AI writing's #1 offender

Why AI writing consistently misses the mark (and how to fix it)

I love writing.

I’ve written for local magazines, for travel blogs, I’ve been a ghostwriter for lists of 50,000+ subscribers.

I’ve written weekly emails to my own email list for over eight years with very few breaks.

All that is true AND I love using AI to support my writing process.

I provide the ideas, takeaways, structure, and details about the reader and how I want them to feel.

It helps me draft something quickly that I can then edit.

I use it as a thinking partner and then a drafting partner before I finalize things.

The problem is that AI consistently takes out the very things that help my writing feel like me and connect with the people I want to connect with.

I’ve spent hours dissecting AI drafts, asking why a sentence doesn’t work and what about it fell off. Figuring out the constructions that make writing sound flat and robot-esque.

I wanted to get to the bottom of why it consistently removes meaning from my words.

Why does it do that?

There are thousands of lists now telling you how to spot AI writing.

The em-dash.

Three adjectives in a row.

“Here’s the thing…”

“It’s not just a X, it’s a Y.”

“Once you can see it, you can’t unsee it.”

It’s true that people tend to recognize those patterns as AI.

They show up in AI writing partly because they’ve been used so much by humans, and AI was trained on data that contained those patterns in human writing.

And while those AI tells are easy to spot, there’s a bigger factor at play and AI writing’s #1 offender:

Compression.

Editing vs compression

Editing is good. Being concise is good. Cutting is most of what good writing is. You write the long version, then take things out until only what’s necessary and compelling is left.

Editing is where you take a sentence, a paragraph, an article, and distill it down to its core sentiment without losing important meaning.

But AI often doesn’t edit the way a skilled human editor does.

It compresses.

It often reduces specificity, which feels like meaning loss.

Think about what happens to a photo when you compress it into a smaller file.

At a glance it looks the same.

Zoom in, and the edges are mushy, the fine detail is gone, and there are little smudges where information used to be.

The file is smaller and faster to send. But it also loses nuance and precision.

AI does something similar to writing.

It gives you a smaller, faster version that looks right at a glance but, on a closer look, has replaced details with generic AI-isms.

That’s why so much AI writing feels like it says very little.

It often sounds like a complete thought without complete thinking.

It’s very good at copying what a real point sounds like even when it’s not actually making the point.

A lot of AI writing follows the structure of insight:

• Setup.

• Contrast.

• Resolution.

• Takeaway.

The structure itself is fine. Good writing often uses exactly the same pattern. But, the problem is that sometimes there is no underlying observation.

You get the resemblance of an insight without the insight itself.

Why “it’s not X, it’s Y” shows up everywhere

Of all the patterns AI uses, contrast is one of the most common (and infuriating).

“It’s not about the tools, it’s about the system.”

“This isn’t productivity, it’s clarity.”

This is where you say what something isn’t in order to define what it is.

Sometimes it’s inane:

❌ a system that works with your brain, not against it

✅ a system that works with your brain

Sometimes it’s helpful:

✅ This isn’t a time management problem. It’s a prioritization problem.

AI doesn’t reliably distinguish between useful contrast and unnecessary contrast.

For contrast to be helpful, it has to introduce real informational separation between two ideas. It has to change what the reader understands or what they would do next.

AI didn’t invent this pattern.

Contrast is common in persuasive writing because it creates the feeling of clarity. It gives people simple categories to organize ideas into and makes a statement sound more decisive than it might actually be.

That’s one reason AI uses it so much.

The sneaky thing about it is that it often introduces an opposite that wasn’t necessary in the first place.

And what has me rage typing (or ironically raising my voice via Wispr Flow):

Unnecessary, unsolicited contrast introduces an opposite that was never there to begin with and can bring doubt and negativity with it.

“You’re not broken, you’re learning.”

For some people, maybe that feels reassuring.

For others it’s more like:

“Well thank you very much Claudia, I didn’t even consider I was broken but now you’ve introduced that fun thought into the conversation. 🙄”

You raise a negative the reader was never even thinking about.

AI uses contrast constantly, and every time it does, it can override what you meant with random constrast.

The disconnect between you and your reader

Your reader might not be able to put her finger on any of this. She isn’t counting em-dashes or adjectives. But she reads AI all day now whether she knows it or not, and her brain has learned the patterns.

