By Ashley Salek, Agency Director, Seventh Element. Last updated 14 October 2026.
The short answer: The signs of AI writing are predictable words (“delve”, “testament”, “seamless”), stock phrases, tidy groups of three, “it’s not X, it’s Y” lines, uniform rhythm and confident claims with no source. They appear because models default to the most likely next word. The fix is a better prompt, a lean tone-of-voice setup and five minutes of human editing, not a humaniser tool.
We have a confession. Somewhere in our shared drive sits a folder of AI drafts that all open with “In today’s fast-paced digital landscape”. Nobody remembers asking for that. Nobody remembers anyone talking like that, ever.
Most teams have a folder like this. The usual response is to blame the robot, buy a “humaniser” and hope for the best.
We think that’s the wrong diagnosis. AI writing sounds like AI because people prompt it lazily and publish it unedited. Both are fixable this afternoon. This guide shows you the tells, gives you the prompts and Claude skills that avoid most of them, shows you how to set up your tone of voice in Claude, ChatGPT and Perplexity, and hands you a five-minute edit routine for the rest.
Why does AI writing sound like AI?
AI writing sounds like AI because a language model predicts the most probable next word, and the most probable word is rarely the most interesting one. Without a clear brief, it falls back on a safe, formal, slightly breathless register that sits halfway between a press release and a school essay.
Google says as much in its own guidance. Its Search Central page on generative AI content notes that models work by predicting likely word sequences, which is also why they can make things up.
Three things push a model towards the robot voice:
- A thin brief. “Write a blog about X” gives it nothing to go on, so it writes about X the way the average of the internet writes about X.
- No voice reference. If you don’t show it how you sound, it can’t sound like you.
- No permission to have an opinion. Models hedge by default. Ask for a balanced overview and you get exactly that, all 900 limp words of it.
The robot voice is a default setting, not a personality. Change the inputs and most of it goes away. Edit the output and the rest goes too.
Does it actually matter if your copy sounds like AI?
Yes, and the evidence points in one direction. Readers don’t need proof that AI wrote something. A suspicion is enough to cost you.
Finding | Figure | Source |
Drop in trust when readers suspect content is AI-generated | 48% lower | Raptive, survey of 3,000 US adults, Aug 2025 (vendor research) |
Drop in perceived authenticity when AI is suspected | 57% lower | Raptive, Aug 2025 (vendor research) |
US adults who say it’s important to tell AI content from human content | 76% | Pew Research Center, Sept 2025 |
US adults not confident they could tell the difference | 53% | Pew Research Center, Sept 2025 |
UK adults comfortable with news made automatically by AI | 11% | Reuters Institute Digital News Report 2025 |
Workers who received low-effort AI “workslop” in the past month | 40% | BetterUp Labs and Stanford Social Media Lab, Sept 2025 |
Put those together and you get an awkward picture. People care a lot about whether something was written by a machine. They aren’t sure they can tell. So they go on feel. And when it feels like AI, they mark you down.
The Raptive figures come from Raptive’s own study, so treat the exact percentages with a little caution. The direction matches everything else we’ve seen, though.
It costs more than clicks
The workslop research from BetterUp Labs and the Stanford Social Media Lab is the internal version of the same problem. Each incident cost recipients close to two hours to deal with. People who received it rated the sender as less capable (50%) and less trustworthy (42%).
Swap “colleague” for “prospect” and “memo” for “proposal” and you can see why this matters for marketers. Sounding like AI doesn’t just make your copy weaker. It makes you look like you didn’t care enough to read it.
Readers forgive AI. They don’t forgive laziness. The penalty lands when the copy shows nobody was paying attention.
What are the signs of AI writing?
The signs of AI writing fall into six families: vocabulary, stock phrases, structure, rhythm and formatting, substance, and tone. One tell on its own means nothing. Humans use “crucial” too. Four or five in a paragraph is a pattern, and readers clock patterns fast.
