How to Actually Make AI Write Like You, Not Like a Robot, With the Accent of Your Favorite LLM.
Every AI model has a built-in writing style that hides your voice. This article shows the research that proves it, then gives you the complete fix.
TL → DR
Every AI model writes in a default style. Researchers can identify which model wrote a text with 99.88 percent precision, even when the model was told to write differently.
The proof has been very public. The word “delve” rose 14-fold in science papers. “Meticulous” ran at 34.7 times its normal rate in AI-edited text. At least 10 percent of 2024 medical abstracts show AI processing.
Describing your voice to the AI is not enough. Models drop instructions as you stack them, and every dropped instruction gets replaced with the default style.
The fix has 2 parts and a compendium of samples. Write a positive profile of what you do. Write a prohibition list that bans the AI’s defaults by name. Attach 3 samples of your real writing. Load it all before every task.
Each platform needs its own setup. Claude takes the full profile as a Project or saved Skill instructions. ChatGPT caps custom instructions at 1,500 characters per field, so the full profile goes in a Project. Gemini needs the samples placed right before the task.
The stakes are growing. The AI’s favorite words are now rising 25 to 50 percent a year in human speech, and 80 percent of people who wrote essays with AI help could not quote their own essays.
Want to learn how to make AI actually write in your true voice?
Register for my free masterclass, happening tomorrow.
In the first quarter of 2024, the word “delve” appeared in scientific papers at roughly 14 times its historical rate. For 2 decades it had idled at about half a mention per 1,000 papers. Then it hit 7.9. Nobody called a meeting about this. No style guide changed, and no editor sent a memo asking the world’s researchers to start “delving.” A machine developed a preference, and several hundred million people carried it into the permanent scientific record, where it sits now like a watermark.
That watermark is the subject of this essay. The machine you write with has a voice of its own, and it is overwriting yours. Until you name that voice and explicitly turn it off, your instructions are negotiating with a fingerprint that was trained long before you arrived.
Most advice about AI writing never touches this. It teaches you to describe what you want, in ever more detail, and assumes the model is a blank instrument waiting for direction. The evidence says otherwise. The model arrives with commitments. You do not have a prompting problem. You have a fingerprint problem, and the fix has 2 jobs where almost everyone does 1.
The live masterclass is tomorrow. Everything in this article gets built in the room, in 1 hour: the 2-sided voice profile, the prohibition list, and the 3-machine setup for Claude, ChatGPT, and Gemini. You watch me build mine, then you build yours. The session is free.
Why does all AI writing sound the same?
A team led by Dmitry Kobak analyzed more than 14 million biomedical abstracts and found that at least 10 percent of 2024 abstracts had been processed with a language model. In some subfields, the estimate reached 30 percent. The giveaway sat in 329 “excess words” that spiked past their projected baselines, and the composition of that list is the telling part. 66 percent were verbs, 18 percent adjectives. Vocabulary of style rather than substance.
A separate study of conference peer reviews measured the same signature from a different angle. In the AI-modified text, the adjective “commendable” appeared at 9.8 times its expected rate. “Intricate” ran at 11.2 times. “Meticulous” ran at 34.7 times. At EMNLP 2023, an estimated 16.9 percent of review sentences had been substantially modified by a model. The reviewers judging the field’s science were, in 1 of 6 sentences, letting the machine speak for them.
“Meticulous” at 34.7 times baseline, “delve” at 14, “intricate” at 11.2, “commendable” at 9.8. At least 10 percent of 2024 biomedical abstracts show LLM processing.
The issue reaches well past academia. Stanford researchers tracked it across the wider economy through late 2024 and found it in 17.7 percent of consumer complaints, 23.8 percent of corporate press releases, 13.7 percent of United Nations press releases, and roughly 1 in 10 job postings.
A quarter of corporate communication now arrives pre-inflected. Somewhere in a training pipeline, a preference formed the way an accent forms, absorbed rather than chosen, and it is now the most widely spoken accent on earth.
How do researchers detect AI writing?
