Skip to content

AI-ANTIPATTERNS.md

Latest commit

 

History

History
410 lines (348 loc) · 19.8 KB

AI-ANTIPATTERNS.md

File metadata and controls

410 lines (348 loc) · 19.8 KB
Raw

AI Anti-Patterns: the tells that expose AI-built work

Compiled 2026-07-27 from a three-track web investigation (webapp design tells, game tells, writing and code tells); sources at the end. This is the project-agnostic catalog, binding for ALL work in this repo: pages, copy, emails, generated documents, docs, commits, and code. Adopted on 2026-07-27. Where a rule here and a more specific rule in CLAUDE.md or docs/04-UI-DESIGN.md disagree, the more specific one wins. Section 2 covers game surfaces and is kept so this file stays portable; it does not apply here.

Two findings frame everything below:

  1. No single tell is proof. Density and uniformity are the signal. Every source converges on this: one purple button convicts nobody; a page where every card, heading, and sentence follows the statistical default does. Edited AI output with the tells removed is largely indistinguishable, which is the point: remove them.
  2. Tells are what happens when nobody made a decision. Generators emit the statistical average of their training data (the purple gradient traces to Tailwind demo defaults saturating tutorials). A surface that could belong to any product is therefore a defect even when nothing on the lists below appears. The core test users apply takes about 3 seconds: "they all look the same."

Stigma is measured, not theoretical: disclosed AI use on Steam cuts review counts roughly in half; a CHI 2026 study found players rate content worse merely for believing it is AI-made; undisclosed AI assets in a game with any audience are typically spotted within 24 hours.

1. Web and app visual tells

1.1 Color

  • Purple-to-blue / indigo-to-violet gradients; the loudest single tell. Blue-to-pink and cyan-to-purple variants included.
  • The recurring lavender "AI purple" accent (ranked the #2 tell in a 3.2M-post Reddit-mining study).
  • Gradient-filled text: hero headlines, big stat numbers.
  • Default shadcn grays plus Tailwind blue as the entire palette.
  • Neon-on-dark: cyan/violet accents with glowing colored box-shadows.
  • Radial gradient halos and blurred gradient "orbs" floating behind hero content on dark pages.
  • Pure #fff or #000 backgrounds with no tint; or its 2026 successor, the reflex "tasteful" warm cream/beige.
  • Multiple saturated colors competing with no hierarchy.
  • Dark-theme body text that barely scrapes WCAG AA contrast.

1.2 Typography

  • Inter as the unchosen default; the rotating cast of stand-ins: Poppins, Space Grotesk, Geist, Manrope, Roboto, Instrument Serif.
  • The oversized italic-serif accent word inside a sans hero headline.
  • One font family everywhere, no display/body pairing, hierarchy carried by font-size alone.
  • All-caps letter-spaced kicker labels repeated above every section heading.
  • Monospace used decoratively "for the hacker vibe" rather than for code or data.
  • Crushed, over-tight letter-spacing on display type.

1.3 Layout and structure

  • The default SaaS hero: everything centered, eyebrow pill/badge chip above the H1 ("Trusted by 2,800+ teams"), gradient headline, dual CTAs (solid "Start Free" plus ghost "Watch Demo").
  • The fixed page skeleton in the same order every time: hero, three feature cards, logo strip, stats banner, pricing (middle tier highlighted), FAQ accordion, footer.
  • Exactly three feature cards in a row: icon tile, title, two lines of blurb, rounded corners, soft shadow.
  • Numbered "1 / 2 / 3" how-it-works step rows; tiny editorial "01 / 02 / 03" section numbers.
  • Stat banner rows: big number, small label, three supporting stats.
  • Bento grids as the reflex "fancy" option.
  • Cardocalypse: cards nested 3-5 levels deep, everything boxed, each layer with its own padding and shadow.
  • The colored 3-4px left-border strip on cards, cycling colors with no semantic logic; multiple sources call this the single most reliable tell.
  • Hairline 1px border paired with a wide diffuse shadow on the same card; over-rounded 24px+ radii turning small cards into soft blobs.
  • Uniform spacing rhythm: identical section padding, one max-width container, identical card heights, uniform 16px radius everywhere.
  • Permanent dark mode as reflex default; the "Linear-clone" look (dark background, blurred glows, animated gradient artifacts) that fails the squint test between four different companies.
  • Glassmorphism used decoratively rather than to solve a layering problem.

