HPR4722: Show and tell: AI tropes

This show has been flagged as Explicit by the host.

Obviously, I'll need to introduce why I'm the best person to explain this concept and how I did my research.

After that, I dive into the core of he episode: the research.

The Ultimate Guide To Spotting AI Generated Text That Every Writer Needs To Read.

That title already has two tells:

• Title case : every word capitalised, which conveniently cannot be read out loud
• Hyperbole : a promise of ultimacy, for a ten minute podcast

For most of this episode I am not myself, I am the machine, and every sentence I say is an example of the thing it describes. This page is the key to that part of the show: the tells in the order you hear them, with what they sound like in the recording.

Most names come from tropes.fyi , so you can look them up there and get the longer explanation.

As also stated in the podcast: these aren't errors. They are ordinary figures of speech that human writers have used for centuries. The tell is the density.

Part one: Vocabulary

• Prestige vocabulary : words nobody uses out loud. e.g. delve, harness, utilise, robust, foster
• Grand nouns for ordinary things : e.g. tapestry, landscape, realm, ecosystem, paradigm
• Magic adverbs : imply significance, carry no information. e.g. quietly, fundamentally, arguably, remarkably
• Avoiding the verb "to be" : plain "is" swapped for something that sounds busier. e.g. serves as, stands as, represents, marks
• Invented concept labels : an abstract problem noun (trap, creep, paradox, inversion) welded onto a domain word, then used as if it were established. e.g. "the fluency trap", "prestige creep"
• Vague attributions : authority without a name. e.g. "experts agree", "industry reports suggest", "observers have noted"
• Appeal to familiarity : consensus claimed rather than proven. e.g. "a classic", "famously", "as we all know"
• Promotional register : describing turns into selling. e.g. unlocks, seamless, unprecedented

Part two: Sentence shapes

• Negative parallelism : the "it's not X, it's Y" pattern, and the most reported tell of all. e.g. "This isn't a stylistic quirk, it's a structural signature."
• Not X. Not Y. Just Z. : the same move tripled, building tension for a reveal that needed none. e.g. "Not a habit. Not a tic. A tell."
• Sets of three : one tricolon is elegant, three in a row is a stuck needle. e.g. "it arrives in threes, it lands in threes, it exhausts in threes"
• Short punchy fragments : emphasis manufactured by full stops instead of earned by content. e.g. "Short. Parallel. Relentless."
• Self-answered questions : asks what nobody asked, answers it immediately. e.g. "The problem? Nobody asked."
• Anaphora : the same sentence opening until it becomes furniture. e.g. "They open sentences the same way. They open the next one the same way."
• Trailing "-ing" analysis : shallow significance bolted onto a finished sentence. e.g. "highlighting its enduring importance", "reflecting broader trends"
• False ranges : "from X to Y" where there is no scale and no middle. e.g. "from cesspit to citation"
• Filler transitions : connect nothing to nothing. e.g. "it's worth noting", "importantly", "notably"
• The interjection reflex : every sentence broken in half, whatever the punctuation. e.g. em-dash, double hyphen, comma pair, colon
• Comma-clipped tail : a short phrase hung off a comma instead of landing the point. e.g. "above the content, and save."

Part three: Tone

• Manufactured suspense : the buildup arrives, the revelation does not. e.g. "here's the thing", "here's the kicker"
• Patronising analogy : teacher mode, uninvited. e.g. "think of it as a Swiss Army knife for condescension"
• Forced figurative language : a metaphor picked for cleverness rather than clarity. e.g. "Picture a bicycle. Now picture that the bicycle is prose."
• Stakes inflation : everything is historic. e.g. "a pivotal moment in the evolution of reading itself"
• False vulnerability : honesty performed, never risked. e.g. "let me be honest with you here. Genuinely honest."
• Assertion in place of evidence : e.g. "history is unambiguous on this point", "the evidence is clear"
• "Imagine a world where..." : the futurism sales pitch, pleasant outcomes if you accept the premise. e.g. "imagine a world where every paragraph carries a quiet intelligence"
• Quotable one-liners : slide bait that survives being pulled out of context because it never had any. e.g. "Every bite is a chance to be here, now, fully."
• The collaborative "we" : one author, plural pronoun. e.g. "we will unpack each one in turn"
• Sycophancy : not a figure of speech but a training artefact, and the fastest way to spot raw chatbot output. e.g. "What a genuinely profound question."

Part four: Formatting

• Title case headings : see the top of this page
• Bullet lists where prose would do : the document turns into a slide deck. e.g. this page, from here to the bottom
• Bold-first bullets : every item opens with a bolded keyword and a colon. e.g. yes, also this page, I know
• Emoji as punctuation : rocket for aspiration, key for insight, sparkles because the paragraph ended
• Em-dashes : mid-sentence, twice, sometimes three times, where a comma would have done the job
• The find and replace tell : double hyphens everywhere, because someone removed the character and kept the habit
• Unicode decoration : arrows and curly quotes nobody typed on a keyboard. e.g. Input → Processing → Output
• One thought per line : paragraphs abolished. e.g. "It happened. Alone. On a Tuesday."