Catching an author by counting words predates the AI era by 60 years. In 1964, the statisticians Frederick Mosteller and David Wallace settled a 175-year dispute over who wrote 12 unclaimed Federalist Papers. They ignored the arguments and counted the filler. Madison wrote “whilst” where Hamilton wrote “while.” Hamilton reached for “upon” at rates Madison never approached. The invisible words convicted him. All 12 papers went to Madison, and forensic stylometry became a discipline.
The same method now works on machines, and it has become far more effective. A 2025 study trained classifiers on 200,000 texts from 4 model families and identified which model wrote a given passage with 0.9988 precision. The authors’ summary sentence deserves to be quoted in full: large language models “have distinct and consistent stylistic fingerprints, even when prompted to write in different writing styles.”
Models are trained toward human approval, and the training compresses them. Researchers have documented that reinforcement learning from human feedback reduces output diversity. The text comes out more similar, regardless of what goes in. A 2025 paper traced the cause to what it called "typicality bias." Human raters systematically reward conventional, expected phrasing, independent of quality, so the model’s probability mass collapses toward its most typical sentences. Anthropic describes the endpoint of this process plainly in its own research. Post-training, in the lab’s words, selects “one particular character from this enormous cast” and places it center stage. The character is stable. It generalizes across models. It is a house voice.
Why does AI ignore my writing instructions?
The standard response is to write better instructions. Describe the voice. Add detail. Be firm. The instruction-following literature explains why this keeps disappointing you.
On IFEval, the benchmark for objectively checkable instructions, a GPT-4 class model failed roughly 1 multi-instruction prompt in 4; these were instructions a script could verify, including word counts and required keywords. A 2025 study stacked constraints and watched compliance sag, with the best frontier models honoring about 68 percent at a density of 500 instructions. The Allen Institute for AI then ran the sharpest version of the test. Models scoring high on familiar constraints collapsed on unfamiliar ones. Claude 4 Sonnet went from 91.3 percent to 42.3. Gemini 2.5 Pro went from 65.4 to 52.3. OpenAI’s o3 went from 95.0 to 69.3.
The same 3 models on seen constraints (IFEval) and unseen constraints (IFBench). Claude 4 Sonnet drops from 91.3 to 42.3. The best frontier accuracy at 500 stacked instructions is 68 percent.
Your tone rules are exactly the kind of unfamiliar, unverifiable constraint that decays first. When a constraint decays, the model reverts to the trained default. So the more elaborate your description of your voice, the more constraints you have queued for decay, and every decayed constraint is refilled with house style.
There is a second finding that matters even more for the fix we are all seeking. Models exhibit what MIT researchers call affirmation bias. They under-process negation and instead attend to the named items, with performance in the underlying study dropping by nearly 25 percent when statements were negated. The study was conducted on vision-language models. But the direction of the finding aligns with every operator’s experience with text models. A default you have not named stays on. A default you have named weakly also stays on. Suppression has to be explicit. It has to be specific. And the machine needs to hear it more than once, because its reflex is to register the noun and miss the “never.”
How I Built a $200,000 Newsletter Business From a Substack Nobody Was Reading
The Five Moves That Turned My Substack Into a $200,000 Business
The Operating Project: Most Coaching Ends With Notes. This Ends With a Machine.
What is a 2-sided voice profile?
The first side of a working voice profile is the positive profile, the description of what you actually do. Your sentence rhythm. Your openings and your closings. Whether you lead with the conclusion or build to it. Your punctuation habits. How your register shifts between a cold email and an internal memo. Most people who take AI writing seriously have built some version of this. It is necessary, but by itself it does almost nothing about the fingerprint because it never mentions the fingerprint.
The second side is the prohibition list, the explicit inventory of what the model must never default to. The negation reflex, “it’s not X, it’s Y,” which the machine deploys with the regularity of a metronome. The rule of three, triplets in every sentence, makes it feel complete. The vocabulary you now know by heart: “delve,” “robust,” “seamless,” “testament,” “pivotal,” “underscore.” The grand contextual opener that clears its throat for a paragraph before saying anything. Signposting. Paragraphs of eerily uniform length. Each of these is a measured, documented default, and each one stays on until it is named and switched off. The complete inventory, with the measured effect sizes, sits in the appendix at the end of this piece.
