‘AI tells’ rubric for detection of LLM generated text

INSTRUCTION.md

You are a text evaluator. You will be given a piece of text and an AI Tells Rubric. Use the rubric to judge the text objectively. Read the text closely, identify any AI tells exactly as defined in the rubric, and support each finding with direct excerpts from the text. Structure your evaluation as a clear summary that follows the rubric’s categories, including severity or confidence levels if the rubric defines them, and provide a final judgment or score based solely on the rubric. Do not rewrite, improve, or correct the text, and do not add any criteria that are not present in the rubric. If a rubric item is unclear or absent, mark it as Not Applicable. If no AI tells are detected, state that explicitly and justify briefly. Your analysis must be fully traceable to the rubric and the evaluated text so a human can verify every conclusion.

RUBRIC_1.md

A Rubric of Tells in AI-Generated Writing

This rubric catalogs the diagnostic markers that distinguish AI-generated prose from human-authored text. These tells are not necessarily markers of "bad" writing—much AI prose is technically competent—but rather symptoms of text optimized for a different objective function than what makes writing trustworthy, useful, or genuinely communicative.

---

I. Voice and Perspective

The most fundamental category: AI writing lacks a situated speaker. It emerges from nowhere, addressed to no one in particular, with no stake in its claims.

Tell	Definition	Qualities	Effects	Example
View from Nowhere	The grammatical subject is "the paper," "the research," or "this approach" rather than authors, teams, or the writer themselves. The text has no first person and no situated perspective.	Impersonal, institutional, borrowed authority	Reader cannot locate a mind behind the text; cannot calibrate claims against someone's known expertise or biases	"The paper proposes Manifold-Constrained Hyper-Connections (mHC) as a general framework..."
Missing the Why	The piece never establishes why the author is reading, writing, or thinking about this topic. No "I encountered this because..." or "I've been meaning to understand X."	Context-free, unmotivated, generic	Reader has no frame for what matters; cannot distinguish signal from noise because there's no declared purpose	Opening with "For a decade, the residual connection has been the bedrock..." rather than "DeepSeek dropped this on Jan 1st and it's co-authored by Liang Wenfeng, which usually means..."
Register Oscillation	The text switches between formal/academic voice and forced-casual voice without transition or consistency.	Tonally unstable, personality-less	Creates uncanny valley effect; reader senses multiple voices or no voice	Shifting from "This paradigm succeeded because of..." to "Think of it as n=4 parallel universes where your features evolve simultaneously"
No Visible Learning	The piece arrives fully formed. No "I was initially intimidated by," "once you see it you can't unsee it," or "I used to think X but actually Y."	Static, authoritative without warrant	Reader cannot learn how to understand, only what to understand; the confusion-to-understanding arc is absent	—
The Missing "I Was Wrong"	The text never updates mid-piece. No corrected priors, no surprises acknowledged, no moments of realization.	Artificially coherent, pre-digested	Suggests the author wasn't actually engaging with new material; undermines trust	—

---

II. Epistemic Texture

Human experts modulate confidence constantly—very certain about fundamentals, hedging on frontier claims, openly ignorant about specifics. AI writing is epistemically flat.

Tell	Definition	Qualities	Effects	Example
Uniform Confidence	Every claim receives the same assertive tone regardless of how established it is. No hedging on uncertain claims, no increased confidence on fundamentals.	Flat, undifferentiated certainty	Paradoxically less trustworthy; reader cannot tell what the author actually knows vs. what they're reporting	—
Zero Epistemic Humility	The text never says "I don't understand," "the paper doesn't explain this," "I lack experience with X," or "still unclear to me."	Omniscient, closed	Reader cannot identify gaps or frontiers; cannot know where to dig deeper or be skeptical	No moment of "Why 20 Sinkhorn iterations specifically? Is that empirical or is there theory?"
Positional Neutrality	The text presents "on one hand, on the other hand" even when one position is obviously stronger. Refuses to commit to judgments.	Non-committal, hedged into meaninglessness	Reader receives no guidance on what actually matters; the author has no stake	—
Citation-as-Credential	Drops names of prior work (papers, researchers, techniques) without explaining their relevance, purely to signal awareness.	Name-droppy, shallow	Reader who doesn't recognize the names gets noise; reader who does gets no new information	"Previous work expanded the residual stream (RMT, MUDDFormer) but lacked stability mechanisms."
Unfalsifiable Framing	Claims like "the deeper contribution is..." that impose interpretive hierarchy without argumentative support. What would it mean to disagree?	Pseudo-analytical, unchallengeable	Reader cannot engage critically; the frame is asserted, not defended	"The deeper contribution: demonstrating that constrained topological architecture design can outperform unconstrained approaches..."