Family | What it looks like | Typical examples |
Vocabulary | Words people rarely say out loud | delve, tapestry, testament, seamless, pivotal, underscore |
Stock phrases | Filler that could open any article | “In today’s fast-paced world”, “It’s important to note”, “When it comes to” |
Structure | The same shapes, every time | “It’s not X, it’s Y”, groups of three, question then instant answer |
Rhythm and formatting | Too tidy to be human | Same-length sentences, em dashes, bold labels on every bullet, Title Case Headings |
Substance | Confident but empty | Hedging, vague examples, “studies show” with no study, inflated significance |
Tone | A brochure that read a self-help book | Relentless positivity, “Great question!”, “revolutionary”, empathy clichés |
The vocabulary is real, and measurable
This isn’t just a vibe. Researchers have counted it.
- In academic writing: Kobak and colleagues analysed 15.1 million PubMed abstracts and estimated that at least 13.5% of 2024 abstracts were processed with an LLM. “Delves” appeared about 28 times more often than the pre-ChatGPT trend predicted. “Underscores” and “showcasing” also jumped (Science Advances, 2025).
- In speech: Researchers at the Max Planck Institute for Human Development analysed around 740,000 hours of YouTube talks and podcasts. They found “delve” and other ChatGPT-favoured words rising in spoken academic talks after ChatGPT launched (arXiv, July 2025).
So yes, people now say “delve” out loud. Partly because they read it so often. That’s the point at which a word stops being vocabulary and starts being a watermark.
Structure gives it away faster than words
Word lists get all the attention, but structure is the bigger tell. Swap every “delve” for “explore” and the copy still reads like AI if every paragraph has the same shape.
The patterns we see most in client drafts:
- “It’s not X, it’s Y.” Once is a point. Three times in one post is a tic.
- Everything in threes. Fast, simple and effective. Clear, concise and compelling. Real life rarely comes in tidy triplets.
- The rhetorical question with an instant answer. “The result? A 40% uplift.” Fine once. Exhausting as a format.
- The intro that restates the headline, then the conclusion that restates the intro.
- Bullets with a bold label and a colon on every line, whether the content needs them or not.
We ran this article against its own list before publishing. It was humbling. We cut four triplets and one “here’s the thing”.
Is the em dash really an AI tell?
Partly. Em dashes became the internet’s favourite AI tell because chatbots used them constantly, often where a comma or full stop would do. We couldn’t find any research showing they reliably identify AI text, so treat this one as folk wisdom with a grain of truth.
The grain of truth matters, though. When the em dash became a running joke, OpenAI’s chief executive Sam Altman publicly celebrated in November 2025 that ChatGPT would now follow a custom instruction to stop using them (as reported by Storyboard18). If the people who build the models are fixing it, your readers have definitely noticed it.
Our house style bans them outright. Not because em dashes are evil. Because they’ve become a distraction, and a distraction is the last thing you want when someone is deciding whether to trust you.
- Use a full stop instead. Short sentences read better anyway.
- Use a comma if the clause genuinely belongs.
- Use brackets for a true aside (sparingly, like this).
Why “humaniser” tools miss the point
Search for how to humanise AI text and the first page is mostly tools that promise to make AI text “undetectable”. They rewrite your copy until a detector stops complaining. That treats the symptom and ignores the reader.
Two problems with this.
Detectors aren’t the audience
OpenAI launched its own AI text classifier in January 2023. It correctly flagged only 26% of AI-written text and wrongly labelled 9% of human text as AI. OpenAI withdrew it in July 2023 for low accuracy. That’s old data, but it tells you how shaky “beat the detector” is as a goal.
Meanwhile, people are getting better at this. A 2025 University of Maryland study by Russell, Karpinska and Iyyer found that a panel of five people who frequently use ChatGPT for writing misclassified just 1 of 300 articles. They beat most commercial detectors, even when the text had been paraphrased or “humanised”. The experts didn’t just spot vocabulary. They also judged originality, formality and clarity.
In other words, your sharpest readers aren’t scanning for “delve”. They’re noticing that nothing in the piece could only have come from you.
Google doesn’t care who typed it
Google’s position has been consistent. Its Search Central guidance says using generative AI to produce many pages without adding value for users may break its scaled content abuse policy. That policy, introduced in March 2024, applies whether the content is made by automation, by humans or by both.