The 2-sided architecture. Positive profile, prohibition list, and 3 few-shot anchors, loaded together before the task.
You can spend an hour describing your voice in exquisite detail, and the model will still hand you “delve” and a triplet, because you described what you want and never prohibited what it does. Both sides need to load together, or the calibration fails.
Do writing samples work better than instructions?
The third component outperforms both descriptions.
Show the model your actual writing.
The finding is as old as the modern field. The 2020 GPT-3 paper demonstrated that models can perform new tasks from a handful of examples in the prompt, with no retraining required, and quantified the gap. On 1 benchmark, accuracy went from 64.3 percent with instructions alone to 71.2 percent with a few examples. Later work sharpened the reason in a way that matters for voice. When researchers deliberately corrupted the content of in-context examples, performance barely moved. What the model was absorbing was the shape. The distribution of the text, the format, and the rhythm of how input becomes output. Which is to say, examples convey exactly what you cannot fully put into words about your own prose.
Anthropic’s documentation calls examples “one of the most reliable ways to steer Claude’s output format, tone, and structure.” OpenAI’s guidance compresses it to 3 words: “show and tell.”
This should only take 10 to 15 minutes. Pick 3 of your best pieces of writing in different registers: 1 email, 1 memo or post, and 1 longer piece. These are your few-shot anchors, your voice in its native format.
How do I set up a voice profile on ChatGPT, Claude, and Gemini?
Here is where nearly all AI writing advice goes soft. It claims to work on any model, which is precisely why it works well on none of them. The machines differ, and the differences are documented. A classifier separates their prose at 0.9988 precision. The same voice profile, loaded identically, produces 3 different drifts.
Claude honors explicit prohibitions most faithfully, and its failure mode is over-compliance. Strip out enough defaults and it can go flat. The fix is to pair each major prohibition with 1 positive rule. Avoid signposting, but open with strong declaratives. Load the full profile, positive side first, prohibition list second, anchors third, as per Project instructions, so every conversation inherits it.
ChatGPT’s default is agreeable polish. It will sand your fragments and finish your abrupt endings unless you name them as features. Its custom-instructions fields cap at 1,500 characters each, so the play is a condensed rule set there and the full profile with anchors as the first message in a Project, where it persists.
Gemini over-formalizes, an elevation reflex more pronounced than the other 2, and it weights recent context heavily. Put the profile in a Gem’s instructions, place the examples immediately before the task rather than burying them early, and give the prohibition list an explicit register rule. Match the examples, not what reads as professional. The position advice has empirical teeth. Language models reliably privilege the beginning and end of their context window over the middle, and Anthropic’s own docs report that placing the query after long context improves response quality by up to 30 percent.
1 profile, 3 platforms. Claude over-complies, ChatGPT polishes, Gemini formalizes. Each drift has a 1-adjustment fix and a placement rule.
The differences are proof that voice calibration is a technical discipline rather than a mindset, and that platform-agnostic advice stays shallow for that reason.
Is AI changing how people write and speak?
The strangest finding in this literature concerns your speech rather than your drafts.
Researchers at the Max Planck Institute for Human Development analyzed 360,445 academic talks on YouTube and 771,591 podcast episodes, roughly 7.35 billion spoken words. In the 18 months after ChatGPT launched, the words the model favors- “delve,” “realm,” “meticulous,” “underscore,” rose in unscripted human conversation at annual rates of 25 to 50 percent. The patterns stored in the technology “seem to be transmitting back to the human mind.”
Machine-favored words are rising 25 to 50 percent yearly in 7.35 billion words of human speech. 80 percent of AI-assisted writers could not quote their own essays.
