Never Sound Like AI
Your clients can smell machine-generated text in two seconds.
The Architecture Series · 8 minute read · By Ed | Founder & CEO, DealiOS
You can spot AI-written text in two seconds. Not because of any single word. Because of the absence of texture.
Human writing has rhythm irregularities. Sentence fragments. Odd word choices that reveal a real person made a real decision. AI writing has an eerie uniformity, like a neighborhood where every lawn is the same height. Nothing is wrong. Everything is wrong.
The uncanny valley isn't just for faces. It lives in your proposals, your contracts, your executive summaries. And your clients can feel it even when they can't name it.
The Uniformity Problem
In a 2025 study from JMIR Medical Education, blinded experts were asked to distinguish AI-generated scientific texts from human-written ones. They identified the AI text correctly 70% of the time. [1] Not because the AI made errors. Because the AI made nothing. No stumbles, no recoveries, no choices that felt like a person choosing. The experts cited three tells: redundancy, repetition, and a strange uniformity in how ideas connected to each other.
A separate analysis published in Scientific Reports confirmed this from the computational side: AI text relies heavily on structural connectors and filler words, while human writers use content-rich, contextually grounded vocabulary. [2] The machine produces grammatically flawless sentences. The human produces sentences that feel like someone was in the room.
This is what researchers in psycholinguistics call the absence of experiential grounding. Human writing reflects episodic memory, thematic coherence, and personal narrative strategies. AI writing mimics these surface-level patterns but lacks the underlying cognitive source. The result, according to a 2025 study connecting stylometric features to cognitive theory, is that AI text shows flat emotional expression and overly uniform readability, while human writing adapts complexity based on audience and intention. [3]
Reading AI text is like talking to someone who has technically correct answers to every question but has never experienced anything. They know the words for "frustrated" and "excited" and "exhausted." They've just never been any of those things. You can tell. You can always tell.
Human writing stumbles. It recovers. It takes chances. AI writing glides. Gliding is how you spot the robot.
AI doesn't get caught because it's bad. It gets caught because it's too perfect. The perfection is the tell.
The linguistic term for this is "burstiness." Human writers vary their sentence complexity dramatically. One short punchy sentence. Then a 40-word monster with nested clauses and a parenthetical aside. Then another short one. AI writers produce sentences of remarkably similar complexity throughout. Plot the sentence-level difficulty on a graph: humans produce spikes and valleys, AI produces a flat line. The flat line is the fingerprint.
The Lexical Fingerprint
There are specific words that flag AI origin. Not wrong words. Overused words. "Comprehensive." "Robust." "Leverage." "Delve." "Seamless." "Transformative." "Streamline."
The AI reaches for these words because they appeared millions of times in its training data as signals of professional quality. They were the vocabulary of corporate blogs, marketing white papers, and consulting decks for two decades. The model learned that when humans want to sound professional, these are the words they use. So when the model tries to sound professional, these are the words it uses. Every time. In every document. Until every AI document sounds like the same LinkedIn post from 2023.
The fix isn't a casual aside in the system prompt. "Please avoid corporate jargon" is the first thing the model forgets when it's 3,000 tokens into generating a 15-page proposal. The fix is architectural: a banned-words table with specific alternatives, structurally embedded in every system prompt as detailed linguistic rules. Not a suggestion. A formatted table mapping each banned word to its replacement, injected into every AI call regardless of document type. "Comprehensive" doesn't get a gentle nudge. It gets a rule: "Name what's included." "Robust" gets: "State the actual standard." The specificity is what makes it work. The model doesn't have to remember a vague instruction. It has a reference table with exact substitutions sitting in its context window for the entire generation.
Then there are the punctuation tells. Em dashes are the single most common AI punctuation fingerprint. The model uses them to split independent clauses because its training data is full of published writing where editors made that choice deliberately. The model makes the choice reflexively, four or five times per page, until the document reads like a transcription of someone who can't finish a thought.