So the question was never “will Google detect the AI?” It’s “would anyone miss this page if it disappeared?” We’ve written more about that in what we tell every client who asks “can’t we just use AI?”.
| Humaniser tool | Prompt plus edit routine |
What it fixes | Detector scores | What the reader actually notices |
Adds experience or opinion | No | Yes, because you add it |
Fixes unsourced claims | No | Yes, minute five checks them |
Fixes structure and rhythm | Partly, often makes it worse | Yes |
Builds a reusable team standard | No | Yes |
Cost | A subscription | Five minutes per piece |
The best AI prompts for humanisation
Most of the robot voice disappears when you give the model a proper brief. Below are the four prompts we actually use, in the order we use them. The full versions, with every bracket explained, are in the free kit.
Before the prompts, five principles. They come from Anthropic’s and OpenAI’s own prompting guidance, and from a lot of trial and error on our side.
Principle | What it means in practice | Where it comes from |
Say what to do, not just what to avoid | “Write in short, plain sentences” works better than “don’t be wordy” | Anthropic prompting best practices |
Explain why a rule exists | “No em dashes, because readers now read them as AI” helps the model generalise | Anthropic prompting best practices |
Show, don’t describe | Three to five real examples of your writing beat a paragraph of adjectives | Anthropic prompting best practices |
Write the prompt in the style you want back | A prompt full of bullet points gets bullet points back | Anthropic prompting best practices |
Remove contradictions | “Be concise” next to “be thorough” makes output worse, not balanced | OpenAI GPT-5 prompting guide |
Prompt one: the humaniser prompt (for every new draft)
You are writing as [brand] for [reader], who already understands [what they know]. Before you write anything, ask me up to five questions about what you would otherwise have to guess: our opinion, a real example, a figure with a named source.
Match the voice of this sample. Copy its rhythm, not its topic: [paste 300 words of your best writing].
Rules:
- Start with the point. No opening that restates the title. No closing summary.
- Use plain words you would say out loud. Never use: delve, tapestry, testament, landscape, seamless, robust, pivotal, leverage, unlock, elevate, navigate, "in today's", "it's important to note", "when it comes to".
- Vary sentence length. Mostly short. The occasional longer one.
- State things directly. No "it's not X, it's Y" lines. No groups of three unless there really are three.
- Take a position. If the evidence is mixed, say which way you lean and why.
- Every statistic needs a named source and a date. If you don't have one, write [SOURCE NEEDED]. Never invent one.
- UK English. Use full stops and commas instead of em dashes, because readers now read em dashes as a sign of AI.
When you finish, list the three weakest sentences in your draft and rewrite them.
Two lines do most of the work. "Ask me up to five questions" forces the model to collect the specifics it would otherwise invent. "[SOURCE NEEDED]" stops it filling gaps with confident nonsense.
Prompt two: voice calibration (run once, reuse forever)
Most teams can’t describe their own voice. This gets you a usable description in two minutes, which then goes into prompt one, your project instructions or your skill.
Here are three pieces of writing we're proud of. Describe our voice in ten rules a new writer could follow: sentence length, vocabulary, humour, how we open, how we close, and what we never do. Back each rule with a short quote from the samples. Don't flatter us. Point out anywhere the samples contradict each other.
Prompt three: rewrite a draft you already have
Rewrite the draft below in our voice. Keep every fact, figure and source exactly as it is. Cut anything that restates the headline or summarises what was just said. Replace any word from our kill list with the plainest alternative. Where a sentence is vague, mark it [NEEDS DETAIL] rather than inventing one. Then list every change you made and why.
Prompt four: the sceptical editor (in a fresh chat)
Use a new chat so the model isn’t marking its own homework.
Act as a sceptical editor who hates generic copy. Don't rewrite anything. Give me a list of: every sentence that could appear unchanged in a competitor's article; every claim without a named source and date; every phrase from this kill list [paste it]; and three alternative opening lines, each more specific than the last.
If you want to go further on prompting, our prompting guide for marketers covers the habits worth unlearning.