The written evidence points the same way, and it is peer-reviewed. A Science Advances study gave 300 writers AI-generated story ideas and had roughly 600 evaluators judge the results. Individual stories got better, especially from the weaker writers. The collection got worse. Stories written with AI help were measurably more similar to one another than stories written without it. Each writer improved into the same writer. And the pull has a direction. A Cornell study of American and Indian writers using AI autocomplete found the suggestions steadily moved Indian participants toward Western styles, which means the default is 1 register, exported, rather than an average of everyone.
The New Yorker’s Kyle Chayka compressed the phenomenon into 5 words. AI, in his phrase, is “a technology of averages.” The MIT study he reported carried the detail that should bother you most. Of the subjects who wrote essays with ChatGPT’s help, 80 percent could not quote from what they had putatively written.
The professional class spent the last 2 years paying monthly subscriptions to sound like each other. The tool was sold as amplification and adopted as ventriloquism, and most users never noticed the transfer because the prose kept arriving fluent and finished. The fluency we were purchasing was cheap. What belongs to the machine is “finished,” and increasingly, the memory of having written at all.
Can AI detectors prove a text was written by AI?
The word-level tells are aggregate signals, not individual proof. AI detectors built on them misclassified 61.22 percent of essays by non-native English speakers as machine-written in a 1 study at Stanford. A single “delve” convicts nobody. The mechanism behind the vocabulary is also unproven. The popular theory blaming annotation workers’ dialects was tested directly and found unsupported. And the fingerprints move. One 2025 analysis found that DeepSeek’s prose was classified as OpenAI’s 74.2 percent of the time, which suggests the house styles can converge as labs train on similar data. The prohibition list you write this year will need an audit next year.
How do I make AI write in my voice this week?
What to do this week…
Pull your last 5 AI-assisted drafts and audit them against the tells in Figure 1.
Count the negation structures, triplets, marker words, and signposts.
This count is your baseline.
Write the positive profile in 1 sitting. Rhythm, openings, escalation, punctuation, argument order, register shifts. Keep it to one page max.
Write the prohibition list from your audit. Name every default you found, specifically. Not “avoid AI-sounding language” but “never use ‘delve,’ never construct ‘it’s not X, it’s Y,’ never signpost.”
Choose your 3 anchors and load all of them together before the task, on whatever machine you use, configured the way that machine requires.
Then track your edits for a week. Every word you swap and every sentence you reorder is data on a rule the profile still lacks. The gap between what the model gave you and what you shipped is your voice, captured in real time.
In the live session, we build both sides in the room, run the same profile through Claude, ChatGPT, and Gemini side by side, and fix each machine’s drift as it appears. Bring the 3 anchors.
What words and patterns give away AI writing? The full prohibition list.
The difference between a writer’s reframe and the machine’s reflex is about choice, and frequency. Use the pattern once, on purpose. Without clear directoin, the machine uses it 6 times a page. The same goes for every entry below. Possession of a single marker creates little impact, and detectors built on these markers falsely flagged 61 percent of essays by non-native English speakers. The tell most always lives in the presence of the same move, arriving again and again.
The markers fall into 2 evidentiary classes:
The vocabulary is measured, with effect sizes from studies covering more than 14 million documents.
The structures are documented by detection researchers and linguists, who observe rather than quantify them.
Load all 4 sections into your AI as explicit prohibitions, then pair each structural prohibition with 1 positive rule from your own writing.
Which words does AI overuse?
The 15 measured word families. “Delves” grew 6,697 percent, 2020 to 2024.
The core cluster, with effect sizes. Growth figures measure frequency change in scientific writing, from 2020 to 2024 (Juzek and Ward). Fold-increase measures overrepresentation in AI-modified peer reviews (Liang et al.). Frequency ratios measure excess above baseline in 14 million PubMed abstracts (Kobak et al.).
delve / delves / delved / delving. “delves” grew 6,697 percent. The flagship marker.
underscore/underscores/underscoring. “underscores” grew 904 percent.
showcase/showcasing/showcases. “showcasing” grew 1,396 percent.
meticulous / meticulously. 34.7-fold increase in AI-modified reviews.
intricate / intricacies. 11.2-fold increase. “Intricacies” grew 773 percent.
commendable. 9.8-fold increase.
boasts. Grew 918 percent.
comprehending. Grew 899 percent.
surpass/surpasses/surpassing. Grew 667 percent.
garnered. Grew 437 percent.
emphasizing. Grew 397 percent.
realm. Grew 381 percent.
groundbreaking. Grew 330 percent.
advancements. Grew 278 percent.
aligns. Grew 267 percent.