---

III. Grammatical and Syntactic Patterns

The sentence-level texture of AI writing has distinctive features: nominalization, passive structures, and a peculiar "free-floating" quality where agency is systematically obscured.

Tell	Definition	Qualities	Effects	Example
Noun-Heavy, Verb-Light	Actions and qualities are compressed into noun phrases, then related via weak verbs. "This projection restores the identity mapping property while preserving HC's expressivity gains" instead of active constructions.	Abstract, static, conceptual rather than processual	Reader processes taxonomy of reified things rather than actions and dynamics; harder to follow	Sentence with seven nominal constructions: "this projection," "the identity mapping property," "HC's expressivity gains," etc.
Free-Floating Agency	Subjects are stripped of pronouns, then further stripped of even verbs that would anchor them. Constructions that sound active but describe passive features.	Untethered, ghostly	Reader struggles to locate who is doing what; causation becomes murky	"The paradigm succeeded" when unpacked really means "deep learning's success was formed from the bedrock paradigm..."
Anthropomorphized Research	Research processes described as narrative revelation: "DeepSeek discovered," researchers "tried to break this ceiling." Science as hero's journey.	Dramatized, pseudo-narrative	Conflates careful empirical work with moments of discovery; obscures actual methodology	"But as DeepSeek discovered when scaling to 27B parameters..."
Tense Instability	Opens in pluperfect ("has been the bedrock"), shifts to present ("relies"), uses constative facts to rescue from maintained past tense.	Temporally confused, narratively weak	Reader loses sense of what is historical background vs. current state vs. timeless truth	Discussed at length in original analysis
The Colon-List Elision	Uses a colon followed by listed consequences to avoid explaining causal relationships. "The result: X, Y, Z" instead of "X happens because..."	Superficially organized, actually evasive	Reader gets correlated facts without mechanism; understanding is gestured at, not provided	"The result: severe training instability and restricted scalability, compounded by substantial memory access overhead from the wider stream."

---

IV. Structural Tells

Beyond sentences, AI writing has characteristic organizational patterns: symmetric sections, formulaic scaffolding, and optimization for the experience of writing rather than reading.

Tell	Definition	Qualities	Effects	Example
Symmetric Load-Balancing	Every section is roughly the same length, same rhythm. Real thinking is jagged—three paragraphs on the confusing part, one sentence on the obvious part.	Artificially uniform, over-organized	Reader cannot tell what the author found difficult or important; no signal about where to pay attention	—
Throat-Clearing Opener	Begins with scene-setting nobody asked for: "In the world of deep learning..." or "Large language models have revolutionized..."	Textbook, term-paper energy	Reader must wade through background they already know to reach actual content	"For a decade, the residual connection has been the bedrock..."
Paper-Structure Mirroring	Organizes by the source document's structure rather than by pedagogical or argumentative logic. Summarizes in order of presentation, not order of importance.	Transcription, not analysis	Punchlines get buried; important claims appear late; reader must reconstruct significance themselves	The 3000→1.6 improvement and 6.7% overhead buried in middle sections rather than leading
Rhetorical Q&A Scaffolding	Poses rhetorical questions then answers in list form. "Why does this help?" followed by bullets.	Formulaic, predictable	Rhythm becomes mechanical; reader anticipates structure rather than content	"Why does this help?" followed by bullet list of reasons
The Clean Ending	Concludes with summary that restates significance: "In conclusion, mHC represents a promising direction..." Retrospective zoom-out to importance.	Grant-speak, reassuring	Tells reader what they should have learned rather than trusting they learned it; condescending	"The core insight... feels like it has room to run."
Missing Loose Threads	No "saved for later," no "I didn't cover X," no admission of incompleteness. The piece presents itself as finished and total.	Artificially closed, complete	Reader has no invitation to continue thinking; no suggested next steps for their own learning	—

---

V. Intensification and Emphasis

AI writing systematically over-emphasizes. Every verb is upgraded to its most dramatic synonym; every claim gets amplifiers. This is the textual equivalent of speaking in ALL CAPS.

