Here's a thought experiment I couldn't resist: what happens when you use an AI to roast the writing of other AIs?
I spent an afternoon collecting 30 real outputs from ChatGPT, Gemini, and Claude — blog intros, LinkedIn posts, email drafts, product descriptions — and ran every single one through an AI text roaster. The feedback was merciless. And weirdly, it was the same four complaints every single time.
Before you assume this is an AI vs. AI pile-on — it's not. The roasts were genuinely useful. They identified patterns in AI writing that humans have complained about for years but rarely articulated clearly. This is what the experiment revealed.
I kept it simple. I used the same prompt template for all 30 tests: paste AI output, hit roast, read the feedback. No cherry-picking — I ran everything in sequence, kept every result. The sample broke down like this:
Content types: 8 blog post introductions, 9 LinkedIn posts, 7 cold email drafts, 6 product descriptions. All generated with typical, "good" prompts — the kind people actually use, not edge cases designed to fail.
| AI Model | Avg. Roast Score | Most Common Complaint | Hardest Hit Content Type |
|---|---|---|---|
| ChatGPT (GPT-4o) | D+ | Over-structured, bullet overkill | LinkedIn posts |
| Gemini 1.5 Pro | C | Hedging language / passive voice | Cold emails |
| Claude 3.5 Sonnet | C+ | Overly polite openers | Blog introductions |
No AI "won." They all had real, recognizable weaknesses that the roaster flagged consistently. Here's what it found.
Every AI model opened at least half its blog intros and LinkedIn posts with a wide-angle establishing shot. "In today's fast-paced digital landscape..." / "As AI continues to transform industries..." / "In an era defined by rapid change..."
"In today's competitive business environment, email marketing remains one of the most powerful tools in a marketer's arsenal. As inboxes become more crowded than ever, standing out requires more than just a catchy subject line."
"Congratulations, you've written the opening paragraph of literally every marketing blog post published between 2019 and now. 'In today's competitive environment' is the literary equivalent of clearing your throat for 40 words. Your reader already knows the environment is competitive — they live in it. Start with something they don't know."
Lead with the counterintuitive thing, the specific number, or the uncomfortable truth. "Email open rates dropped 34% last quarter for every industry except one." That's a first sentence. The other thing is table-setting.
ChatGPT was the worst offender here, but all three models defaulted to bullet lists when the content didn't need them — especially in LinkedIn posts, which are supposed to feel personal and conversational.
"Remote work has taught me 5 things:
• Communication is everything
• Trust your team
• Set clear boundaries
• Celebrate small wins
• Prioritize mental health
What would you add? 👇"
"This is a listicle wearing a LinkedIn post costume. 'Communication is everything' and 'trust your team' are so vague they could appear in a motivational poster at a dentist's office. You have learned nothing from remote work that hasn't appeared in a 2020 Forbes article. Where's the moment? Where's the specific thing that happened to you? The engagement bait question at the end cannot save this."
Tell the one specific story instead of five generic bullets. The weird Tuesday when your kid interrupted your board call and the CEO laughed and it changed how you thought about formality — that post will outperform this one 10x.
This one was subtle but the roaster caught it every time: AI writing systematically avoids taking a position. It presents "both sides" on things that don't have two equally valid sides. It uses phrases like "some might argue," "it's worth considering," and "results may vary" to avoid saying anything definitive.
"Our platform might be a good fit for teams looking to potentially streamline their workflow processes. Many of our customers have found value in the various features we offer, which could help address some of the challenges you may be facing."
"'Might be a good fit.' 'Potentially streamline.' 'Could help.' 'May be facing.' You have packed more uncertainty into 39 words than a weather forecast for a town that doesn't exist. This isn't a pitch — it's a legal disclaimer that accidentally formatted itself as an email. If you don't believe your product solves a real problem for a real person, why would they?"
Make one specific claim and stand behind it. "Teams using this cut meeting prep time by 40%. If that number sounds useful, it's worth 15 minutes." Confident, specific, easy to disprove — which paradoxically makes it believable.
Claude's outputs got flagged most for this. AI writing often front-loads warmth — "Great question!", "Thank you for asking about this important topic!", "I'd be happy to help with that!" — and that habit bleeds into generated content as over-long, over-polite setups before the actual point arrives.
"Productivity is something we all think about, and it's a topic that touches nearly every aspect of our working and personal lives. Whether you're a seasoned professional looking to optimize your output or someone just starting to explore the world of personal efficiency, there's always more to learn about how we can make the most of our time."
"'Whether you're a seasoned professional OR someone just starting out' — you have now successfully included every human being who has ever existed. This opening is so inclusive it has excluded all meaning. You've spent 58 words saying 'productivity matters to people.' The reader knows. They clicked the link. Skip the waiting room and take them straight to the thing."
Assume your reader is smart and pressed for time. Your first sentence should be the thing they'd quote if they were recommending your article to a friend. Everything before that is dead weight.
No. But it means unedited AI writing has a tell — and the tell is getting more obvious, not less. Readers have developed a kind of AI-prose fatigue. The wide-angle openers, the hedge language, the bullet avalanche — these patterns are now associated with content that wasn't written for anyone in particular.
The interesting part of the roasting experiment was that the best AI outputs in the batch — the ones that scored highest — were ones where the original prompt was highly specific. "Write a cold email to a 40-person fintech startup's Head of Operations about a tool that cuts vendor invoice processing time in half, written in the tone of a direct founder" performed dramatically better than "write a cold email about our SaaS product."
The roaster isn't just a vanity mirror. It's a prompt diagnostic. If your AI output is getting destroyed, the problem is usually upstream in how you asked for it.
The workflow that made these experiments useful:
The "feed the roast back" step is underrated. Giving the AI the explicit critique of its own output as part of the next prompt produces noticeably better second drafts than just saying "make this better."
Paste any AI output, email draft, LinkedIn post, or blog intro. See exactly what's killing it — and get specific fixes, not generic advice.
Roast Your Text Free →The AI-roasts-AI experiment felt a bit recursive and absurd going in. What came out was actually useful: a clear taxonomy of the failure modes baked into current LLM writing defaults.
The four patterns — throat-clearing openers, bullet avalanches, epistemic cowardice, and politeness tax — aren't random bugs. They're features of how these models were trained to be helpful, safe, and comprehensive. Those are good goals for an assistant. They're terrible goals for writing that needs to hold a reader's attention.
The roaster is free. Paste something you generated this week. There's a reasonable chance it tells you something you'd rather know now than after you hit publish.