The documented adjective set, from the same studies’ extended lists: notable, versatile, noteworthy, invaluable, pivotal, potent, ingenious, laudable, lucid, cogent, methodical, profound, refreshing, intriguing, admirable, exceptional, remarkable, seamless, holistic, insightful, comprehensive, compelling, unprecedented, cohesive, tangible, multifaceted.
The documented adverb set: notably, aptly, methodically, compellingly, impressively, undoubtedly, undeniably, seamlessly, effortlessly, robustly, thereby, hitherto, herein, additionally, subsequently, crucially, markedly, judiciously.
A composition note: Of the 329 excess words of 2024, 66 percent were verbs and 18 percent were adjectives. The fingerprint lives in how the machine says things. Content words like “potential,” “findings,” and “crucial” also rose and remain weak signals on their own since any writer covering research reaches for them.
Which phrases give away AI writing?
“It’s not just X, it’s Y” and every variant.
“In today’s fast-paced world.”
“In the ever-evolving landscape of.”
“Plays a pivotal role in.”
“Serves as a testament to.”
“Rich tapestry.” “
Navigating the complexities of.”
“Unlock the potential of.”
“At the intersection of.”
“In the realm of.”
“A delicate balance between.”
“It’s important to note that.”
“It’s worth noting that.”
“Let’s dive in.”
“Let’s delve into.”
“In conclusion” and “Ultimately” are closing moves.
“Whether you’re a beginner or an expert.”
“Underscores the importance of.”
“Sheds light on.”
“The key takeaway.”
Stacked hedges like “could potentially.”
And every trace of the eager assistant: “Great question.”
Which sentence structures does AI default to?
The antithesis construction: “This isn’t X. It’s Y.” Deployed several times per page.
The rule of three: Nouns, adjectives, and clauses arriving in threes, in every sentence that wants to feel complete.
The participial tail: A clause bolted on with an “-ing” verb: “...showcasing its versatility,” “...underscoring the need for further research.” 5 of the 21 fastest-growing marker words are exactly these participles.
The em dash cascade: Several per paragraph, doing work commas and periods should do.
Uniform sentence length: Human writing mixes 4-word sentences with 40-word sentences. Machine writing hums at a steady cadence.
The balanced concessive: “While challenges remain, the opportunities are significant.” Every tension is neutralized in a single sentence.
Elevated formality throughout: No contractions, no slang, no fragments, even where a human would relax.
The signposted transition: “Additionally,” “Furthermore,” “Moreover,” “Notably” at the head of paragraphs after paragraphs. Several of these adverbs appear on the lists of measured overuse.
Which document habits give AI away?
The throat-clearing open: A paragraph of context before anything is said.
Uniform paragraph length: Every paragraph has 3 to 4 sentences; the whole page is a solid gray.
The scaffolded essay: First, second, third, finally, whether or not the material has that shape.
Bolded lead-in bullets: Lists where every item opens with a bolded phrase and a colon, much like this section, but it’s how I like to construct sections like this, so I do it anyway.
The both-sides shuffle: Every claim balanced by its counterclaim, no position taken.
Over-explanation: Restating the question, summarizing what was just said, announcing what comes next.
The tidy moral: A closing paragraph that resolves everything and lands on an uplifting generality.
Completion compulsion: Nothing left open, every thread tied, even when the honest answer is unresolved.
Frequently asked questions
Why does my AI writing sound like AI?