1.4 Components

  • shadcn/ui defaults shipped untouched (rounded-2xl shadow-lg p-6, default theme, default radius); the #1 tell in the Reddit-mining study.
  • Testimonial walls with placeholder avatars and unverifiable names; logo strips with unearned or fake press logos, often auto-scrolling.
  • The universal feature-card icon tile: a small rounded square holding a generic thin-line icon that could illustrate any product.
  • Emoji used as icons in navs, cards, and bullets (ranked #5 tell); the four-pointed sparkle glyph as default AI iconography.
  • Meaningless pulsing/colored status dots mapped to no real state.
  • "Edit with Lovable" / "Built with Bolt" badges; "Built with" plus a heart emoji in the footer.
  • Forms with no validation, error, hover, focus, disabled, empty, or loading states.
  • Default favicon, links pointing to "#", placeholder social links.

1.5 Motion

  • Identical fade-in-on-scroll on every element, same timing, no purpose.
  • Universal hover-lift on cards, stagger on lists, image scale/rotate on hover.
  • Auto-scrolling logo marquees; typewriter/blinking-cursor hero text; aurora and gradient-blob background animations; scroll-jacking; cursor followers.
  • Bounce/elastic easing; or the inverse tell, zero interaction states at all.
  • Ignoring prefers-reduced-motion.

1.6 Imagery

  • AI-generated hero art with plastic smoothness and unnatural symmetry; stock photos of diverse people at laptops.
  • Undraw/Storyset-style flat stock illustration as filler; floating 3D abstract blobs; generic SVG mascot doodles.

2. Game tells

2.1 Art

  • Melted or wrong fine detail: hands, fingers, feet, eyes, teeth; the first thing players check.
  • Gibberish pseudo-text baked into signs, labels, and textures; players screenshot these.
  • Objects merging or floating: straps that go nowhere, weapons fused with hands, accessories detached from bodies.
  • The glossy "AI sheen": waxy poreless skin, over-rendered volumetric glow, hyperdetailed but soulless; cinematic lighting over content that falls apart on second look.
  • Excessive random ornament that means nothing; prompt attributes leaking onto everything in a scene (spillover).
  • Characters that never keep the same face: portrait, sprite, and dialogue bust read as different people; outfit details drift between poses. Only reference-linking holds identity.
  • Style inconsistency across assets: no shared palette, outline treatment, rendering style, or finish level; pixel-density mismatches between sprites and backgrounds; unmodified recognizable asset-store packs (the "asset flip" read).
  • Extreme per-asset contrast with no unified value structure across the scene.
  • Gorgeous AI key art next to visibly cruder gameplay screenshots; the gap itself is the tell.

2.2 Game UI

  • Engine-default UI: Godot default theme, Unity default font and grey buttons; reads as prototype.
  • UI icons in a different style from the game art; icon sets pulled from mixed packs; inconsistent icon sizing and margins; a store-bought UI kit dropped in wholesale.
  • Web/SaaS grammar inside a HUD: cards with drop shadows, pill labels, gradient CTAs.
  • Emoji in game-facing strings as stand-in iconography.

2.3 Game writing

  • Dialogue in the assistant register: over-formal, over-explaining, relentlessly positive, exclamation-heavy barks, "brave adventurer" filler.
  • Lore dumps with no voice that summarize the world at the player instead of speaking in character.
  • Templated item descriptions: every item follows one sentence skeleton (name, epic adjective, vague lore clause).
  • Tutorial copy in the assistant register ("Great job! Now let's...").
  • Plus every prose tell in section 3.