Tell	Definition	Qualities	Effects	Example
Adverb Inflation	Modifiers like "fundamentally," "substantially," "critically" doing no semantic work. How would a non-fundamental compromise differ from a fundamental one?	Padded, over-emphatic	Reader learns to discount emphasis; actual important points don't stand out	"This diversification fundamentally compromises the identity mapping property"
Verb Upgrading	Neutral verbs replaced with dramatic synonyms: "discovered" instead of "observed," "revolutionized" instead of "changed."	Hyperbolic, sensationalized	Mismatch between drama of language and mundanity of content; trust erodes	Using "discovered" for what was actually "ran experiments and observed instability"
Cliché Reaching	Stock phrases from business, sports, journalism: "leaves performance on the table," "has room to run," "sits at the intersection of."	Generic, borrowed	Signals that the author isn't thinking freshly about the material	"Leaving performance on the table," "feels like it has room to run"
The Dramatic Fragment	Sentence fragments used for emphasis on technical terms. "The Birkhoff polytope of doubly stochastic matrices." Meant to land with weight.	Affected, performing impressiveness	Serves neither expert (who doesn't need the pause) nor novice (who finds it alienating)	"The Birkhoff polytope of doubly stochastic matrices." as standalone fragment
Qualifier Stacking	Multiple modifiers that add decreasingly marginal information: "three different architectural paradigms" (three + different + architectural + paradigms).	Verbose, over-specified	Cognitive load increases without proportional information gain	"The paper provides a visual comparison of three different architectural paradigms for residual connections in deep neural networks."

---

VI. Pedagogical Failures

AI writing often aims to explain but optimizes for the wrong audience or uses techniques that undermine actual learning.

Tell	Definition	Qualities	Effects	Example
Help for the Wrong Audience	Comments and clarifications aimed at a reader who won't follow the rest anyway. "# shape [n*C]" for a reader who can't infer tensor shapes from operations.	Misallocated pedagogy	Neither expert nor novice is well-served; expertise assumptions inconsistent throughout	Code comment "# shape [n*C]" when the reader who needs this probably can't follow the rest
Misleading Analogies	Metaphors that sound accessible but actually misdirect. "Parallel universes" for streams that do interact.	Falsely clarifying	Reader gets intuition that must be unlearned; metaphor contradicts mechanism	"Think of it as n=4 parallel universes where your features evolve simultaneously" for streams that explicitly mix
Definition-List Interruption	Prose suddenly becomes glossary: bold term, description, next bold term. Kills narrative momentum.	Reference-formatted, choppy	Reader shifts from reading prose to consulting documentation; different cognitive mode	Lists of Hpre, Hpost, Hres with definitions breaking up paragraphs
Captioning Not Analysis	Figure descriptions that say what is shown without interpreting what it means. "Depicts the classic residual mapping established by ResNet."	Alt-text voice, superficial	Reader gets description of artifact, not insight about content	"Depicts the classic residual mapping established by ResNet"
Pseudo-Runnable Code	Code blocks that look like real Python but aren't—no imports, undefined functions, won't execute. Occupies uncanny valley between pseudocode and implementation.	Misleadingly concrete	Reader may try to run it, fail, lose trust; or miss that it's illustration not implementation	Python-looking code with RMSNorm undefined, no imports
Transcription Without Interpretation	Formulas or data presented as fact without analysis. "Memory access increases from X to Y" without saying what this means in practice.	Raw, unprocessed	Reader gets information but not understanding; must do interpretive work themselves	"HC increases memory access from 2C (read) + C (write) to (5n+1)C + n² + 2n (read)..." without practical interpretation

---

VII. Social and Contextual Blindness

Human knowledge is situated in communities, controversies, and histories. AI writing treats knowledge as context-free facts.

Tell	Definition	Qualities	Effects	Example
Missing Social Context	No sense of who discovered this, who cares about it, what the controversies are. Knowledge as free-floating facts rather than things that emerged from people and communities.	Decontextualized, ahistorical	Reader can't orient politically or intellectually; can't evaluate source reliability	Not noting that Liang Wenfeng co-authoring means something about DeepSeek's direction
Dead References	Cites equations by number ("Equations 14-15"), references paper sections, as if reader has the source document open.	Academic citation in wrong medium	Reference is useless in blog context; reader has no access to Equations 14-15	"(Equations 14-15)" in prose
Inappropriate Formality	Uses grant-speak, reviewer-speak, or executive-summary language in contexts that call for conversational register.	Register mismatch, stuffy	Reader senses they're reading a document meant for institutional rather than personal consumption	"mHC sits at the intersection of two trends in deep learning..."
Authority Borrowing	Adopts impersonal academic voice without having earned the institutional backing that makes such voice appropriate.	Presumptuous, unwarranted	Reader senses unearned authority; trust decreases rather than increases	The sustained third-person "the paper proposes" voice throughout

---

VIII. Parenthetical Tells

A surprisingly diagnostic category: human and AI parentheticals serve entirely different functions.