Claude skills for humanising AI writing: build your own or borrow one
A prompt you paste every time is a habit. A skill is the same thing, saved, so it loads itself when the job needs it. If your team writes in Claude Code or Claude, a writing skill is the single biggest upgrade we’ve made to our own output.
A skill is a folder with a SKILL.md file inside. The file starts with a short name and description, then the instructions. Anthropic’s documentation explains that until a skill is triggered, only its name and description sit in Claude’s context. The rest loads when it’s needed. That’s what makes skills more efficient than pasting a 2,000-word brand guide into every chat.
How to build a humaniser skill in Claude Code
Time to get geeky. It's less geeky than it sounds.
01 Create the folder
For a personal skill, make:
~/.claude/skills/our-writing-voice/ For a skill your whole team shares through a code repository, use this inside the project:
.claude/skills/our-writing-voice/ 02 Write the description first
This is what decides when the skill switches on, so say what it does and when to use it. Anthropic's limit is 1,024 characters. You'll need far fewer.
03 Keep the main file short
Anthropic recommends keeping SKILL.md under 500 lines.
Put the long stuff (your full voice guide, your kill list, example pieces) in separate files in the same folder and point to them.
04 Add your best writing examples
Add three to five examples of your best writing as their own files.
Recent, edited, approved. Not your 2019 brochure.
05 Test it on a real brief
Test it on a real brief, then read the output with the five-minute routine below.
Every time it gets something wrong, fix the skill, not just the draft.
06 Using Claude in the browser or desktop app?
Zip the folder and upload it under Settings → Capabilities → Skills.
Anthropic lists skills as available on Pro, Max, Team and Enterprise plans, with code execution switched on.
--- name: our-writing-voice description: Writes and edits copy in [brand]'s voice. Use for any blog post, LinkedIn post, email, landing page or rewrite of an AI draft, and whenever someone asks to make writing sound less like AI. --- # Our writing voice 1. Read voice.md and kill-list.md before writing anything. 2. Ask up to five questions about anything you would otherwise guess: our opinion, a real example, a source. 3. Draft in the voice described in voice.md. Use the files in examples/ as the benchmark for rhythm and tone. 4. Check the draft against kill-list.md. Rewrite every sentence that contains a hit. 5. Every statistic needs a named source and date, or write [SOURCE NEEDED]. 6. Finish by listing the three weakest sentences and rewriting them.
YOUR FOLDER STRUCTURE
The folder ends up looking like this:
The best third-party humaniser skills
You don’t have to start from scratch. These are the open-source skills we’d look at first. Star counts are from GitHub as of October 2026 and change constantly.
Skill | What it does | Good for | Watch out for |
Checks text against 26 AI writing patterns based on Wikipedia’s “Signs of AI writing” guide, then rewrites, self-critiques and finalises. Around 57,000 stars, MIT licence | A strong first pass on any draft | It’s generic. It knows AI’s voice, not yours. Its own README says beating detectors isn’t the goal | |
Banned phrases, structural clichés and sentence rules, plus a score out of 50 across five dimensions. Around 18,000 stars, MIT licence | Teams who want a pass mark before publishing. Also works pasted into project instructions | Some rules are stricter than your brand may want. Edit them | |
Packages Strunk’s 1918 The Elements of Style as a “write clearly and concisely” skill | Tightening long, wordy copy | About 12,000 tokens of reference text. American and old-fashioned in places | |
Anthropic’s official examples, including doc-coauthoring, internal-comms, brand-guidelines and skill-creator | Using skill-creator to build your own voice skill properly | brand-guidelines applies Anthropic’s own brand. Treat it as a template | |
A curated directory of community skills | Browsing what exists | It’s a list, not a quality check |
Two rules before you install anything. Read the SKILL.md first, because skills are instructions (and sometimes code) that your AI will follow. And skip anything whose pitch is “undetectable”. That’s a detector-evasion tool, and we’ve covered why that misses the point.
Our setup, for what it’s worth: a third-party humaniser for the first pass, our own voice skill for the second, and a human for the third. The third pass is not optional.