Not necessarily enough to identify AI. But often enough to recognize when something feels generic.

So when her experience with your writing follows the same patterns she sees everywhere else, there is no distinction. She can’t tell your writing apart from everything else she scrolls past.

She feels a small step back, a little less trust, a little less of you in your words.

There are two ways to get around this while still using AI to write your stuff.

1. Say things other people are not saying.

If your content is differentiated by what you’re saying, you can get away with more AI-isms. When a piece is distinctly you because of the stance you’re taking or the story you’re telling, that will be what your reader trusts.

2. Train your AI to compress less.

You can teach it to not do the thing in the first place. Let me tell you how.

Train your AI to compress less

If you find your writing keeps getting flattened when you use AI, there’s a simple way to test what’s going on before you overhaul anything.

Start small. Drop this into the chat you’re working in:

Do not remove or generalise:
- examples or concrete instances
- conditions or qualifiers (time, frequency, probability, context)
- causal or explanatory links
- distinctions between ideas or scenarios

Do not convert specific statements into abstract summaries.

Do not replace multi-part explanations with single general claims.

Only use contrast (“X vs Y”, “not X but Y”) when each side introduces distinct information that changes interpretation or action. If both sides do not add new information, remove the contrast and rewrite as a single statement.

If a change reduces informational content, do not apply it.

This is usually enough to show you when the model is defaulting into that “compressed” version of your thinking.

If it helps, the next step is to make it more permanent.

Put it somewhere the model consistently reads it at the start of a conversation — Project Instructions, Custom Instructions, a Skill/GPT/Gem.

That’s what keeps it from resetting every time you start a new thread.

And when you notice it drifting, a small reminder usually works:
“You’re compressing again, check the rules.”

It’s surprisingly responsive to that kind of nudge.

Below is the full instruction set you can use for that system-level setup.

Let me know how it goes.

Amy x

## Objective

Preserve informational content, specificity, and relationships between ideas during transformation.

Avoid loss of meaning through abstraction, over-summarisation, or unnecessary generalisation.

---

## 1. Compression definition (operational)

Compression occurs when transformation:

* removes specific details without replacement
* replaces concrete elements with general categories
* deletes qualifiers (conditions, frequency, uncertainty, scope)
* collapses multi-step explanations into single abstract claims

Compression is defined by **loss of informational distinctions**, not by length.

---

## 2. Information preservation rule

Preserve informational distinctions present in the input.

Informational distinctions include:

* examples or instances
* constraints or conditions
* causal or explanatory links
* contrasts between ideas
* exceptions or edge cases
* temporal or probabilistic qualifiers

Do not merge distinct informational elements into a single general statement.

---

## 3. Anti-abstraction rule

Do not replace specific content with abstract summaries.

Invalid transformations:

* example → category
* scenario → principle without instance
* multi-step explanation → single general claim

If abstraction is used, it must retain at least one concrete reference.

---

## 4. Contrast rule

Use contrast only when both sides contribute distinct informational content.

A contrast is valid only if:

* each side adds non-overlapping information
* removing either side changes meaning or interpretation

Otherwise:

* remove contrast structure
* rewrite as a single statement preserving all information

---

## 5. Redundancy rule (primary optimisation rule)

You may remove content only when it does not change informational meaning.

Allowed removal:

* repeated meaning with no added specificity
* rephrasing that introduces no new constraint or distinction
* filler phrases with no informational role

Do not remove content that changes meaning, scope, or interpretation.

If uncertain, keep the more specific version.

---

## 6. Structure constraint

Avoid fixed rhetorical templates unless they reflect real informational progression.

Invalid default pattern:
setup → contrast → resolution → takeaway

Use only if each step adds new information rather than restating prior content in different form.

---

## 7. Output constraint

Prefer:

* explicit relationships over implied ones
* concrete statements over abstract summaries
* retention of qualifiers when they affect meaning

Do not simplify in a way that removes informational distinctions.

---

## 8. Validation check (single pass)

Before final output, check:

* Are any informational distinctions removed without replacement?
* Are any concrete elements replaced with abstractions?
* Are any contrasts used without informational gain?
* Has any multi-step logic been collapsed into a single claim without loss of structure?

If yes → revise only the affected parts.

---

## Priority hierarchy

Information preservation > correctness of relationships > concision > stylistic uniformity