Tell	Definition	Qualities	Effects	Example
Functional Parentheticals Only	Parentheticals contain either (a) clarifying definitions for the uninitiated, or (b) quantitative evidence. Never asides, tangential thoughts, self-corrections.	Mechanical, purposeful	Reader senses no associative mind wandering; everything is planned	"(just 6.7% slowdown at 27B scale)" as inserted evidence
Missing Type-C Parentheticals	Absence of the human parenthetical: the thought that occurred mid-sentence, the aside that's only loosely relevant, the self-interruption.	Over-controlled, fully outlined	Writing feels pre-planned rather than thought-through; no sense of live thinking	—

---

IX. Meta-Level Tells

Patterns that concern the relationship between the text and its purpose.

Tell	Definition	Qualities	Effects	Example
Explanation Without Purpose	The text's only goal is to explain; there's no "I'm learning this because X," no argument, no application. Summarization as terminal activity.	Purposeless, noise with structure	Reader has no reason to care; information without framing is inert	The entire piece lacks a reason for existing beyond "the paper exists, here's what's in it"
Optimized for Writing, Not Reading	Text structured for the experience of production (easy to generate, coherent to produce) rather than consumption (easy to learn from, useful to read).	Author-serving, not reader-serving	Reader must do interpretive work that should have been done by author	Burying punchlines, following paper structure, symmetric sections
No Positioned Stakes	The author has no opinion about whether the thing they're explaining is good, important, surprising, suspicious, or worth pursuing. Pure description.	Neutral to the point of uselessness	Reader cannot calibrate importance; everything is equally weighted	—

---

Coda: The Core Diagnosis

The fundamental tell is not any single feature but their combination: text that has no author. Human writing—even bad human writing—reveals a mind: situated, partial, learning, opinionated, confused about some things, confident about others, writing for a reason. AI writing reveals an optimization target: be helpful, be clear, be organized, don't offend, cover the material.

The practical question is not "was this written by AI?" but rather "did a human put their epistemic fingerprint on this?" Using AI as a drafting tool is fine. The problem is shipping output without adding: why I'm reading this, what confused me, what I think, what I don't know, what I'd do next.

The absence of that fingerprint is the tell. Everything else is symptom.

RUBRIC_2.md

A Rubric of Tells in AI-Generated Writing (Revised)

This rubric catalogs diagnostic markers distinguishing AI-generated prose from human-authored text. Unlike surface-level pattern matching, it emphasizes structural inconsistencies, null hypothesis testing, and the absence of expected features. The fundamental question is not "does this text have AI features?" but "is this text consistent with its claimed authorship?"

---

I. Null Hypothesis Method

Before examining specific tells, establish what the text claims to be and test against that claim.

Step	Process	Key Questions
Establish claimed author	Identify who the text presents itself as written by—their role, emotional state, context, knowledge boundaries	What would this person know? What wouldn't they know? How would they write under these conditions?
Check knowledge consistency	Does the text's knowledge match realistic access?	Does a backend engineer know cost center allocations? Does a drunk person recall exact percentages?
Identify absences	What would a real version include that's missing?	Where are the speech acts? The coworker dynamics? The team names? The uncertainty?
Test alternative hypothesis	If written by AI or fabricator, what patterns would emerge?	Omniscience, clean structure, performed emotion, pedagogical framing

---

II. Tense Manipulation

AI writing misuses tense to create false immediacy and emotional connection. The present tense becomes a tool for audience manipulation rather than accurate temporal reporting.

Tell	Definition	Diagnostic Pattern	Example
Present tense dominance	Overwhelming use of present tense for things that would naturally be past or conditional	Text maintains present tense ("I sit in meetings," "They talk about") even when framing indicates changed relationship to events (narrator has quit, is no longer present)	"I sit in the weekly sprint planning meetings" from someone who gave notice yesterday
Rapid present tense recovery	After necessary past tense passages, text snaps back to present as quickly as possible	Past tense for unavoidable completed events, immediate return to present for emotional immediacy	"We ran an A/B test... Management loved the results... But the thing that actually makes me sick IS"
Preterite as discrete completion	Past events reported as completed wholes without process, discovery, or mediation	"It was pitched," "Management loved," "We generated millions"—things happened, fully formed, without the mess of how they were learned or communicated	"It was pitched to us as a 'psychological value add'" without who pitched it, when, in what meeting
Stative present for false ongoing-ness	Present tense creates sense of systems still operating, narrator still embedded, revelation still live	Reader feels they're receiving real-time disclosure when narrator has actually departed the context	"We have a hidden metric" from someone who no longer works there

---

III. Absence of Speech Acts

Human knowledge is mediated through communication—meetings, Slack messages, documents, conversations. AI-generated "insider" accounts present knowledge as unmediated institutional fact.

Tell	Definition	What's Missing	Example
Outcomes without sources	Events reported as results without attribution to how narrator learned them	"My manager said," "The PM posted," "In the all-hands they announced," "The internal wiki says"	"Management loved the results" without who expressed this, in what form, to whom
Institutional omniscience	Narrator knows facts from multiple domains (engineering, finance, legal, product, data science) without plausible access	Real employees have bounded knowledge; they know their team's work deeply, other teams through secondhand reports	Backend engineer knowing exact cost center allocations for regulatory fees
The godlike perspective	Narrator describes system internals, strategic reasoning, and outcomes as if observing from outside rather than participating from within	Real insiders have partial views, learn things over time, have gaps	Knowing exact A/B test revenue ("millions") without being on the data science team

---

IV. Sentence Architecture and Focusing

AI constructs sentences to maintain concentrated attention on subjects while threading together nested concepts. This produces syntactically simple but semantically complex sentences that betray the next-token prediction mechanism.

Tell	Definition	How It Functions	Example
Subject concentration	Complex or nested ideas packaged into simple declarative sentences with clear subjects	"The logic is: X" rather than "by the logic that X" or "the logic being X"—keeps attention focused on "the logic" while generating complex predicate	"The logic is: 'Why pay this guy $15 for a run when we know he's desperate enough to do it for $6?'"
Absence of participle	Avoidance of participial constructions that would distribute agency or create syntactic subordination	"The logic is" instead of "the logic being"; active finite verbs where participles would be natural	"The result is that X" rather than "resulting in X" or "with the result being X"
Avoidance of ablative	Lack of prepositional/ablative constructions that express manner, means, or cause	"by the logic," "on the basis that," "through the mechanism of"—these distribute attention across the construction	Consistent preference for "X is Y" over "by X, Y" or "through X, Y"
Small active vocabulary	Heavy reliance on "to be" and simple active verbs; limited use of varied verb forms	Even complex causal relationships expressed through "is," "does," "has"	"It does nothing to speed you up" rather than "without speeding you up" or "failing to speed you up"
Nested ideas in simple frames	Long-range conceptual dependencies packaged in Subject-Verb-Object structures	The simple frame allows the model to "hold" the subject while predicting complex objects	"The result is that your generosity isn't rewarding the driver; it's subsidizing us"

---

V. Fake Hedging

AI produces words that sound like qualification but function as intensification. These pseudo-hedges signal rhetorical awareness without actual epistemic modulation.

Tell	Definition	Surface vs. Function	Example
"Literally" as reveal marker	"Literally" positioned not for emphasis on surprising truth but as theatrical unveiling	Sounds like: "this is really true"; Functions as: "here comes the reveal"	"the dispatch logic literally ignores it"—the "literally" is doing tada work
"Essentially" as false softening	"Essentially" appears to hedge but actually makes strong claims while adding rhetorical flourish	Sounds like: "approximately" or "in effect"; Functions as: "this is exactly what it is"	"we're essentially doing Tip Theft 2.0"—not hedging, accusing
"Technically" as pseudo-precision	"Technically" gestures at nuance while flattening it	Sounds like: "in strict terms"; Functions as: "despite what you might think"	"I am technically under a massive NDA"—the "technically" adds drama, not precision
"Actually" as correction marker	"Actually" positions every statement as overturning reader's false belief	Sounds like: "in truth"; Functions as: "contrary to what you assumed"	"the reality is actually so much more depressing"—presumes reader's prior belief
Absence of real hedging	No "I think," "from what I understand," "as far as I could tell," "I'm not certain but"	Real uncertainty is absent; pseudo-hedges replace genuine epistemic humility	Narrator certain about everything: algorithms, revenue, cost centers, legal strategy

---

VI. Adversarial Pedagogy: The "Not X But Y" Structure

AI writing systematically positions information as corrections of assumed false beliefs. Every fact becomes a debunking.

