The Math Behind Why Everyone Sounds Like a Corporate Robot (And How to Fix It)

AI writing is everywhere. I write two or three articles a week, and yes, I use AI to scale my thoughts. But I edit. A lot. Enough that it sounds like me, not like every other person outsourcing their thinking to the same large language model.

The problem isn’t that people use AI. The problem is that nobody bothers to make it theirs. All the prose is incredibly identical. What is “prose”? Just continuous writing. Sentences flow into paragraphs without bullet points, without subheadings breaking things up every forty words. The way you’d write if you were thinking out loud on paper.

Scroll through LinkedIn for thirty seconds. Read a few status updates. Look at the proposals crossing your desk, the RFPs, the pitch decks. They all sound identical.

Now this is another long article. Please try to finish the entire read. At the end, there is an anti-AI writing prompt that you can use.

Now lets get into the math of sameness. We will start by taking a few examples from LinkedIn. I am sure you will read these and have a deja-vu moment…

The same person could’ve written every sentence.

Large language models don’t think. They predict. When you prompt an AI, it’s running a probability calculation across billions of text fragments it’s seen during training. For any given word, it’s asking: what word is most likely to come next?

I like to use math as much as possible to explain things that fascinate me. I hope you find it helpful. If not and it’s confusing, just skip over it.

This reads as: “The probability of word n, given all the words that came before it.”

A real example helps. Say the AI has written: “In today’s landscape, businesses must...”

What comes next? The model scans its training data and finds:

• “adapt” appeared after this phrase 47,000 times

• “innovate” appeared 39,000 times

• “leverage” appeared 52,000 times

• “reconsider” appeared 3,400 times

The model assigns probabilities:

P(”leverage”) = 0.31
P(”adapt”) = 0.28
P(”innovate”) = 0.23
P(”reconsider”) = 0.02

It picks “leverage” because it has the highest probability. Not because it means anything. Because it appeared most often in similar contexts during training.

Once “leverage” is chosen, the next word calculation begins:

“In today’s landscape, businesses must leverage...”

New probabilities:

P(”synergies”) = 0.34
P(”technology”) = 0.29
P(”resources”) = 0.21
P(”staplers”) = 0.00003

The model picks “synergies.” Again, highest probability. This is why you see “leverage synergies” everywhere. Not because it’s good writing. Because those two words appeared together thousands of times in the training corpus.

The chain continues:

Each selection is based on what’s most statistically likely given everything before it. The output feels smooth because it’s optimized for predictability. Predictability is the enemy of distinctive voice.

Dissecting Example 1: Where You Can See the Math

“Here’s the thing about innovation: it’s not about having the best idea—it’s about execution.”

Sentence length: 16 words. Right in the sweet spot of 15-20 that dominates most business writing.

Let’s apply our equation to the opening phrase. The model has generated: “Here’s the thing about...”

What comes next?

P(w₆ | “Here’s the thing about”)

The model calculates:

• P(”innovation”) = 0.38

• P(”leadership”) = 0.29

• P(”success”) = 0.24

• P(”staplers”) = 0.001

It picks “innovation” because P = 0.38 is highest.

Now the structure reveals everything:

• “Here’s the thing” = Scaffold phrase with P = 0.43 (appears in 43% of similar openings)

• “it’s not about X—it’s about Y” = Binary contrast structure with P = 0.61

• “innovation” + “execution” = High co-occurrence pair in business corpus

The AI isn’t creating an argument. It’s stacking high-probability templates.

Next sentence: “In today’s landscape, organizations that leverage cross-functional collaboration will outpace those that don’t.”

Length: 14 words. Almost identical to the previous sentence.

Let’s trace the probability chain through this sentence:

P(”In” | [previous sentence]) = 0.42
P(”today’s” | “In”) = 0.51
P(”landscape” | “In today’s”) = 0.48
P(”organizations” | “In today’s landscape”) = 0.37

And so on. Each word is selected based on maximizing:

Phrase probabilities aggregate:

• “In today’s landscape” = P = 0.52 (appears in 52% of business intros)

• “leverage” + “collaboration” = P = 0.38

• “outpace those that don’t” = Comparison closer with P = 0.44

Notice the pattern? Template phrases, stacked sequentially, each with probabilities above 0.35. Word by word, the model is playing statistical averages.

Final sentence: “The signal is clear: adapt or fall behind.”

Length: 8 words. Short for emphasis, which the model learned creates “impact.”

