Are LLMs Making You Sound Less Certain Than You Are?
A closer look at how LLMs can influence writing style, confidence signals, and communication patterns over time.
In this post:
• How AI tools can unintentionally make confident writers sound less confident
• The difference between communicating observations and communicating reactions
• Real examples from my recent LLM Visibility Test project
• A practical editing technique for reducing unnecessary hedging
• What AI users should watch for as these tools become part of daily communication
What Hedging Looks Like in AI-Assisted Writing
A few weeks ago, I started noticing something strange while writing my recent LLM Visibility Test series. The project itself was straightforward. I was testing ChatGPT, Claude, Gemini, Copilot, and Perplexity to see what they could infer about my work from years of public content. I wanted to understand how these systems interpreted expertise, recurring themes, and professional identity.
The results were interesting enough on their own but somewhere in the middle of writing those posts, another pattern started showing up. Not in the LLM outputs or in my own drafts. As I worked through edits with ChatGPT, I started seeing more phrases like:
“Honestly, that surprised me.”
“That made me pause.”
“That fascinated me.”
“That genuinely surprised me.”
Nothing dramatic or wrong but after seeing variations of these phrases over and over again, I realized that the language kept shifting attention away from the observation itself and toward my reaction to the observation when I didn’t want it to. It was drafting and editing texts to include far more hedging than I’d ever use.
When I say hedging, I don’t just mean words like “maybe” and “perhaps.” I also mean language that subtly softens an observation by wrapping it in surprise, uncertainty, or self-reflection.
For example, one of the findings from the visibility experiment was that multiple LLMs identified similar themes in my work despite being asked slightly different questions. They consistently connected me to podcasting, communication, language learning, expat life, and helping people make complex information easier to understand.
An example:
The important part was of the below sentence is that the systems arrived at similar conclusions despite the prompts being different. But look at how ChatGPT softened this idea that it moved the focus from the LLMs agreement to my feelings.
BEFORE:
“Honestly, that surprised me a bit because I didn’t heavily force those terms in the prompts themselves.”
AFTER:
“I didn’t heavily force those terms in the prompts themselves, which made the consistency across the LLMs especially interesting.”
The two sentences have the same information but a very different emphasis. One focuses on my reaction and the other focuses on the result. As soon as I started spotting this pattern, I couldn’t stop seeing it.
Gender Bias in AI Writing?
Before going any further, I should be clear about something. I’m not sure of this is a is a sexist LLM problem, although I have my suspicions. In fact, every time I wrote this post with that slant, ChatGPT rewrote it to be a more vague all people problem. This could be true and it might not be true. The blanket papering over of my suspicion was even more annoying than the initial hedging though.
Even so, my experiment wasn’t designed to test the is this sexist question, and I don’t have enough evidence to make that claim. In fact, I suspect at least some of this behavior affects everyone to some degree. So I am writing with that lens for this post. However, I’ve got a poll on LinkedIn to see what other people’s experiences are with this. If I see a gender bias, I’ll report back to you.
What I do know is that I noticed the pattern in my own writing. The shift wasn’t dramatic. It wasn’t turning strong conclusions into weak ones. It was more subtle than that. Observations gradually became wrapped in reactions. Findings became framed through surprise. Statements became slightly softer than I would normally write them.
LLMs Often Treat Exploration and Uncertainty as the Same Thing
One thing that bothers me about some AI-generated writing is that it often treats curiosity and uncertainty as if they’re the same thing. They’re not.
I run experiments all the time.
I retitled 92 podcast episodes and tracked download changes.
I tested transcript strategies for podcast discoverability.
I compared five different LLMs using the same set of prompts.
I spent 3 years documenting my curiosity with hanzi characters (Mandarin Chinese writing system)
I’m testing out different podcast SEO and GEO methods all the time
In the LLM Visibility Test I wasn’t doubting myself, I was exploring and gathering information.