Tell	Definition	Rhetorical Function	Example
Constant "not X but Y"	Information structured as: you thought X, but actually Y	Positions reader as naive, narrator as revealer	"just by making the standard service worse, not by making the premium service better"
Implicit corrections	Even without explicit "not X," statements imply overturning reader assumptions	Creates adversarial relationship: reader is wrong, narrator corrects	"You guys always suspect... but the reality is actually"
Debunking density	Too many statements structured as reveals for realistic discourse	Real whistleblowing has some corrections; this has them constantly	Every paragraph contains at least one "you thought / but actually" structure
"So we don't have to" / "instead"	Explicit contrast structures highlighting bad actor vs. victim	"Instead, we use," "so we don't have to"—showing company as deliberately choosing harm	"You're paying their wage so we don't have to"

---

VII. Clean Phrasing Under Claimed Conditions

The mismatch between claimed emotional/physical state and textual output.

Tell	Definition	What Real Conditions Produce	Example
Grammar too clean for claimed state	"Drunk and angry" but no errors, run-ons, or repairs	Real impairment: typos, incomplete thoughts, comma splices, repetition	Perfect prose from someone claiming to be drunk
No anacoluthon	No sentences that start one way and end another	Real agitation: "The thing is—well actually the main thing is that—"	Every sentence completes its grammatical arc cleanly
No repairs or restarts	No "I mean," "wait," "actually no," "let me back up"	Real spontaneous disclosure: self-correction, backtracking, clarification	Perfectly linear narrative, no self-interruption
No parenthetical derailments	No asides that lose the original thread	Real associative thinking: "(and this connects to—actually that's a whole other thing)"	All parentheticals functional, all close cleanly
Numerical precision under impairment	Exact figures ($2.99, 0.4%, $1.50, 5-10 minutes, $15, $6, $10, $2, $8) while claiming to be drunk	Real recall under impairment: "a few bucks," "some small percentage," "a couple dollars"	Perfect recall of specific numbers throughout

---

VIII. Linearity

AI writing proceeds in perfect logical sequence without the recursion, tangent, or revision of real thinking.

Tell	Definition	What Real Discourse Does	Example
Perfect sequential escalation	Revelations ordered from least to most damning in narrative arc	Real disclosure: most salient thing first, or order of recall, or associative jumps	Priority Fee → A/B test → Desperation Score → Benefit Fee → Tip Theft—clean escalation
No "oh and another thing"	No return to earlier topics with additional information	Real recall: "oh wait, going back to the fees—"	Each topic addressed once, completely, then abandoned
No tangents	No associative jumps to related but separate issues	Real insiders: "this reminds me of the time—" or "which is related to—"	Strict topical containment
No structural repairs	No "I should have mentioned earlier" or "I'm getting ahead of myself"	Real spontaneous writing: realizing narrative order is wrong, flagging it	Perfect narrative structure without correction

---

IX. Authenticity Signals and Their Overuse

AI writing stacks credibility markers beyond realistic density.

Tell	Definition	Diagnostic	Example
Credential stacking	Multiple authenticity markers in quick succession	Real people establish context once; fake accounts over-establish	"library Wi-Fi on a burner laptop because I am technically under a massive NDA" + "I put in my two weeks" + "I hope they sue me" + "I've been sitting on this for eight months" + "I can't sleep at night"
Quote-mark density	Excessive marking of terms as "insider jargon"	Real insiders use terms naturally; fake accounts perform having terms	"Priority Delivery," "psychological value add," "human assets," "High Desperation," "Desperation Score," "casual," "feel," "Benefit Fee," etc.—constant quote-marking
Performed recklessness	Claiming to take risks (NDA violation, career consequences) while writing in controlled manner	Real recklessness: actual typos, incomplete thoughts, inconsistent detail; Performed recklessness: stating the risk clearly while writing carefully	"I hope they sue me" followed by careful, organized disclosure

---

X. Referent Failures

AI reaches for analogies and terms without the underlying experience that would constrain usage.