How to set up your tone of voice in Claude, ChatGPT and Perplexity
Every major AI tool now lets you save instructions and reference files so you don’t start from zero each time. The names differ. The principle doesn’t: give it a short, current, high-signal picture of how you write.
It took us a long time to get here. We tried dozens of prompts, built and rebuilt our own writing skills, and kept feeding our projects tone-of-voice notes and memory until the output finally sounded like us. There was no single magic prompt. It was a system, and we had to build it one frustrating draft at a time.
Tool | Where your voice lives | What to put there | Useful limits and quirks |
Claude | Project instructions and project knowledge | Instructions: the ten voice rules from prompt two. Knowledge: voice guide, kill list, three to five recent examples | On paid plans, Claude switches to retrieval when project knowledge nears the context limit (Anthropic says up to 10x more content) |
Claude | Styles (Create custom style, then Add writing example) | Two or three of your best pieces | A quick win for one-off chats outside a project |
Claude | Skills (see above) | Your voice skill and a humaniser skill | Only the name and description load until needed |
ChatGPT | Settings, Personalization, Custom instructions | Your top ten voice rules and kill list, nothing else | 1,500 characters on Free and Go, 5,000 on paid plans |
ChatGPT | Personality presets | Pick the closest base tone (for example Professional, Candid or Efficient) | OpenAI says it doesn’t govern emails or code |
ChatGPT | Projects | Project instructions plus voice guide and examples as files | Project instructions override your global custom instructions. File limits run from 5 (Free) to 40 (Pro and business plans) |
ChatGPT | Custom GPTs | Avoid starting new ones | OpenAI plans to retire custom GPTs in favour of plugins, starting December 2026 for affected Enterprise workspaces |
Perplexity | Projects (which some of you will know as Spaces) | Instructions with your voice rules, plus your sources | Up to 8,000 characters of instructions |
Perplexity | Profile, in settings | Your preferred communication style and formatting | Best for research tone, not finished copy |
A practical note on Perplexity: we use it for research and sourcing, not final drafts. Set the voice there so summaries come back readable, then write in Claude or ChatGPT with your full setup.
Context dos and don'ts: why more isn't better
The instinct, once you discover project knowledge, is to upload everything. Brand book, old guidelines, every blog you’ve ever published, three competitor examples for “inspiration”. We did exactly this. It made things worse.
The mistake we made early on was thinking more context meant better output. We stuffed our projects with every brand document we had. It burned through tokens, the documents contradicted each other, and the copy got blander, not better. Now we keep it lean, update it every month, and feed it the copy we're proudest of. Your project should learn from your best recent work, not your oldest.
The research backs this up.
- Context is a budget. Anthropic’s engineering team describes context as “a critical but finite resource” with diminishing returns, and recommends the smallest possible set of high-signal information (September 2025).
- Longer inputs, shakier answers. Chroma’s “Context Rot” study tested 18 models and found performance grew increasingly unreliable as input length grew, even on simple tasks. Related-but-irrelevant text made it worse (July 2025).
- The middle gets lost. Researchers at Stanford and others found models use information best when it sits at the start or end of the input, and worst in the middle (TACL, 2024; the study dates from 2023).
- Contradictions hurt. OpenAI’s GPT-5 prompting guide warns that contradictory or vague instructions can be more damaging to its newer models than older ones.
Do | Don’t |
Keep three to five high-signal files: a one-to-two page voice guide, a kill list, three to five recent examples | Upload every brand deck, old guideline and blog archive “just in case” |
Put your most important rules first in the instructions | Bury the rules that matter in the middle of a 20-page document |
Update your project monthly. Date the voice guide and replace examples as your best work changes | Set it up once and wonder why it still writes like you did last year |
Feed it the edited, approved copy you’re proud of | Feed it raw AI drafts. It will learn its own bad habits back |
Delete anything that contradicts your current voice | Keep the 2019 guidelines next to the 2026 ones and hope it picks the right one |
Pair every ban with what to do instead | Write a list of 40 “nevers” and nothing it should do |
Keep one project per brand or client | Mix three clients’ voices in one project |
Check what you upload is safe to share with the tool under your organisation’s policy | Paste in client data, contracts or anything confidential without checking |
The five-minute edit routine
A good prompt, skill and project setup get you most of the way. Editing gets you the rest. This routine takes five minutes on a 1,000-word draft. We time it. It genuinely fits.