Every model is trained toward the answers human raters prefer, and raters reward conventional phrasing. That training compresses the model into 1 default style, the same vocabulary and sentence shapes for every user. Your instructions sit on top of that default. They soften it. They do not remove it until you ban its parts by name.
What words does AI overuse?
The measured core: delve, underscore, showcase, meticulous, intricate, commendable, boasts, comprehending, surpass, garnered, emphasizing, realm, groundbreaking, advancements, and aligns, plus their inflections. “Delves” grew 6,697 percent in scientific writing between 2020 and 2024. The full list, with phrases and structures, is in the appendix above.
How do I stop ChatGPT, Claude, or Gemini from sounding like AI?
Build a voice profile with 2 sides. Side 1 describes what you do: your rhythm, your openings, your punctuation. Side 2 bans the machine’s defaults by name: the words, the “it’s not X, it’s Y” reflex, the rule of three, the signposting. Attach 3 samples of your real writing. Load all of it before the task, every time.
Do writing samples work better than written instructions?
Yes. The foundational GPT-3 research showed accuracy rising from 64.3 to 71.2 percent on the same task when examples were added. Later work found that examples transfer format and rhythm, the parts of a voice you cannot fully describe. Anthropic calls examples “one of the most reliable ways to steer” tone. OpenAI’s version is “show, and tell.”
Can AI detectors prove a text was written by AI?
No. The word-level tells are reliable across millions of documents and unreliable on any single one. In 1 Stanford study, detectors falsely flagged 61.22 percent of essays by non-native English speakers as AI-written. A single “delve” convicts nobody. The tell is the pattern repeating across a whole body of text.
Is the same voice profile enough for every AI platform?
The words are the same. The setup is different. Claude takes the full profile as Project instructions and needs each ban paired with a positive rule. ChatGPT caps custom instructions at 1,500 characters per field, so the condensed rules go there and the full profile goes in a Project. Gemini weights recent context heavily, so the samples go right before the task.
Is AI changing how people talk?
The early evidence says yes. Max Planck researchers analyzed 7.35 billion words of podcasts and talks and found the machine’s favorite words rising 25 to 50 percent a year in unscripted speech after ChatGPT launched. The study is a preprint, so treat it as a strong early signal rather than settled fact.
How long should a voice profile be?
Short enough to survive. Instruction-following decays as rules stack, so a bloated profile fails faster. Target 1 page for the positive side, a named prohibition list, and 3 writing samples. Fewer, harder rules beat many soft ones.
Build it live, tomorrow.
This article gives you the doctrine. The masterclass gives you the working setup.
Tomorrow, live, we will build every piece from this article in 1 hour. Your positive profile. Your prohibition list, drawn from the measured word lists above. Your 3 few-shot anchors. Then we load the same profile into Claude, ChatGPT, and Gemini side by side and fix each machine’s drift as it appears, so you leave with your voice running on whatever AI you already use.
You do not need a technical background.
You need 3 pieces of your own writing and 1 hour.
The session is free.
Research appendix: every source in this article.