Somewhere along the way, LLMs seem to have learned that people exploring ideas should sound perpetually surprised by their own findings. I don’t think that’s how most experienced professionals actually communicate. Researchers don’t write, “Wow, I can’t believe that happened” in their research. Neither do consultants, business leaders, etc. “Honestly, this surprised me” is more casual spoken language. So dare I say it’s surprising that is kept showing up in these drafts.
People who work with data, systems, audiences, customers, and communication patterns usually say:
“Here’s what happened.”
“Here’s what stood out.”
“Here’s what changed.”
“Here’s what I think this means.”
That’s a very different voice.
How AI Writing Tools Shift Attention From Observations to Reactions
Once I noticed this, I started making a simple editing pass through my drafts. I looked specifically for reactions. Not because reactions are bad. Sometimes they’re important and they can help readers understand why something matters. But many of them weren’t adding value.
Example 1:
“That honestly made me pause for a second because it revealed something interesting about AI visibility.”
became
“That revealed something interesting about AI visibility.”
Example 2:
“What surprised me was how much weight those older identity signals still carried across the LLMs.”
became
“What stood out was how much weight those older identity signals still carried across the LLMs.”
The second versions feel more grounded to me. They’re still human and conversational but they don’t accidentally undermine the observation they’re trying to highlight.
How AI-Assisted Writing Can Change How Expertise Is Communicated
At first glance, this may sound like a small writing issue but it isn’t. The more I looked at these edits, the more I realized they were changing how expertise was being communicated. Most people don’t have direct access to our expertise. They experience it through our communication. Clients, readers, and potential employers don’t know what we know. Increasingly, AI systems don’t know what we know either.
They only have access to how we express it.
If AI-assisted writing consistently encourages us to wrap observations in surprise, hesitation, or self-reflection, that can subtly change how our expertise sounds, even when our actual knowledge hasn’t changed at all.
It’s not always bad but depending on what you’re writing, who you’re writing for, etc, it could be something to consider removing from your AI co-written text. As always, you are the final editor of your text. So once you notice it, you can decide whether those changes reflect your voice or the system’s.
A Simple AI Editing Technique for Reducing Unnecessary Hedging
Lately, I’ve been asking myself a very simple question when reviewing LLM-assisted writing:
Is this sentence communicating an observation or a reaction?
Sometimes the reaction stays because it fits. But surprisingly often, removing the reaction makes the sentence stronger.
For example:
“I was surprised that older content still influenced the results.”
tells readers how I felt about what happened.
becomes
“Older content still influenced the results across all five systems.”
tells readers what happened
Those are not always the same thing. And they don’t always deserve equal space on the page.
How I Keep My Own Writing Voice When Working With LLMs
LLMs don’t just generate content, they influence how we communicate. Sometimes that’s obvious but it’s subtle enough that sometimes you don’t notice until you’ve worked through a few dozen drafts.
For me, one of those patterns was the tendency to frame observations through surprise. The fix wasn’t complicated.
I added this to my ChatGPT custom instructions
“remove hedging. Instead of ‘I think’ or ‘ It could be’ or other uncertain moments during certain data sharing, write plainly and confidently like ‘ The data...’ or ‘ It was clear that...’
Knowing that custom instructions are not always followed, after a draft is ready, I add a check for these types of sentences during the HI (human intelligence) edit and change them as I go. No need to prompt my way through it since it’s apparently a default writing style now in ChatGPT and some other LLMs.
That small shift made the writing sound much more like me. I have no problem talking about my feelings but not when it comes to geeking out about an experiment. Not in the reporting of the steps and results. That’s for facts and information, not emotions. And once again ChatGPT reminded me how I like to communicate by doing something that I don’t like in our co-writing. This will happen and it’s the awareness that can make sure that we keep our own voice in our co-writing sessions.
📬 Know someone who uses LLMs regularly for writing, strategy, research, or communication? Share this post with them. They may be picking up communication habits they haven’t noticed yet.
💬 Have you noticed LLMs influencing your writing style, vocabulary, confidence level, or communication patterns? Leave a comment. I’d love to compare notes.
See you next week,
Stephanie