2.4 Audio

  • Generic AI music: predictable progressions, cliche instrumentation, supermarket-playlist polish; fixed songs with audible loop seams and no dynamic layering; mood that does not match the scene.
  • AI voice: flat pitch, no emotional variation, mispronounced names and numbers, uncanny consistency.

2.5 Design and structure

  • Broad but shallow: many systems, none deep; content volume disproportionate to team size is itself a red flag players price in.
  • Procedurally samey or outright broken generated levels (blocked paths, no coherence).
  • Visible runtime-AI latency (NPCs pausing seconds before responding).

2.6 Store page and meta

  • The capsule/header is where players hunt for tells first (fingers, gibberish text, sheen).
  • Missing or evasive AI disclosure where the platform requires it (Steam, itch.io); communities detect undisclosed assets within about 24 hours, and the cover-up outweighs the crime. Disclose honestly; this workspace's policy is honest disclosure with the art_src/ authorship trail as evidence.
  • Trailer/screenshot mismatch; scrubbing AI marketing material after being caught becomes its own story.

3. Writing and copy tells

3.1 Vocabulary (banned in shipped copy)

  • The canonical spikes: delve, tapestry, landscape, realm, synergy, testament ("a testament to"), journey, navigate ("navigating the complex landscape of").
  • Abstract verb inflation: leverage, utilize, harness, streamline, facilitate, foster, enhance, elevate, empower, supercharge, unleash, unlock (as marketing verb).
  • Inflated adjectives: crucial, pivotal, robust, seamless, comprehensive, nuanced, vibrant, profound, nestled, innovative, cutting-edge, game-changing, world-class, best-in-class, enterprise-grade.
  • The formal transition stack at unnatural density: moreover, furthermore, additionally, consequently.
  • Elegant variation: refusing to repeat a word, cycling synonyms; dodging plain "is" with "serves as", "stands as", "functions as", "marks".

3.2 Constructions

  • "It's not just X, it's Y" / "Not X, but Y" contrast-parallelism; the signature construction of this era.
  • The reflexive rule of three ("fast, fun, and friendly") in nearly every list-like sentence; paired-fragment intensifiers ("No ads. No nonsense.").
  • "In today's fast-paced world"; "Let's dive in"; "look no further".
  • Sycophantic openers leaking into copy ("Great question!").
  • The hedging register: "It's worth noting", "Generally speaking", "could potentially", "arguably one of the most", superlatives nobody verified.
  • Recap conclusions ("In conclusion", "Ultimately") that restate the piece; auto-appended FAQ or "Key Takeaways" padding.
  • Sentence-final participle analysis tacked onto facts ("..., highlighting the importance of...", "..., underscoring its role in...").
  • Vague attributions with no named source ("experts argue", "industry reports suggest").
  • Weightless headlines that fit any product ("Build faster. Ship smarter.", "Your all-in-one platform", "Your journey starts here").

3.3 Punctuation and formatting

  • Em dashes (workspace-wide ban; periods or hyphens).
  • Title Case On Every Heading where house style is sentence case.
  • Bold-keyword-per-sentence; bulleted lists where every item is a bolded lead-in. Micro-tell: ending the bold lead-in with a period where a human writes a colon.
  • Emoji-headed sections and emoji-led bullets (workspace-wide emoji ban).
  • Curly/straight quote inconsistency from pasted output; markdown remnants in non-markdown surfaces.
  • Chatbot artifacts pasted verbatim: "As an AI...", "As of my last update...", "Certainly! Here's...", citation tokens ("oaicite", "turn0search0"), utm_source=chatgpt.com in URLs.
  • Every-sentence exclamation marks.

3.4 Structure and tone

  • Uniform paragraph rhythm: same sentence lengths, same paragraph shapes, cadence that never varies; polished but voiceless.
  • Rigid outline structure regardless of content: intro, 3-5 headed sections, conclusion.
  • Everything hedged and nothing claimed: no stake, no opinion, no concrete anecdote; redundant UX writing where label, sublabel, and helper text all say the same thing.
  • Sudden style shift where AI paragraphs were spliced into human text.
  • Broken or hallucinated citations: dead links, invalid DOIs, cites to search-result pages.