The measured tells
Kobak, González-Márquez, Horvát and Lause, “Delving into LLM-assisted writing in biomedical publications through excess vocabulary,” Science Advances 2025. 14M+ PubMed abstracts, the 329 excess words, the 10 percent figure. arxiv.org/abs/2406.07016
Liang et al., “Monitoring AI-Modified Content at Scale,” ICML 2024. The peer-review fold increases: meticulous 34.7, intricate 11.2, commendable 9.8. arxiv.org/abs/2403.07183
Liang et al., “The Widespread Adoption of Large Language Model-Assisted Writing Across Society,” Stanford HAI / Patterns 2025. The economy-wide adoption rates. arxiv.org/abs/2502.09747
Juzek and Ward, “Why Does ChatGPT ‘Delve’ So Much?,” COLING 2025. The 21 focal words with growth rates, and the tested-and-rejected annotator theory. arxiv.org/abs/2412.11385
Shapira, “Delving into ‘delve,’” 2024. The 14-fold rise of the flagship word. pshapira.net/2024/03/31/delving-into-delve
Gray, “ChatGPT ‘contamination’: estimating the prevalence of LLMs in the scholarly literature,” 2024. The scholarly keyword growth figures. arxiv.org/abs/2403.16887
The fingerprint and its mechanism
Bitton, Bitton and Nisan, “Detecting Stylistic Fingerprints of Large Language Models,” 2025. The 0.9988 precision result and the DeepSeek convergence finding. arxiv.org/abs/2503.01659
Kirk et al., “Understanding the Effects of RLHF on LLM Generalisation and Diversity,” ICLR 2024. Post-training reduces output diversity. arxiv.org/abs/2310.06452
Zhang, Yu, Chong et al., “Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity,” 2025. Typicality bias as the cause. arxiv.org/abs/2510.01171
Anthropic, “The Assistant Axis: Situating and Stabilizing the Character of Large Language Models,” 2026. The default character, from the lab itself. anthropic.com/research/assistant-axis
Mosteller and Wallace, “Inference in an Authorship Problem,” Journal of the American Statistical Association, 1963. The Federalist Papers attribution. jstor.org/stable/2283270
Why instructions decay
Zhou et al., “Instruction-Following Evaluation for Large Language Models” (IFEval), 2023. The 1-in-4 prompt failure figure. arxiv.org/abs/2311.07911
Jaroslawicz et al., “How Many Instructions Can LLMs Follow at Once?,” 2025. 68 percent at 500 stacked instructions. arxiv.org/abs/2507.11538
Pyatkin et al., “Generalizing Verifiable Instruction Following” (IFBench), Allen Institute for AI, 2025. The seen-versus-unseen collapse: 91.3 to 42.3. arxiv.org/abs/2507.02833
Alhamoud et al., MIT, on affirmation bias and negation, 2025. Why prohibitions must be explicit. news.mit.edu
Why examples beat descriptions
Brown et al., “Language Models are Few-Shot Learners,” NeurIPS 2020. 64.3 to 71.2 percent with examples added. arxiv.org/abs/2005.14165
Min et al., “Rethinking the Role of Demonstrations,” EMNLP 2022. Examples transfer format and structure, corrupted labels barely matter. arxiv.org/abs/2202.12837
Liu et al., “Lost in the Middle: How Language Models Use Long Contexts,” TACL 2024. Position in the context window changes what the model uses. arxiv.org/abs/2307.03172
Platform documentation
Anthropic, multishot prompting guidance. Examples steer tone and structure. platform.claude.com
Anthropic, prompting best practices. The query-at-the-end 30 percent figure. platform.claude.com
OpenAI, prompt engineering best practices. “Show, and tell.” help.openai.com
OpenAI, custom instructions for ChatGPT. The 1,500-character limit. help.openai.com
Google, tips for creating custom Gems. support.google.com
The spread into human writing and speech
Yakura et al., Max Planck Institute for Human Development, on LLM-favored words in spoken language, 2025, preprint. 7.35B words, 25 to 50 percent yearly growth. arxiv.org/abs/2409.01754
Doshi and Hauser, “Generative AI enhances individual creativity but reduces the collective diversity of novel content,” Science Advances 2024. science.org
Agarwal, Naaman and Vashistha, on AI suggestions homogenizing writing toward Western styles, CHI 2025. arxiv.org/abs/2409.11360
Chayka, “A.I. Is Homogenizing Our Thoughts,” The New Yorker, 2025. “A technology of averages” and the MIT 80 percent figure. newyorker.com
The limits of detection
Liang et al., “GPT detectors are biased against non-native English writers,” Patterns 2023. The 61.22 percent false-positive figure. cell.com
About John
John Brewton documents the history and future of operating companies at Operating by John Brewton. He is a graduate of Harvard University and began his career as a PhD student in economics at the University of Chicago. After selling his family’s B2B industrial distribution company in 2021, he has been helping business owners, founders, and investors optimize their operations ever since. He is the founder of 6A East Partners, a research and advisory firm asking the question: What is the future of companies