Ellipses signal trailing off. Be definitive. Never trail off. Professional confidence does not end with three dots. Exclamation marks signal enthusiasm that nobody asked for. Professional confidence does not need exclamation marks either. These punctuation patterns aren't style preferences. They're detection vectors.
A banned-words table isn't a style guide. It's an immune system.
A banned-words table in the prompt isn't a style guide. It's an immune system.
The Document-Type Voice Split
Here's where most AI platforms get it wrong: they treat voice as a single setting. One persona. One tone. One set of rules. Everything comes out sounding the same whether it's a proposal, a contract, or a presentation deck.
That's architecturally lazy. And it's wrong. Because the same word that destroys trust in a proposal builds energy in a presentation.
"Transformative" in a document read by a procurement committee signals "this was written by a machine." The committee has seen that word in every vendor's pitch for three years. It triggers pattern recognition. Not the good kind.
"Transformative" in a slide deck delivered by a confident speaker signals conviction. The speaker owns the word. Their body language, their pacing, their eye contact makes it land. The word isn't the problem. The context is.
This means the architecture needs document-type-specific linguistic profiles. Not templates. Profiles. Behavioral rules that change how the AI writes based on what it's producing.
Proposals: "Sound like a senior sales professional." Measured. Evidence-based. Sentences that earn trust through specificity, not through adjectives.
Presentations: "Sound like a creative director." Punchy. Confident. Permission to use emphasis tools like em dashes (re-allowed for this format specifically) because a speaker's rhythm is different from a reader's rhythm.
Contracts: "Sound like outside counsel who bills by the hour and wastes no words." Precise. Formal. No color. No brand. No decorative language. Every word does legal work or it gets cut.
Three document types. Three linguistic profiles. Three different definitions of what "good writing" means. Treating them as one is how you get contracts that sound like marketing copy and presentations that read like legal briefs.
The discipline is architectural, not instructional. Vague prompt suggestions are probabilistic. Structured linguistic rules with specific alternatives are reliable.
The TrustRadius 2025 B2B Tech Buying report found that content increasingly falls flat when it feels "overly polished" or "disconnected from real-world experiences." [4] B2B buyers aren't asking for better marketing. They're asking for writing that sounds like it came from someone who understands their business. Polished isn't the goal. Appropriate is the goal. And appropriate changes based on what you're writing.
The Structural Fix
Beyond word-level and punctuation-level tells, there's a third layer: structural tells.
AI-generated documents have specific formatting failure modes that are invisible in a single generation but become patterns across thousands. Numbered lists that split across lines when they should flow as prose. Bold headers glued directly to paragraph text with no spacing logic. Orphan bullets that dangle without context. Section transitions that feel abrupt because the AI generated each section independently and bolted them together.
These aren't random bugs. They're the specific failure modes that emerge in production output after thousands of generations, discovered through iterative battle-testing.
The approach is two layers working together. First, detailed structural formatting rules in the prompt: when to use bullets vs. prose, how to structure paragraphs, bold lead-in patterns, page balance rules. These aren't vague instructions. They're specific, formatted guidelines with examples that sit in the AI's context window during generation. Second, a formatting normalizer that runs after generation: a deterministic regex pipeline that catches the structural artifacts the AI still produces despite the prompt rules. Numbered lists where the number ended up on one line and the content on the next. Bold headers glued directly to the previous sentence with no paragraph break. Orphan dashes. Excess blank lines. The normalizer fixes these mechanically, every time, without judgment.
This is separation of concerns applied to AI output. The prompt carries the voice, the identity, the behavioral rules, and the structural guidelines. The normalizer catches the formatting edge cases that slip through. The prompt handles what the AI writes. The normalizer handles how it renders. Each layer does what it's good at. The AI and the linter have different jobs.
Why This Matters Beyond Sales
The linguistics of AI detection aren't unique to sales documents. Researchers have identified the same patterns across academic writing, journalism, legal documents, and medical communications. An MIT Press survey on LLM-generated text detection catalogs the problem across every domain where AI text is replacing human text. [5]
The core finding is consistent: AI text is detectable not through what it says wrong, but through what it doesn't say at all. The absence of experiential anchors. The uniformity of sentence rhythm. The overreliance on structural connectors instead of content-rich vocabulary.