Let’s see the probability calculation for “signal”:

P(”signal” | “The”) = 0.41

Compare this to:

• P(”message”) = 0.18

• P(”answer”) = 0.12

• P(”evidence”) = 0.09

“Signal” is pure AI vocabulary. In human writing before 2018, “signal” appeared in business contexts about 12,000 times per million words. In AI-generated content after 2022? 89,000 times per million. The model absorbed it from its own synthetic training data and now overuses it.

“Adapt or fall behind” is a closing binary with P = 0.67. The model loves binaries because they reduce perplexity (the measure of how “confused” the model is). Lower perplexity = higher confidence = more likely to generate.

Example 2 Breakdown:

“In an increasingly competitive market, businesses face a critical decision.”

Applying our equation to the first phrase:

P(”increasingly” | “In an”) = 0.71
P(”competitive” | “In an increasingly”) = 0.68

• “Increasingly competitive” = Modifier pair with combined P = 0.71

• “face a critical decision” = Setup phrase with P = 0.58

• Sentence length: 11 words

“This is not simply a question of resources—this is a question of strategy.”

Another binary scaffold. AI loves these because they’re symmetrical and resolve tension instantly. Length: 15 words. Em dash used to create false profundity.

The probability chain:

P(”This is not”) × P(”simply”) × P(”a question of”) = 0.59

“Our framework enables stakeholders to navigate complexity and unlock value at scale.”

Every single word in this sentence is jargon with high business-context probability. Let’s trace it:

P(”framework” | “Our”) = 0.81
P(”enables” | “Our framework”) = 0.76
P(”stakeholders” | “Our framework enables”) = 0.63

Continuing through:

• “navigate complexity” = P = 0.59

• “unlock value” = P = 0.72

• “at scale” = P = 0.77

This sentence is statistically inevitable. Five high-probability phrases stacked into one 12-word package. You could generate this sentence by rolling weighted dice.

The Pattern Across All Three Examples:

1. Sentence lengths cluster around 12-16 words (the statistical average)

2. Scaffold phrases appear at high frequency (”Here’s the thing,” “This is not X”)

3. Abstract jargon dominates (leverage, framework, landscape, signal)

4. Binary contrasts structure every argument

5. Em dashes create artificial pauses that feel “thoughtful” but add nothing

The math underneath:

OK, this looks ugly and scary but it’s not. You read this as: “The output equals the sum of the probability of each word given all the words before it, from the first word to the last word.”

In fact, you can also rewrite it as simple arithmetic that looks like this:

Or even simpler:

Total Sentence = (Probability of Word 1) + (Probability of Word 2 given Word 1) + (Probability of Word 3 given Words 1 and 2) + ... and so on

The AI generates each word by maximizing the probability given the prior context. The result reads smoothly because it’s statistically average. “Average” means “like everything else the model has seen.”

That’s why everyone sounds the same.

So what can we do to sound less generic?

I began reverse-engineering the syntax of machine fluency and comparing it to more natural, human writing. And, there are patterns you can interrupt. Below is a prompt I built to break the algorithmic sameness.

Use this before everyone else figures it out and “ungeneric” becomes the new generic:

Anti-AI Writing Prompt:

• Remove formulaic scaffolds and generic transitions (“Here’s the thing,” “This is not X this is Y,” “signal,” “noise,” etc.).

• Avoid em dashes unless grammatically necessary; use commas, semicolons, or sentence breaks instead.

• Vary sentence lengths within each paragraph following a natural, non-repetitive rhythm. Example pattern (not fixed): 12 words → 7 → 8 → 18 → 5 → 15 → 9 → 22 → 6 → 11, then shuffle or skip as needed. Never keep two sentences of identical rhythm back-to-back.

• Maintain 4–7 sentences per paragraph, with visible contrast in length.

  Use short, plain sentences to punctuate longer, information-dense ones.

• Write factually and precisely, keeping tone neutral but human.

  No unnecessary adjectives or imagery unless they clarify meaning.

• Avoid grand conclusions, sweeping morals, or rhetorical closures (“In the end,” “Ultimately,” etc.).

• Favor tangible detail, active verbs, and verifiable data over conceptual phrasing or abstract metaphors.

• Treat “voice” as a product of clarity, not flourish. The reader should feel presence, not performance.

• Accept tension and incompleteness; not every paragraph must resolve.

Copy that into your workflow. Paste it above whatever you’re asking the AI to write. Watch what happens to the rhythm. It won’t fix everything, but it’ll stop your writing from sounding like it came off an assembly line.

At least until everyone starts doing this. Then we’ll need a new trick.

Try the prompt. Hope it helps. Let me know in the comments how it turned out.