Part five: Composition

• Announce then answer : says what the text is about to do instead of doing it. e.g. "This guide will explore what those signs look like."
• Rapid-fire historical analogies : authority borrowed from a list, with no argument underneath. e.g. "Marconi. Sony. Apple. Print, radio, web, mobile."
• The metaphor that will not end : e.g. the garden that stays a garden for six sentences, then becomes a north star that moves the needle
• Self-echo : a word from earlier returning as if it were being paid off, when it is the same narrow vocabulary surfacing again
• One point diluted : the same idea in five coats, sounding like progress. e.g. "put differently", "to rephrase", "in other words"
• Fractal summaries : every section summarises itself, including the one you are inside. e.g. "In this section we explored composition."
• The tie-back : closes by looping the answer back to the original question, long after the point was made. e.g. "So, to answer your question..."
• Concede and dismiss : e.g. "despite its fluency, it faces challenges typical of automated systems. Despite those challenges, it endures."
• Signposted conclusion : competent writing does not need to announce that it is ending. e.g. "in conclusion", "to sum up", "in summary"
• The conclusion that refuses to end : clause stacked on clause, quieter and more meaningful each time. e.g. "the real AI generated text was the paragraphs we scrolled past along the way"
• Engagement bait : e.g. "Agree?" plus a pointing emoji

Tells on the list that did not fit in the episode

The recording would have run to forty minutes. These are worth knowing anyway.

• Reasoning leak : the text narrating its own decisions instead of just making them. e.g. "I want to be exact about my own role here."
• Premise stacking : a paragraph of evidence in front of the point, so the point has been made twice before it arrives
• Compulsive counting : states the number of items before listing them. e.g. "Five things we wish to discuss"
• Excessive enumeration : a listicle disguised as prose. e.g. "The first wall is... The second wall is..."
• Belaboring the unnecessary : defending a minor point against an objection nobody was going to raise
• Synonym cycling : refusing to repeat a noun. e.g. the dashboard, then the interface, then the portal, then the analytics hub
• Wh-headings : the default shape a model reaches for when naming a section. e.g. "What this means", "Why it matters"
• Content duplication : the same paragraph appearing twice, because the model lost track of what it had written

Why this is not a detector guide

Liang, Yuksekgonul, Mao, Wu and Zou, Stanford, published in Patterns (Cell Press) on 10 July 2023. Seven commercial GPT detectors were run over 91 TOEFL essays written by non-native English speakers and 88 US eighth grade essays.

• Average false positive rate on the TOEFL essays: 61.3 percent.
• 89 of the 91 essays were flagged as AI generated by at least one detector. All seven agreed on 18 of them.
• The US eighth grade essays came back near perfect. Plainer word choice, lower perplexity, reads as machine to the software.
• Feeding the TOEFL essays through ChatGPT with the prompt "enhance the word choices to sound more like that of a native speaker" dropped the false positive rate to 11.6 percent.

So the way to stop being accused of writing like a machine was to have a machine rewrite you. Do not take my word for it, the paper is open access:

• https://www.cell.com/patterns/fulltext/S2666-3899(23)00130-7
• Preprint: https://arxiv.org/abs/2304.02819

The pointers, for real

• Write with things in it: names, dates, prices, the model number of the drive that died.
• Say something a machine has no way to know, because it happened to you.
• Be willing to be wrong out loud.
• Keep one weird sentence that your editor would cut.

Links and references

• The list this episode leaned on: tropes.fyi , by ossama.is
• The detector study: Liang et al., "GPT detectors are biased against non-native English writers", Patterns, 2023
• "Not an Addict", K's Choice, 1995. It's not a habit, it's cool, I feel alive.
• "Friends, Romans, countrymen" and "I come to bury Caesar, not to praise him": Shakespeare, Julius Caesar, act 3 scene 2
• "Veni, vidi, vici", attributed to Julius Caesar by Suetonius and Plutarch

Over to you

There is a comment section under this episode. Tell me which tell I missed, and which one you are guilty of. Better yet, record it. Read your worst, most machine sounding paragraph out loud and post it to Hacker Public Radio. The queue is always hungry.

Provide feedback on this episode.

HPR Comments

Antoine says: AI writes very well

Antoine says: Re #7 - Brazilian ... logic?

Ben says: Great text

Steve Barnes says: Re: This is so true

Trey says: This is so true

Trollercoaster says: Re: AI writes very well

Trollercoaster says: Re: Great Text

Trollercoaster says: Re: This is so true

Trollercoaster says: Re: cool ep

candycanearter07 says: cool ep

Mastodon Comments

More Information...