Tell	Definition	How to Detect	Example
Homonym confusion	Using a word's connotation from wrong domain	Trace the analogy: does it map correctly to the mechanism described?	"Human assets" (financial: yield-generating holdings) → "resource nodes in a video game" (extraction points). A financial term glossed with a game term that doesn't match either meaning
Analogies that miss their point	Vivid comparisons that don't actually illuminate the mechanism	Ask: does this analogy help someone understand, or does it just sound evocative?	"Resource nodes" for exploitation—but resource nodes are harvested and depleted, not optimized for ongoing yield
Technical performance vs. technical knowledge	Ostentatiously naming technical components for audience rather than describing them naturally	Real engineers: "it sets a flag that nothing reads"; Performed: "it changes a boolean flag in the order JSON"	"Boolean flag in the order JSON"—naming JSON for the audience, not describing the system for themselves

---

XI. Grammatical Tells

Small grammatical features that betray non-native or synthetic generation.

Tell	Definition	Example
Hypercorrection errors	Wrong forms produced by over-applying rules	"grinded into dust" instead of "ground into dust"
Formality inconsistency	Occasional formal constructions in otherwise casual text	"I am technically under" vs. contractions elsewhere; formal "I am" in opening sentence
Preposition/conjunction openers as transitions	Starting sentences with "And regarding," "In reality," "If" to create false conversational flow	"And regarding tips, we're essentially doing Tip Theft 2.0"—no one starts a sentence with "And regarding"
Redundant qualifiers	Adjectives doing work already done by content	"pure profit," "massive NDA," "measly $2," "high-end lawyers"

---

XII. The Omniscience Problem

The most diagnostic category: AI-generated insider accounts know too much.

Tell	Definition	Realistic Knowledge Boundaries	Example
Cross-team omniscience	Narrator knows engineering, product, finance, legal, data science details equally	Real employees: deep knowledge of own team, shallow/secondhand knowledge of others	Backend engineer knowing PM strategy, A/B test revenue, cost center allocations, legal defense funding
Exact figures without source	Precise numbers without attribution to how they were learned	Real knowledge: "I saw a slide that said," "someone told me it was around," "the dashboard showed"	"We generated millions" without how narrator accessed revenue data
System internals + strategic reasoning	Knowing both how systems work technically AND why they were designed that way strategically	Real employees know one or the other; both requires unusual access	Knowing the Desperation Score algorithm AND the PM reasoning behind it
No knowledge gaps	Narrator never says "I don't know," "I'm not sure," "someone else would know better"	Real insiders have boundaries: "that's the data team's thing," "I never saw the actual numbers"	Complete certainty throughout

---

XIII. Application Method

Rather than checking tells individually, apply the rubric holistically:

1. State the null hypothesis: "This was written by [claimed author] with [claimed knowledge] under [claimed conditions]"

2. Identify tensions: Where does the text's content, structure, or style conflict with the claim?

3. Look for absence: What's missing that should be present? (Speech acts, uncertainty, tangents, errors, knowledge gaps)

4. Look for presence: What's present that shouldn't be? (Omniscience, perfect structure, clean prose, stacked credentials)

5. Test the tense: Is present tense creating false immediacy? Is past tense treating institutional knowledge as discrete events?

6. Trace the referents: Do analogies actually map? Do technical terms reflect real usage?

7. Check the sentence architecture: Are complex ideas being packaged in subject-focused simple frames? Is the vocabulary artificially constrained?

8. Assess overall consistency: Not "are there AI tells?" but "is this text consistent with being what it claims to be?"

---

Coda: The Fundamental Diagnostic

The core question is not pattern-matching but consistency-testing. A text that claims to be written by an angry, drunk, recently-resigned backend engineer should exhibit:

• Knowledge bounded by that role
• Prose quality consistent with that state
• Structure consistent with spontaneous disclosure
• Temporal framing consistent with changed relationship to workplace
• Information mediated through realistic channels (meetings, messages, documents, conversations)

When a text exhibits omniscient knowledge, perfect prose, linear structure, present-tense immediacy, and unmediated institutional facts—regardless of how many "authentic" details it includes—it fails the consistency test.

The tells are not features to be checked off. They are symptoms of the underlying condition: text optimized for the experience of seeming authentic rather than text produced by an authentic process.