I've been reading AI-generated content for months now, and something started bothering me. Not the obvious stuff, the robotic tone or the weird formality. What got under my skin was subtler: I could feel when text was machine-generated before I could explain why.
So I started keeping notes. Every time I spotted something that made my hacker's brain twitch, I wrote it down. The list got long fast.
The Words That Give It Away
"Delve" was the first one I noticed. Nobody delves anymore. We look into things, we examine them, we study them. But AI? It delves. Constantly. I've seen it in hundreds of articles, always in the same context: "Let's delve into the complexities of..."
Then came "tapestry." Everything became a tapestry. Ideas formed a tapestry. History was a tapestry. Communities were tapestries. I checked my own writing from the past decade. I'd used "tapestry" exactly twice, both times describing actual woven fabric.
The pattern became clear. These weren't just overused words. They were statistical artifacts, the most probable next token, selected over and over by models trained on billions of pages of mediocre web content.
The Persona Problem
AI doesn't just use certain words too much. It tries on personas like badly fitted costumes.
There's the Grizzled Expert persona, which shows up when the text says "the hard truth is." Nobody earned that authority. The model just calculated that this phrase precedes certain types of statements in its training data.
There's the Folksy Storyteller, who appears when you see "here's the kicker" in formal business writing. Real people use this phrase in conversation, usually after building up to a punchline. AI drops it into product descriptions.
There's the Didactic Helper, who cannot resist telling you that something "plays a crucial role" or that "it's important to note that." Everything gets emphasized. Nothing can speak for itself.
Why This Happens
I spent time looking at research on model training. The explanation is simpler than I expected: regression to the mean.
AI models are trained on the vast middle of human writing. Not the brilliant stuff, not the terrible stuff. The average. That average includes millions of SEO articles, corporate press releases, and forum posts. The phrases that appear most frequently in that massive dataset become the phrases the model reaches for first.
Add in safety training, teaching the model to be helpful and harmless, and you get hedging. Lots of hedging. "To some extent," "it can be argued," "in many cases." The model learns that qualified statements are safer than definitive ones.
The result is prose that sounds polished but stiff. Comprehensive but soulless. Like it was written by a very diligent student who read the assignment requirements but didn't have anything particular to say.
What I Do About It
I don't run every piece of content through detection software. That's an arms race nobody wins. Instead, I edit.
I built myself a searchable list of problem phrases. When I'm reviewing anything that might have been AI-assisted, I run through the list. Find every "moreover" and "furthermore." Cut them. Find every instance of something "underscoring" something else. Rewrite it.
The phrases themselves aren't forbidden. A human writer might occasionally use "moreover" when it's genuinely the right transition. But when I see five instances in 500 words? That's a machine optimizing for formal connectives.
The Real Tell
Here's what I've learned: the definitive marker isn't any single word. It's density.
One "delve" in a 2,000-word article? Could be human. But "delve" plus "tapestry" plus "moreover" plus "it's worth noting that" plus "in conclusion"? That's not a person. That's a prediction engine selecting the statistically most probable sequence of tokens.
The machine doesn't choose words. It calculates probabilities. That's why the prose feels calculated, because it is.
I can spot it now within a paragraph. Not because I'm memorizing lists, but because I've trained myself to hear when writing lacks choice. When every sentence feels like the safest possible version of itself. When the text is optimized for being correct but not for being interesting.
The models will get better at hiding these tells. They'll learn to avoid "delve" and "tapestry." New patterns will emerge. But the fundamental issue won't change: statistical compulsion produces different prose than human choice does.
My job now is to put the choice back in or...
...to use clever prompt engineering to make the output sound more unique and more "human".
Growth hacking means finding advantages - if everyone knows the trick, it's no longer an advantage, does it? Everyone "knows" they are supposed to use AI for automation, but they put very little effort into it.
So here's a short and sweet result of a LOT of research that is compiled into this nice addition to your prompting:
Follow these principles for good writing:
Write in plain, direct English using simple, clear sentences
Use first-person perspective ("I") when appropriate and write with the humility of a first-hand observer
Describe only what can be observed, avoiding claims about what is "profound" or "pivotal"
Eliminate redundancy and tighten sentence structure for clarity
Avoid the "Rule of Three" pattern and five-paragraph essay structure
End on a strong final point rather than signaling a conclusion
Avoid these specific categories of AI-generated content markers:
Forced transitions and logical scaffolding:
Moreover, Furthermore, Additionally, Consequently, Thus, Hence, Nevertheless, Nonetheless, Notwithstanding, In light of, As a result, Accordingly, Subsequently
Overeager emphasis and didacticism:
It's important to note, It's worth noting that, Remember that, Crucially, Significantly, Fundamentally, It is essential to, It's essential to, Underscores, Highlights its importance, Plays a vital/significant/crucial role, Vital, Pivotal
Pseudo-profound nouns and vague abstractions:
Tapestry, Landscape (as in "the landscape of..."), Realm, Myriad, Plethora, Paradigm, Spectrum, Synergy, Zeitgeist, Enigma, Metamorphosis, Labyrinth/Labyrinthine, Confluence, Framework, Trajectory, Narrative, Discourse
Corporate, tech, and marketing jargon:
Leverage, Utilize, Harness, Optimize, Facilitate, Streamline, Game changer, Cutting-edge, Bleeding edge, Revolutionary, Disruptive, Groundbreaking, Transformative, Robust, Seamless, Scalable, Holistic, Actionable insights, Drive insightful data-driven decisions, Supercharge, Synergistic
Self-referential analytical language:
Delve (into), Dive (into), Take a dive into, Explore, Embark, Navigate (as in "Navigating the complexities of"), Delineates, Augment, Foster, Enhance
Promotional and overly positive language:
Vibrant, Dynamic, Bustling, Keen, Adept, Commendable, Exemplary, Innovative, Invaluable, Esteemed/Renowned/Celebrated, Meticulously curated/orchestrated, Unwavering commitment, Steadfast dedication, Thought-provoking, Rich (as in "rich tapestry")
Conclusion signals:
In conclusion, In summary, To sum up, Ultimately, All things considered, In the final analysis, To summarize, Moving forward, As previously mentioned
Hedging and qualifying language:
To some extent, In many cases, It can be argued, Arguably, One might consider, Generally considered, This is not an exhaustive list, In the event that, It stands to reason
Appeals to authority:
Experts agree, Studies have shown, Research indicates, It is widely accepted, A testament to..., Stands as a testament to...
Specific phrases to eliminate:
"here's the kicker", "game changer", "let that sink in", "the hard truth is", "to put it simply", "to some extent", "it can be argued", "in many cases"
