated. I see it in LinkedIn posts and business emails and Substack posts and Twitter ads, and at this point, if I have to see it everywhere, so do you.
First up, an exercise for the reader:
One of the following paragraphs about the bird flu was pulled from a recent BBC article. One was generated using a bare-bones prompt for an article about a fictional virus called ‘bird flu’. One was written by me. Which is which?
#1:
The first U.S. case of bird flu has emerged in Idaho's Snake River Valley, marking a grim milestone as the deadly pathogen makes its first confirmed jump from birds to humans on American soil.
The case, identified in a 42-year-old poultry processing worker at the Valley Pride processing facility in Twin Falls, has triggered an aggressive response from state and federal health officials. The facility, which processes over 100,000 birds weekly, has been shuttered indefinitely for emergency containment procedures.
"We've identified an unusual pattern of mortality in the facility's bird population over the past two weeks," says Dr. James Harrison, Idaho's State Veterinarian. "Initial testing confirms these deaths were linked to H5N1, suggesting the virus had been circulating in our domestic bird population undetected."
The infected worker, whose name is being withheld for privacy, is currently in isolation at St. Luke's Magic Valley Medical Center.
#2:
Bird flu infected cattle for the first time ever in the US at the beginning of last year. Since then, there have been nearly 1,000 confirmed cases across 17 states. There also have been 68 confirmed cases among humans, most of whom worked closely with animals.
Then, in February, a new variant of the virus - called D1.1 - was detected in dairy cows and an infected worker in Nevada for the first time. The strain also caused two severe infections in humans in North America, a teenager in hospital in British Columbia and a patient in Louisiana who died from the virus.
The new variant also means there are now at least two instances in which cows have caught the virus from wild birds, and not other dairy herds.
"We're seeing these little red flags, real signs that the virus is making some inroads," Mr Pekosz said.
#3:
Following last week’s reports of poultry culls across Utah to curb the spread of bird flu, the first confirmed human case of H5N1 in the state has been identified in a middle-aged farmhand.
“Preventing crossover infection is going to be difficult,” admits James Morgan, owner of Stonyriver Farms, a mid-sized Utahn commercial farm. “We’re monitoring the situation. Keeping our workers safe is our top priority.”
State representatives confirm that government officials are working towards the deployment of emergency relief for affected farms and families, but warn that consumers might bear the burden of supply shortages until the virus is under control.
In a Tuesday statement, Governor Spencer Cox called for bravery: “We’ve seen this kind of thing before, and we’ve beaten it. We can do it again.”
Answer and discussion after the break.
1: AI; 2: BBC; 3: Me! I made it all up on the spot, and then checked who the Utahn Governor was, and whether I’d spelled “Utahn” right.
So, this one is sort of unfair. None of the tricks I’m about to describe are present in the AI-generated text; the only reliable distinguisher between the paragraphs is prior knowledge on the topics discussed or explicit fact-checking (and even fact-checking could be misleading, considering ChatGPT name-dropped what appear to be real places in Idaho and I just made up a farm).
A couple points I’d like to emphasize here:
There is no silver bullet for identifying AI-generated text. Assuming that you’ll be able to tell the difference is setting yourself up for failure. Oh, and any program or website that says that they can is lying or selling you a garbage product.
Human-written text is not inherently safer than AI-generated text. Well-intended human comments can be just as wildly confusing and misleading as AI-generated ones, and often humans are intentionally dishonest. It just takes a whole lot longer for humans to write.
Ultimately, the most important cognitive security skill here is general textural analysis: being able to carefully evaluate meaning in text, determine authorial intent, and analyze the soundness of arguments in the pursuit of distinguishing truth.
I don’t presume to be able to teach you to distinguish truth reliably from fiction, especially on today’s Internet, but I can try to help you identify suspicious content. With that disclaimer, let’s get into it.
In the interest of providing specific examples, I’m going to go ahead and dissect one of the Claude 3.5 Sonnet pieces I generated for my last post, comparing it to suspicious text I recently found out in the wild.
Obviously, the Claude piece is completely AI-generated. I would know, I generated it myself, with the prompt “please generate a funny post for a programming subreddit about a plucky gen-z intern finding an important bug in a hilariously unconventional way”.
This exercise should provide some helpful examples of what AI-generated text tends to look like, but, again, please be aware that just as it’s impossible to capture the breadth of human writing habits in a single example, LLM writing habits are similarly difficult to pin down to a science. That being said, if you work on training your eye, you, too, will soon have the distinct pleasure of noticing it everywhere for the rest of your life. :D
The Introduction
So our summer intern (let's call her Emma) has been with us for about a month. Super bright kid, but definitely operates on a different wavelength than our mostly millennial dev team. Yesterday she saved our entire sprint in the most Gen Z way possible.
Alright, so I’m gonna preface this by saying that, in hindsight, this was probably a mistake. I understand that now. But at the time, it felt like the funniest thing I could possibly do. So, last night, I went to Olive Garden with my buddies. We were already a little elevated, if you catch my drift, and at some point, I got this brilliant idea: What if Olive Garden was even more unlimited?
Both pieces start off relatively innocuous. Obviously, the timeline is off in my Claude programming piece – I generated this in February, about as far from a month into summer as we could possibly be. The important structural note here is that both posts have a conversational introduction that still manages to set up a very tidy story.
The Claude text immediately sets up the silly millennial vs. Gen Z theme and our intern as an unconventional hero with exactly one personality trait: being the Internet’s Gen Z stereotype personified.
The Reddit text is a little different, since our hero is also the narrator of the story, but the introduction still sets up the silly extra-unlimited Olive Garden theme and our narrator as the unconventional hero with exactly one personality trait: being the Internet’s stoner stereotype personified.
Progression of the Story
Throughout the rest of the Claude story, our Gen Z intern hero’s only personality trait continues to be Gen Z intern.
Enter Emma, who had been quietly observing our debugging sessions while making TikToks about "corporate girlie life" (which, btw, have apparently gone viral in tech TikTok).
She literally found a race condition because she was trying to nail the timing of a TikTok transition. 💀
The best part? She explained it to the team using the actual TikTok as a visual aid, complete with a "Think Fast!" sound effect at the exact moment where the race condition could occur. Our tech lead just sat there in stunned silence before going "...that's actually exactly what's happening."
Similarly, throughout the rest of the Reddit story, our stoner hero’s only personality trait continues to be stoner.
Now, I happen to have in my possession a very potent THC concentrate. Like, this isn’t your average dispensary gummy—this is a substance of unknown origin that I got from a guy named Kyle who lives in a van. I don’t even know if it’s technically legal.
Anyway, I sneak into the kitchen and start drizzling this stuff into everything. The breadsticks? Laced. The alfredo sauce? Absolutely infused. The minestrone soup? Bro, I blessed that minestrone.
I was careful, though. Real sneaky. A little drop here, a little swirl there. I was like the Walter White of weed. Except instead of cooking meth, I was enhancing people’s tour of Italy.
Both of these characterizations seem to draw from the shallowest connections to the defining personality trait:
The Gen Z intern is TikTok viral
The stoner self-describes as the ‘Walter White of weed’ and gets his drugs from a guy named Kyle who lives in a van
These simple characterizations make sense in the context of large language models, which generate content based on statistical associations from vast amounts of internet text and are thus primed to reflect the most common of stereotypes.
Also, these sentences show off a couple of technical tells:
ChatGPT’s little flag is the spaceless em-dash. (i.e., “this isn’t your average dispensary gummy–this is a substance of unknown origin…”) Using an em-dash as such to denote natural breaks in speech is common in professional writing, but not casual internet writing. Additionally, many internet users that do use them (like me!) often surround the dash with spaces. But ChatGPT loves spaceless em-dashes, and will put them absolutely everywhere it can.
Additionally, large language models love overusing rhetorical questions. Look for “The best part?” in the Claude post (and usually “The worst part?” or other “The [___]est part?” immediately after, but that’s not demonstrated here). Oftentimes, these rhetorical questions will come in groups of three – which of course, you can see in ‘The breadsticks? The alfredo sauce? The minestrone soup?” in the Olive Garden post.
The Punchline
tl;dr: Gen Z intern debugged our production system by accident while making TikTok content, and I've never felt more simultaneously impressed and ancient.
So now I’m sitting at home, absolutely losing my mind. Did anyone catch me? Is there security footage? Am I about to be the first person to serve federal time for enhancing an unlimited soup, salad, and breadstick experience?
The beginning of a good conclusion in an essay is effectively a restatement of the thesis, and both of these punchlines follow exactly that logic. Our Claude punchline threads back in the millennial vs. Gen Z divide with our narrator’s ‘ancient’ statement, and in the Reddit story, we’re back to the theme of unlimited Olive Garden. Each story is wrapped up just as neatly as it began from a technical standpoint.