- Minute one: search and destroy. Ctrl+F your kill list. Delete or replace every hit. Don’t swap “delve” for “dive into”. Rewrite the sentence.
- Minute two: cut the top and tail. Delete the first paragraph and the last paragraph. Read it again. Nine times out of ten, nothing is missing.
- Minute three: make one thing specific. Replace one vague example with a real one. Add one detail only your team would know: a client conversation, a number from your own data, a mistake you made.
- Minute four: read it out loud. Wherever you stumble or run out of breath, cut. Break up any run of same-length sentences. Kill any triplet that isn’t earning its place.
- Minute five: check the claims. Every number has a named source and a date. Every quote has a speaker. Anything you can’t verify goes.
Five minutes of editing beats five subscriptions. The tool can rewrite the words. Only you can add the bit that makes it worth reading.
Instead of this, try this
Instead of | Try |
“In today’s ever-evolving digital landscape, brands must…” | Start with the problem: “Your traffic is down 30% and nobody can say why.” |
“This plays a pivotal role in driving growth.” | Say what it actually did: “This page brings in a third of our demo requests.” |
“It’s not just a tool, it’s a partner.” | Pick the one that’s true and say it plainly. |
“Studies show that 70% of consumers…” | Name the study, the year and the sample, or cut it. |
“Whether you’re a startup or an enterprise…” | Write for one reader. Everyone else will cope. |
What this means for your team (and what to do this week)
One person editing well is a habit. A whole team editing well is a system. If AI is part of how your team writes, four things make the difference between occasional good copy and consistently good copy.
- A shared kill list. One list, in one place, baked into every prompt, project and skill.
- A shared voice setup. One voice guide and one example bank, used in every tool, rather than five people with five different custom instructions.
- A named editor. Someone owns the final read. “We all kind of check it” means nobody does.
- A monthly refresh. Someone owns updating the project and the skill with your latest best work.
This matters most in regulated and mission-led sectors. A charity appeal, a healthcare explainer or a professional services proposal that reads like AI doesn’t just underperform. It quietly signals that nobody with expertise was involved. That’s the opposite of the trust you need, and we’ve argued before that AI content has made your experts more valuable, not less.
It also matters for discovery. AI answer engines increasingly cite content that’s specific, sourced and clearly authored. Generic copy is easy to summarise and easy to ignore. Our generative engine optimisation services are built around making your content the version worth citing.
If you’d rather fix the whole process than patch individual drafts, our AI-powered content systems work does exactly that. We’ve also written about why a better content system beats more content.
Action plan
Eight practical actions to make your AI-assisted writing sound more like you.
Keep improving the system, not just the next draft.
Frequently asked questions
Clear answers to common questions about AI writing, humanisers and finding your brand voice.
How can you tell if something is written by AI?
Look for clusters, not single words. Stock vocabulary (delve, testament, seamless), "it's not X, it's Y" lines, everything in threes, uniform sentence length and confident claims with no named source tend to appear together. One tell proves nothing. Five in a paragraph is a strong signal.
What words does AI overuse?
Research on 15.1 million PubMed abstracts found "delves" used about 28 times more often than expected in 2024, with "underscores" and "showcasing" also surging (Kobak et al., Science Advances, 2025). In marketing copy, watch for tapestry, testament, landscape, seamless, robust, pivotal, leverage, unlock and elevate.
Do AI humaniser tools work?
They can change detector scores, but they don't fix what readers notice. A 2025 University of Maryland study found that frequent ChatGPT users spotted AI text with only 1 error in 300 articles, even after the text had been "humanised". A better prompt, a voice setup and a human edit fix the underlying problem.
How do I make ChatGPT or Claude write in my tone of voice?
Give it a short voice guide, a kill list and three to five examples of your best recent writing. Put them in a project (Claude or ChatGPT) or a skill (Claude), not pasted into every chat. Keep it lean, because too much context and contradictory documents make output worse. Update it monthly with your latest approved copy.