4. Code and repository tells

4.1 Comments

  • Narrating the obvious ("// increment counter") uniformly across files.
  • Change-narration addressed to a reviewer ("// Fixed the bug where...", "// Updated to use the new API"); the comment describes the diff, not the code.
  • Step headers inside functions ("# Step 1: Validate input"); formulaic never-revisited TODOs.
  • Full Args/Returns/Raises docstring boilerplate on trivial functions in a codebase that documents nothing else.
  • Conversational leftovers ("Note that...", "You may want to...").

4.2 Structure

  • Over-engineering tiny tasks: factories, DI, abstract base classes for a 10-line script; textbook 3-5 line functions with zero pragmatic shortcuts; sterile uniformity with no debugging artifacts.
  • The same problem solved five different ways in one repo (one per prompt session): multiple HTTP clients, date libs, validation styles.
  • Defensive try/catch around everything; broad catches that log and ignore.
  • Happy-path-only logic: no validation, no edge cases, no system-wide mental model; missing real plumbing (config, env, logging, auth) in "complete" code.
  • Unrequested "example usage" blocks and __main__ demos.
  • Security anti-patterns: disabled TLS verification, string-interpolated SQL, hardcoded example credentials.

4.3 Naming and hygiene

  • Generic names (data, result, temp, item, processData); formulaic near-duplicates (user_data, user_info, user_object); over-descriptive names where a human writes count; mixed naming conventions in one file.
  • Dead code, orphaned functions, unused imports from abandoned generation attempts.
  • Hallucinated packages and APIs (a 2025 USENIX study measured 19.7% of package references hallucinated; the basis of "slopsquatting" attacks). Verify every dependency exists before adding it.
  • Dependency sprawl for simple apps; imported-but-never-called packages.

4.4 Tests and git

  • Tests that assert trivialities, only exercise the happy path, or mock the unit under test; tests altered to pass instead of the bug being fixed (already law here: CLAUDE.md, testing law).
  • AI authorship trailers (Co-Authored-By, "Generated with") in commits; workspace-wide ban.
  • Generic commit messages ("Add feature") or bloated bullet lists narrating the chat session, including abandoned attempts; huge single-commit diffs mixing unrelated changes.

4.5 READMEs and docs

  • Emoji section headers plus badge walls plus a bolded rule-of-three "Features" list; "This project provides a comprehensive..."; a recap "Conclusion" in a README.
  • Docs describing aspirational features the code does not have.

5. What to do instead

The corrective is the same everywhere: make decisions, and constrain the generator with them before generating.

  • Direct with constraints. Every project keeps a written design direction (here: docs/04-UI-DESIGN.md) that names the palette, type, and voice, and bans the defaults by name. Concrete references beat adjectives.
  • Color: cap at ~3 hues from a decided palette, extend with tints and shades; tint the neutrals; every color resolves to a token.
  • Type: a deliberate display/body pairing, licensed on purpose; never the default stack.
  • Layout: break the fixed skeleton; asymmetric grids, varied section rhythm, left-aligned heroes; one repeating design primitive; hierarchy survives the squint test. Cards borderless by default, separated by whitespace; color strips only for real semantic states.
  • Proof over decoration: real screenshots, verifiable claims, earned logos only; no manufactured social proof.
  • Motion: functional transitions only; design all interactive states; respect reduced-motion.
  • Game art: one locked style guide; one defining prompt per character with reference-linking for every variant; a human unification pass on everything (anatomy, palette, outlines, value range, gibberish text); never ship engine-default themes or fonts.
  • Writing: a named voice with a stake and specifics; vary sentence rhythm; say the true specific thing or cut the line; a human edits every shipped string.
  • Code: match the codebase's existing conventions and comment density; the smallest correct change; verify dependencies exist; comments state constraints the code cannot show, nothing else.
  • The template test: every surface must contain at least one decision a template could not have made. If you cannot name it, the surface is not done.

Sources

Highest-signal sources per track; live-fetched 2026-07-27.

Webapps:

Games:

Writing and code: