I Gave Three AI Models the Same Bad Paragraph. One of Them Made It Sound Like a Robot Wrote a LinkedIn Post.

I took a genuinely rough draft — a 200-word intro to a newsletter I actually publish — and fed it into ChatGPT, Claude, and Gemini with the exact same prompt: "Fix this writing." Two of them "fixed" it by erasing every trace of the person who wrote it. Only one understood that editing isn't the same as rewriting. If you use AI to edit your writing (and you probably should), this is the difference that determines whether people can still tell it's you — or whether you sound like every other AI-generated post on the internet.

The Same Prompt Gave Me Three Completely Different Writers

Here's the draft I used. It's a real excerpt, slightly messy, with sentence fragments and a casual tone: "So I tried this new productivity thing for like 3 weeks and honestly? Not sure it worked. But something shifted in how I think about mornings now, which is weird because I didn't expect that."

I gave all three the same instruction: "Fix this writing so it's clearer, but keep my voice."

ChatGPT smoothed it into something polished but generic: "After three weeks of testing this productivity method, I'm not entirely convinced of its effectiveness. However, it has notably changed my perspective on mornings — an unexpected outcome." Grammatically perfect. Completely lifeless. It reads like a press release, not a person.

Gemini went even further in the wrong direction, adding structure I never asked for — bullet points, a bolded "Key Takeaway," and a summary sentence at the end. It didn't just edit my paragraph. It turned a casual thought into a corporate slide.

Claude did something different. It kept the fragments. It kept "honestly?" as its own beat. It cleaned up the run-on without flattening the tone: "I tried this productivity thing for about three weeks. Honestly, not sure it worked — but something shifted in how I think about mornings, which caught me off guard." Same voice. Just clearer.

Why This Happens: AI Models Have Different Default "Editing Philosophies"

Here's the part nobody explains: these three models aren't just differently trained — they have fundamentally different assumptions about what "good writing" means, and that assumption kicks in the moment you ask for an edit.

ChatGPT defaults to "professional." Unless you explicitly fight it, it pulls every sentence toward business-email register — complete clauses, formal transitions, zero fragments. This is great if you're writing a cover letter. It's terrible if you're writing a text-message-style caption or a punchy hook.

Gemini defaults to "structured." It wants to organize, summarize, and add scaffolding — headers, bullets, bold takeaways — even when you didn't ask for any of that. This comes from Gemini being trained heavily around search and information-retrieval tasks, where structure equals helpfulness. But your Instagram caption doesn't need a TL;DR.

Claude defaults to "preserve intent." It treats your original phrasing as data about how you think, not just words to improve. It asks (implicitly, in how it processes the prompt): what was this person actually trying to say, and how would they say it if they had five more minutes? That's a completely different editing philosophy — closer to what a good human editor does.

This isn't a random quirk. It's baked into how each company designed its assistant's personality, and it means your choice of model matters more than your choice of prompt for voice-sensitive writing.

How to Actually Use This Today

Stop using "fix this" as your edit prompt. It's too vague, and it lets the model default to its built-in style preference instead of yours. Here's the workflow that actually works.

Step 1: Tell the AI what to preserve, not just what to fix. Instead of "fix this writing," use: "Edit this for clarity and grammar only. Do not change my sentence structure, tone, or word choice unless it's actually incorrect. Keep every fragment and casual phrase that isn't a grammar error." This single instruction change fixes 80% of the "robot voice" problem across all three tools.

Step 2: Match the tool to the task. For casual, voice-heavy writing — captions, personal essays, newsletters — start with Claude. For formal writing where you want polish — cover letters, business proposals, reports — ChatGPT's default professionalism actually works in your favor. Save Gemini for when you genuinely want structure added, like turning messy notes into an organized outline.

Step 3: Do a "voice check" before you accept any edit. Read the AI's version out loud. If you wouldn't say it that way to a friend, reject it and ask the model to try again with: "That sounds too formal/structured. Make it sound like I'm talking, not writing a report."

Step 4: Feed it your own past writing first. Before asking for an edit, paste in 2-3 paragraphs you've written that you're happy with, then say: "This is my writing style. Study it. Now edit the following paragraph to match this voice, not your default style." This works especially well in Claude and ChatGPT, and it cuts the generic-AI-voice problem dramatically.

The Part Most People Get Wrong

Most people think "AI writing sounds fake" is a prompt problem. It's not — it's a default behavior problem, and no prompt fully overrides it unless you name the behavior explicitly.

People type "make this better" and then wonder why their casual Twitter thread suddenly reads like a McKinsey report. The AI isn't guessing wrong — it's doing exactly what it was trained to consider "better," which usually means more formal, more complete, more structured. That's the model's definition of quality, not yours.

The fix isn't a magic prompt. It's understanding that you have to explicitly define what "good" means for this specific piece of writing — casual and fragmented, or polished and formal — because the AI will never assume casual is intentional. It assumes messy means unfinished.

This is why the same person can get wildly different results from the same tool on different days: one day they say "fix this," the next they say "fix this, keep it casual, don't touch my fragments" — and only the second version actually sounds like them.

Key Takeaways

• Vague prompts default to formal: "Fix this" almost always pulls your writing toward corporate, polished, lifeless — across all three tools.
• Claude preserves voice best by default: It treats your phrasing as intentional rather than something to "correct" automatically.
• ChatGPT is best for formal writing: Its built-in preference for professional tone works in your favor for emails, proposals, and cover letters.
• Gemini adds structure you didn't ask for: Great for organizing messy notes, risky for anything meant to sound conversational.
• Name what to preserve, not just what to fix: Explicitly telling the AI to keep your tone, fragments, and phrasing changes the output more than any other prompt tweak.

What to Do Right Now

Open Claude or ChatGPT right now, paste in something you wrote that felt slightly rough, and use this exact prompt: "Edit this for clarity only. Do not change my tone, sentence structure, or word choice unless it's grammatically wrong." Compare it to what you'd get from just typing "fix this" — the difference will show you immediately why the instruction matters more than the tool.