Does Google penalise AI-written content?
Not for being AI-written. Google's guidance targets content produced at scale without adding value for users, whether a human, a machine or both made it. Helpful, accurate, original content is fine however it was drafted. Thin, mass-produced pages are the risk.
Get the full humaniser kit.
This article covers the highlights. The free kit has the complete set:
What's inside?
EXPLORE THE KIT ↓The four humanisation prompts in full
A ready-to-use SKILL.md template
A setup checklist for your AI projects
60 AI writing phrases and patterns
The context dos and don'ts
The five-minute edit routine
A one-page team style sheet
Ready to make your AI writing sound like you?
Get the complete toolkit in one free PDF.
Free PDF resource
Need something more tailored?
If you'd like help building a content system that sounds like you at scale, talk to us.
Sources
- Raptive: The AI stink is real and it’s costing brands – survey of 3,000 US adults on trust, authenticity and emotional connection when content is suspected to be AI (Aug 2025, vendor research)
- Pew Research Center: AI and its impact – 76% say telling AI from human content is important; 53% not confident they could (Sept 2025)
- Reuters Institute Digital News Report 2025 – 11% of UK respondents comfortable with news made automatically by AI (June 2025)
- BetterUp Labs and Stanford Social Media Lab: Workslop – prevalence, time cost and perception of senders of low-effort AI work (Sept 2025)
- Kobak et al., Delving into LLM-assisted writing in biomedical publications through excess vocabulary – at least 13.5% of 2024 PubMed abstracts LLM-processed; excess frequency of “delves” (Science Advances, 2025)
- Yakura et al., Empirical evidence of large language model’s influence on human spoken communication – rise of ChatGPT-favoured words in spoken academic talks (Max Planck Institute for Human Development, July 2025)
- Russell, Karpinska and Iyyer, People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text – expert panel misclassified 1 of 300 articles (ACL 2025)
- OpenAI: New AI classifier for indicating AI-written text – 26% true positive rate, 9% false positives, withdrawn July 2023 (older than 24 months)
- Google Search Central: Using generative AI content – scaled content abuse and accuracy guidance (updated Oct 2026)
- Google: March 2024 core update and new spam policies – scaled content abuse applies to automated, human or combined content (Mar 2024, older than 24 months, still in force)
- Storyboard18: Sam Altman celebrates no em dash update – reported comments on ChatGPT following “no em dash” instructions (Nov 2025, secondary source)
- Anthropic: Claude prompting best practices – say what to do, explain why, use examples, match prompt style to output
- OpenAI: GPT-5 prompting guide – contradictory or vague instructions are especially damaging (Aug 2025, now archived)
- Anthropic: Agent Skills overview – SKILL.md structure, description limits, progressive loading
- Claude Code docs: Skills – personal and project skill folders, keep SKILL.md under 500 lines
- Claude Help Center: Using skills in Claude – plans, uploading skills as a ZIP
- Claude Help Center: RAG for Projects – retrieval when project knowledge nears the context limit
- Claude Help Center: Styles – custom styles from writing examples
- OpenAI Help: Custom instructions – 1,500 and 5,000 character limits
- OpenAI Help: Personality – base style and tone presets
- OpenAI Help: Projects in ChatGPT – project instructions, file limits, project-only memory
- OpenAI Help: GPT builder – planned retirement of custom GPTs in favour of plugins
- Perplexity Help Center: What are Projects – 8,000-character instructions and file limits
- Perplexity Help Center: Profile settings – communication style and formatting preferences
- Anthropic: Effective context engineering for AI agents – context as a finite resource with diminishing returns (Sept 2025)
- Chroma: Context Rot – 18 models grow less reliable as input length grows (July 2025)
- Liu et al., Lost in the Middle – models use information in the middle of long inputs worst (TACL, 2024; 2023 study, older than 24 months)
- blader/humanizer, hardikpandya/stop-slop, obra/the-elements-of-style, anthropics/skills, travisvn/awesome-claude-skills – third-party skills reviewed (GitHub, October 2026)