Casual human storytelling is usually not nearly this neat. Take it from a writer: it’s incredibly difficult for most humans to write the perfect conclusion. In the process of writing, we start off with one idea, get sidetracked throughout by marvelous new ideas, perhaps realize that at least one of those ideas – or worse, your original idea – is wrong and bad, rewrite, write some more, and eventually end up with a first draft. In order to have a sensical final piece with a neat introduction and conclusion you likely need to rewrite the entire thing at least once.
Current generations of large language models are primed to skip this mess via attention mechanisms. Given an idea and an acceptably large context window for the output size, models can generate a hook, however much body text, and then eventually a conclusion, never straying too far off topic.
The Conclusion
Edit: Since so many people are asking - yes, we offered her a full-time position for after graduation. No, she hasn't accepted yet because she's "waiting to see if her TikTok channel takes off first." 😭
So Reddit, am I going to prison? Should I just skip town and live in a van with Kyle? Or should I ride this one out and hope for the best? I never thought I’d be asking this, but I might just be the first guy ever to be arrested for enhancing an all-you-can-eat deal at Olive Garden.
The last line of a perfect essay should ideally leave the reader with a lingering sense of interest in your topic. The Olive Garden post takes the call-to-action approach and directly addresses the audience, restating the punchline one more time for effect. Our Claude post acknowledges the audience and simply ends with another Gen Z joke. Either way, both of these are technically perfect final lines.
A few other random things:
Paying attention to narrative consistency is helpful: the fear of serving federal time is inconsistent on a meta level with a narrator that isn’t sure whether a high-potency THC substance is technically legal. A human writer that knows what a dispensary is but doesn’t know off-the-bat that ‘high-potency THC substance of unknown origin’ is federally illegal doesn’t exist. But next-token prediction trained on our entire Internet absolutely would generate that sentence in the context of getting a substance of unknown origin from a guy named Kyle who lives in a van, as well as come up with the idea of writing “The Walter White of weed, except instead of cooking meth, I was [doing stuff with weed]” in 2025.
I threw down a $20 (because I still respect service workers), grabbed my friends, and ran.
The narrator doesn’t actually pay for the table’s meal, but it’s a faux-pas on the Internet to leave without tipping. A human wouldn’t write this sentence because of the logical inconsistency, but it’s perfectly generatable by chatbots that tend to avoid going against widely-held Internet ethics.
Also, since neither of these texts were examples of argumentative text, I’m going to pull a quote from a literary analysis test I had ChatGPT o3 do (with light censoring to avoid accidentally spoiling the book in question).
While the text leaves room for interpretation—inviting debate over whether the events are to be read as historical fact or symbolic allegory—I lean toward the view that the murder of the original guy in the woods was an actual event within the narrative framework. [author’s] masterful blend of ritual, memory, and myth is designed to make the reader feel the profound, tangible consequences of that act, regardless of its elusive details.
Yeah, so, this quote means nothing. o3 sucked at the literary analysis exercise. For context, the question being asked was not whether the guy was murdered, it was who did it. o3 was able to construct a vaguely nice-sounding argument, but was unable to actually connect evidence to interpretation, so we ended up with a conclusion that uses a lot of good words, but is entirely meaningless upon close reading.
Many arguments generated by AI will lean into this same pattern – nice-sounding argument, no substance.
Which, of course, is the entire problem. Text generated for entertainment’s sake, like the stories we analyzed, ultimately doesn’t matter terribly much. It’s low-value content, sure, but the chances of an AI-generated subreddit joke posts causing real-world harm is low. Well-constructed, confident arguments for nonsensical stances, however, could absolutely cause harm.
So, in conclusion, suspicious text can generally be identified via grammatical clues (spaceless em-dash overuse, overuse of rhetorical questions, groupings of three), technical tidiness in contexts that shouldn’t be quite so tidy, non-logical adherence to well-known Internet stereotypes and ethics, and general lack of substantive meaning behind confident, pretty sentences. As AI generation continues to improve, these tells will likely become increasingly less useful. Eventually, we’re going to have to commit to consistently doing the truly difficult thing: judging text not by our well-honed heuristics of trusting professional or well-executed writing, but entirely by the merit of its content.
Next time, we’ll take a closer look at generative argumentative text and analyze the implications of usage patterns that are beginning to appear both online and in the real world.