How to Use AI to Write Without Sounding Like Everyone Else
A practical guide to drafting with AI tools while avoiding LLM clichés, based on workflows using voice dictation, Codex, and Claude.
How do you use AI to write without sounding like an AI wrote it?
The short answer: start with your own unfiltered thinking, feed that into the model instead of asking it to invent ideas, and then treat the draft as raw material to argue with rather than a finished product. The core failure mode isn’t that people use AI to write. It’s that they skip the step where they have to believe in what they’re saying. Writers who keep their voice intact tend to front-load a session with messy, personal, spoken-out-loud thinking, then spend real time stripping out the phrasing patterns that make a draft feel machine-generated.
This has become a live problem across the internet, not just a writing-nerd concern. Substack CEO Chris Best has pointed to third-party detection data (from the AI-text detection company Pangram) suggesting that around 40 percent of long-form writing on LinkedIn shows signs of being fully AI-generated. That number isn’t a perfect measure of quality or effort. Plenty of AI-assisted writing is thoughtful, and plenty of human writing is lazy. But it signals a real shift in how much of what people read online was actually thought through by a person.
TL;DR
- AI slop isn’t defined by tool use, it’s defined by intent: content nobody actually believes in, whether a human or a model produced it.
- Detection tools like Pangram can estimate whether text passed through an LLM during generation, and Substack has built this directly into its app so readers can check any piece of long-form writing.
- LLMs converge on a detectable style shaped by pretraining and reinforcement learning, which is exactly what detectors are trained to spot, and why generic AI writing tends to read the same regardless of topic.
- A practical anti-slop workflow starts with voice, dictating a long, unstructured transcript of your actual thinking before any model touches the piece, so the draft inherits your reasoning instead of the model’s.
- Editing against LLM-isms is a separate pass from idea development, meaning writers who care about voice spend real, deliberate effort removing dramatic contrasts, false urgency, and stock phrasing after the thinking is already solid.
- Transparency tools don’t ban AI use, they let readers see how something was made and decide for themselves what they’re comfortable with, shifting the norm toward disclosure rather than prohibition.
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
Why does AI-written text sound the same no matter who uses it?
Large language models are trained on enormous amounts of text and then shaped further through techniques like reinforcement learning from human feedback. That process doesn’t just teach a model facts, it nudges it toward certain patterns of phrasing, sentence rhythm, and rhetorical habits that reviewers or reward models tended to favor during training. The result is a kind of statistical center of gravity in how models write: dramatic before-and-after contrasts, headline phrasing that oversells stakes, and a recognizable cadence that shows up regardless of the topic or the person prompting it.
This is precisely what detection tools exploit. Pangram, for example, builds its detection around identifying that stylistic fingerprint rather than checking facts or sourcing. It’s also why the fingerprint can be reduced. Fine-tuned models built specifically to sound “more human” have been shown to score much lower on these detectors, which means the tell isn’t some unavoidable property of machine text. It’s a byproduct of how most models are trained and used by default, and it can be deliberately edited or trained away.
What’s the actual difference between using AI and gaming the system?
The distinction that keeps coming up in conversations about AI and online writing isn’t tool versus no-tool. It’s whether the person using the tool is still exercising judgment. Best draws the line this way: using AI to help express something you already believe and have thought through is one thing. Telling a model to generate a thousand posts with no mistakes and publishing them without reading them is another. The first is a writer using a paintbrush. The second is closer to an automated flood aimed at gaming search engines or platforms, with no human intent behind it at all.
The practical consequence isn’t just that some AI content is low quality. It’s that once readers can’t tell which pieces involved real thought, they start discounting everything, including work that took real effort. People have reported abandoning comment sections entirely after realizing a “conversation” they’d been having for a while was with a bot set up to auto-reply and pad an account’s engagement. That erosion of trust is the actual cost of unchecked slop, more than any individual bad post.
What does a voice-preserving AI writing workflow actually look like?
One workflow described by a writer working in this space starts with dictation, not typing. Before touching any AI tool, the writer records a long, unbroken spoken transcript, sometimes 10 to 15 minutes, using a voice-to-text tool (Whisper Flow was mentioned specifically). The goal is to get the full shape of an idea out of their head before any model reshapes it. Modern models handle that much input without trouble, so length isn’t a limiting factor the way it might have been with earlier context windows.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
That transcript, in the writer’s own words and structure, becomes the seed the model works from, rather than a bare prompt asking the model to generate an opinion from scratch. From there, the process splits into two distinct kinds of work:
Idea development. This is about whether the core argument or concept is actually clear, whether it says something the writer means, and whether it lands the way they intend. This part gets argued over, sometimes across many rounds, until the thesis is sharp.
Prose cleanup. This is a separate pass focused specifically on removing LLM-isms once the ideas are already solid. That means cutting dramatic contrast structures, inflated headlines, the model’s tendency to narrate its own reasoning (“Let’s break this down”), and other stock patterns that make text instantly recognizable as machine output. The target isn’t literary flourish. It’s plain, clear, workmanlike prose that reads like something a competent person actually wrote and meant.
This same writer maintains a large set of custom instructions, described as more than 20 different “skills” used across tools like Codex and Claude, specifically built around identifying and removing these patterns. The point isn’t to hide AI involvement. It’s to make sure the AI involvement serves the writer’s actual voice instead of replacing it.
Is disclosure the right fix for AI slop, or does it need stronger rules?
Substack’s initial response to this problem has been transparency rather than restriction. The platform integrated Pangram’s detection directly into its app, letting anyone check a long-form piece and see an estimate of whether it was generated through an LLM. Writers also get the option to add a note describing their own process. Nothing is blocked or banned. The idea is to put the fact on the table and let readers form their own opinion about what it means for them.
Whether disclosure alone is enough is an open question. Best has framed this explicitly as a starting point, not a finished policy, on the reasoning that platforms don’t have the luxury of waiting to see how things shake out while usage patterns shift quickly. The risk he’s named directly: a major platform could start to feel flooded with generic content within a short window if nothing changes. Transparency tools are a first move, not a guarantee, and they depend on enough writers and readers actually caring enough to use them.
Frequently Asked Questions
What is AI slop, exactly?
It refers to content, AI-generated or not, that nobody actually believes in or put real thought into. Chris Best describes it broadly as spam, clickbait, and generic material presented as if it were carefully made when it wasn’t. Not all AI-assisted writing counts as slop, and not all slop involves AI.
How does Pangram detect AI-generated text?
Pangram estimates whether a piece of text passed through a large language model during generation, largely by identifying statistical patterns in phrasing and style that LLMs tend to converge on due to their training process. It measures likely generation method, not quality, accuracy, or how much a human edited the result.
Can AI detectors be fooled or bypassed?
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
Yes. Models fine-tuned specifically to write in a more natural, less “AI-typical” style have produced much lower detection scores. This shows the detectable pattern is a byproduct of common training and generation choices, not an unavoidable signature of machine-written text.
Does using AI to write mean the writing is automatically slop?
No. The distinguishing factor described by practitioners in this space is intent and engagement with the material. Someone who dictates their real thinking, argues with a model over clarity, and edits out generic phrasing is using AI as a drafting tool. Someone who asks a model to mass-produce content unread is doing something categorically different.
What’s a practical first step to keep AI-assisted writing sounding like you?
Start by talking through your idea out loud and transcribing it, rather than asking a model to generate the idea itself. That gives any AI tool your actual voice, reasoning, and structure to work from, which makes the later editing pass about sharpening your thinking instead of erasing someone else’s.