And the detectors are getting better. Stanford HAI's 2025 benchmark methodology tested 15 detection tools and found that ensemble methods now achieve 96% accuracy on AI-generated content, with false positives under 3%. [6] The arms race between generation and detection is accelerating, and detection is winning. If your architecture doesn't address the linguistic fingerprint, your customers won't need a tool to notice. They'll feel it. The way you feel it when someone is reading a script instead of talking to you.
This matters for every company generating customer-facing AI text. Not just proposals. Product descriptions. Customer communications. Marketing content. Support responses. Any artifact where the reader's trust depends on feeling like a person was involved in creating it.
The Principle
Fighting AI detection isn't a prompt problem. It's a systems problem. And the system has three layers.
Layer one: the lexical immune system. A banned-words table structurally embedded in every system prompt, mapping flagged vocabulary to specific alternatives. Not a casual "please avoid jargon" instruction that competes for the model's attention. A formatted reference table with exact substitutions that sits in the context window for the entire generation. "Comprehensive" maps to "Name what's included." "Delve" maps to "examine, explore, cover." The specificity is the architecture. The AI doesn't need to remember a vague rule. It has a lookup table.
Layer two: document-type voice profiles. Behavioral rules that change how the AI writes based on what it's producing. The proposal voice uses measured, evidence-based language with room for personality. The contract voice uses precise, formal language with no decoration. The presentation voice uses punchy, confident language with emphasis tools the other formats prohibit. These aren't style guides. They're prompt-layer behavioral constraints that fundamentally change the AI's output based on artifact type.
Layer three: structural formatting. Detailed formatting rules in the prompt (paragraph structure, bullet patterns, page balance) combined with a deterministic normalizer that catches the formatting artifacts the AI still produces. Numbered lists split across lines. Bold headers stuck to paragraph text. Orphan bullets. Excess whitespace. The prompt tells the AI what good structure looks like. The normalizer catches the cases where the output doesn't match.
All three layers work together because any single layer leaves gaps that the other two cover. Banned words don't fix structural tells. Voice profiles don't fix lexical fingerprints. Structural formatting doesn't change how the AI argues.
The discipline is architectural, not instructional. These aren't suggestions in a system prompt that the model might ignore under pressure. They're detailed linguistic rules with specific examples, alternatives, and exceptions, structurally embedded in every AI call. And the formatting normalizer catches what slips through deterministically, every time, without relying on the model to self-correct.
You wouldn't ship code without a linter. Don't ship AI text without one either.
The takeaway: Fighting AI detection isn't a prompt problem. It's a systems problem requiring three architectural layers. A lexical immune system with specific word-level alternatives, document-type voice profiles that change how the AI writes, and structural formatting rules backed by a deterministic normalizer that catches what slips through.
References
1. Doru, B., et al. "Detecting Artificial Intelligence-Generated Versus Human-Written Medical Student Essays: Semirandomized Controlled Study." JMIR Medical Education, 2025. doi:10.2196/62779
2. Scientific Reports. "Classifying human vs. AI text with machine learning and explainable transformer models." Nature Scientific Reports, 2025. doi:10.1038/s41598-025-27377-z
3. ArXiv. "Distinguishing AI-Generated and Human-Written Text Through Psycholinguistic Analysis." 2025. arXiv:2505.01800
4. TrustRadius. "Bridging the Trust Gap: B2B Tech Buying in the Age of AI." TrustRadius for Vendors, 2025.
5. MIT Press. "A Survey on LLM-Generated Text Detection: Necessity, Methods, and Future Directions." Computational Linguistics, Vol. 51, Issue 1, 2025.
6. Stanford HAI. AI Detection Benchmark 2025. Methodology covering 15 detection tools; ensemble methods achieving 96% accuracy with false positive rate under 3%.
This paper is part of The Architecture Series by Ed, Founder & CEO of DealiOS. For the weekly series and more technical deep-dives, visit dealios.app/library.
See how these principles come together. Explore the platform →