Why ChatGPT Can’t Stop Saying “Delve Into”
AI writing has gotten complicated with all the filler and corporate noise flying around. “Delve into.” “It is worth noting.” “In today’s fast-paced world.” I started using ChatGPT for client work sometime around early last year — paid for the Plus subscription, $20/month — and within three days I was ready to throw the whole thing out. Every response sounded like it had been drafted inside a 2003 HR department.
But what is actually happening here? In essence, ChatGPT learned these phrases from its training data — professional documents, business articles, formal writing from across the internet. The model absorbed the pattern that filler signals “serious tone,” so it reaches for it constantly. But it’s much more than a quirk. It’s pattern recognition that calcified into habit. Not laziness. A feature working exactly as intended, just not for you.
The good news is you can break this without switching tools. Once you understand where these phrases come from, designing prompts that stop them is surprisingly straightforward. So, without further ado, let’s dive in.
The 15 Phrases That Make AI Writing Obvious
Not all filler is created equal. Some phrases are the telltale marks of AI output — the ones that make editors wince instantly:
- Delve into — Nobody actually “delves.” It’s vague, it adds nothing, and it’s everywhere. Use “explore,” “examine,” or just cut it entirely.
- It is worth noting — The AI’s method of flagging importance without earning it. If something genuinely matters, that’ll be obvious from context.
- In today’s fast-paced world — Every world moves at some pace. This phrase has never added meaning to a single sentence. Cut it.
- At the end of the day — A transition that transitions to nothing. State the point directly and move on.
- Leverage — Business jargon that bleeds into AI output relentlessly. “Use” works in almost every case. Seriously, just use “use.”
- Navigate — A favorite for describing anything non-literal. “Handle,” “deal with,” or “manage” are all clearer.
- Shed light on — Metaphorical filler. “Explain” or “show” gets you there faster.
- In conclusion — Your final paragraphs need real thoughts, not a transition marker that signals the AI ran out of ideas.
- Going forward — Future-tense stuffing. Just say what’s next.
- Best practices — So vague it’s useless without specifics attached. Name the actual practice instead.
- In this article — Your reader knows they’re reading your article. Meta-commentary breaks the spell immediately.
- Cutting-edge — Corporate speak for “new.” Sounds forced in almost every context outside a press release.
- Deep dive — The AI equivalent of throat-clearing. Either examine something thoroughly or don’t mention it at all.
- Stakeholders — Often dangerously vague. Name who you actually mean — clients, managers, end users, whoever.
- Provide value — “Value” is invisible until you show the reader something concrete. Describe what they actually get.
Fix It Before You Write — Prompt Instructions That Actually Work
Frustrated by filler appearing in every single draft, I started building prompts with hard constraints — not vague guidance, but specific bans that change the output before it reaches you. That’s what makes prompt engineering endearing to us writers who rely on this stuff daily.
Here’s what works.
The Banned Phrases Instruction
Before prompt: “Write a 400-word article about managing remote teams.”
After prompt: “Write a 400-word article about managing remote teams. Do not use these words or phrases: delve, leverage, navigate, at the end of the day, in conclusion, it is worth noting, shed light on, best practices, deep dive, going forward, cutting-edge, stakeholders. Avoid vague abstractions — be specific about people and actions. Use direct language.”
The difference is stark. The first version reads like this:
“Managing remote teams requires the ability to navigate complex communication challenges. At the end of the day, it is worth noting that best practices around engagement can shed light on improving productivity. Leveraging the right tools will help stakeholders collaborate more effectively.”
The second reads like this:
“Managing remote teams means being intentional about how you communicate. Your team needs clear expectations, regular check-ins, and tools that don’t create friction. Asana, Slack, and Google Docs work well because people already know them. Set recurring one-on-ones and make space for non-work conversation.”
Shorter. Clearer. Actually useful. Banning specific phrases forces ChatGPT to find real words with real meaning attached to them.
The Tone Specification
Layer voice requirements in alongside your banned phrases list. Add something like: “Write in a conversational, direct tone. Sound like you’re explaining this to a colleague over coffee — not drafting a memo. Use short sentences. Avoid corporate language.”
Combining constraints — banned phrases plus tone guidance — locks output into something that reads like a person wrote it. Both instructions together. Neither alone does enough.
Fix It After — A Fast Edit Pass for AI Drafts
Sometimes filler gets through anyway. Here’s a five-minute routine that cleans it up reliably.
Step One: Find and Replace
Open your draft. Use Ctrl+F — or Cmd+F on Mac — and search for your personal top-five offenders. Mine are “leverage,” “navigate,” “deep dive,” “at the end of the day,” and “in conclusion.” Every time one appears, ask: does this sentence function without it? Usually yes. I’m apparently wired to hate the word “leverage” specifically, and cutting it every single time works for me while leaving it in never does.
Don’t make my mistake of trying to fix these by rewriting around them. Just delete and move on.
Step Two: Read It Out Loud
Probably should have opened with this section, honestly. Reading aloud catches awkwardness that your eyes skip entirely. You’ll hear filler phrases the moment you hit them — they’re rhythmically clunky in a way that’s immediately obvious when spoken. If you stumble on something while reading, that’s your signal.
Step Three: The Rewrite Prompt
If a section still feels off after reading through it, paste just that paragraph back into ChatGPT with this instruction: “Rewrite this paragraph using simpler, more direct language. Remove any corporate jargon or filler words. Make it sound conversational.”
This fixes what remains without requiring a full regeneration. Surgical. Takes maybe 90 seconds per paragraph.
Train ChatGPT on Your Patterns Over Time
ChatGPT’s custom instructions feature — buried in Settings under the account menu — lets you embed your preferences permanently. Write instructions that apply to every single conversation going forward:
“When I ask you to write, avoid these phrases: [your list]. Prefer simple words over corporate language. Write as if speaking to a professional peer, not a formal audience. Use specific examples instead of generalizations.”
I built a roughly 300-word instruction set covering industry-specific jargon I never use, tone preferences, and my full banned phrases list. One setup session — maybe 45 minutes total. After that, every response respects those boundaries without me repeating myself.
The real shift happens when you stop treating ChatGPT as a black box. It might be the best available option for fast content drafts, as this kind of work requires consistent output at scale — that is because the tool responds to constraints in ways most writers don’t initially expect. Feed it specific rules. Watch the filler disappear — not because the AI got smarter, but because you told it exactly what you need.
Leave a Reply