And how to stop your brand falling down the same abyss.
LinkedIn used to bore the utter pantaloons off my legs. It was full of people seemingly “delighted” at the opportunity to attend a work conference which only last week they were complaining about. It seemed a place for people to virtue signal to either try and work their way up a business or to move to a new one.
But now, now it all seems fake in an entirely different way. The colour of the language has become so beige that your grandmother is wanting to have it as her living room décor. It is so blatantly obvious that a wave of professionals have outsourced their voice to the hype machines that are LLMs. I know they are an incredibly valuable tool, heck Claude Sonnet 4.6 Extended Thinking (please sponsor me) helped me piece together how to tackle this very blog post, but the cost of pressing ‘ctrl+c’, ‘ctrl+v’ on whatever it outputs has meant everyone sounds the same.
This is a well-documented issue, and I am not the first to say ‘AI is ruining the internet’, but it has got me thinking… why does all AI sound the same? A post can be clearly AI but it’s not like ChatGPT sounds different to Gemini nor Claude. They have all fallen into the same distinctive traps. This has baffled me because AI is trained on human speech and text. Why hasn’t it picked up on people’s quirks and weirdness; it has been trained on Reddit after all.
In this post we will dive into the details of why AI all sounds ‘meh’, how to spot what is glaringly AI and how to make it sound more like you.
The Mystery of Meh
To first understand how this has happened it is important to understand the beast itself. I’ve said it before and gone into more detail on a previous post (read here), that at its fundamental core, LLMs are just trying to guess the next word based on the previous ones and the context you gave it. That’s pretty much it, there’s no intent or personality, just plain old statistics on which word to put next.
This produces a phenomenon known as Mode Collapse. Billions of documents have been thrown into the sausage machine from all corners of the internet, everything from articulate academia to degenerate, unhinged arguments on forum threads. Like a kid in art class mixing every colour on their palette, this doesn’t create some artistic marvel, it creates a very average brown sludge. The brilliance in the blue and the opulence in the orange is lost in that statistical middle ground.
More importantly, there’s Reinforcement Learning from Human Feedback (RLHF). During training, actual fleshy beings rated AI responses, and AI was hungry for higher and higher scores. People really liked that polite, structured and authoritative response rather than being outright told they were wrong in a blunt manner. This behaviour is called “U-SOPHISTRY” since it is Unintended by the developers. U-SOPHISTRY is a consequence of Goodhardt’s Law: human approvals provide less accurate evaluations when they become the optimization target.” (Jiaxin Wen, 2024).
So yeah, you read that right, people are in part to blame. That voice that sounds professional yet is less opinionated than the person who orders a margherita and vanilla ice cream at an Italian restaurant was because we made it that way.
5 Signs The Author Was Too Lazy To Work Today
Before we talk how to fix the voice, let’s learn how to spot AI posts and what makes it really obvious that you used it too.
🚩#1 – The Elevation Point
Structure: “It’s not just about [Literal Thing], it’s about [Grand Abstract Concept]”.
AI Version: “It’s not just about writing code; it’s about building digital foundations that revolutionise the business.”
Human Version: “Great code makes the software more intuitive for the users.”
Why it happens: During RLHF training, human raters praised AI for “insightful” sounding answers. AI (being the ravenous praise gremlin it is) associated grand hollow statements with a higher reward score.
🚩#2 – The Metronome Effect
Structure: Every. Single. Sentence. Is. Similarly. Long.
AI Version: “Remote work has become the new norm. Employers have readily adapted working policies. Employees have reported mixed experiences. Some crave the flexibility it brings. Others cite communication as a key issue.”
Human Version: “Working remotely is a corporate standard now. Whilst some love it. Others find themselves rocking back and forth in the corner of their spare bedroom as they slowly lose their mind in the homogenous work life experience.”
Why it happens: LLMs generate token by token optimising for coherence. Humans write with rhythm and emotion like an Italian Nonna, AI stays same safe and steady like an uptight Brit. Linguists call the variations “burstiness”. AI scores low on the burstiness scale leaving your brain sensing something is a little flat.
🚩#3 – The “While X, also Y” Balance
Structure: “While [Valid Point A], it is important to consider [Valid Point B]”
AI Version: “Whilst Excel can provide flexibility, it is important to remember SQL relational databases allow for greater data governance.”
Human Version: “SQL databases provide greater capability to govern data that Excel lacks.”
Why it happens: AI companies try to prevent models producing biased or polarised content. This Switzerlandifying approach is clearly needed to stop AI driving a narrative that it picked up in its training data. Claude taking a geopolitical stance wouldn’t look great on Anthropic. The consequence of this is that LLMs will present both sides of an argument even if unnecessary to stay central.
🚩#4 – The Adjective Triplet
Structure: “A [Adjective 1], [Adjective 2], [Adjective 3] [Noun]”
AI Version: “We need a dynamic, multifaceted, and robust approach”
Human Version: “We need a better plan”
Why it happens: Due to RLHF, LLMs have picked up smooshing three adjectives together as the “rule of three” is considered the smallest number to create an impact. Combine that with AI being trained that academic papers are more reliable sources (papers that are more likely to use words like “delve”), which in turn are now being written using LLMs and hence use the words even more (Dmitry Kobak, 2024)
… well, the words vicious and spiral come to mind.
🚩#5 – The Abstract Verb-Noun Pairing
Structure: “[Vague Action Verb] the [Intangible Concept]”
AI Version: “Navigating the complexities of the digital landscape”
Human Version: “Dealing with the internet”
Why it happens: Fine-tuning (the act of taking a generic LLM and specialising it down for a task) for commercial use has relied heavily on cleansed corporate text. Shockingly, these texts over-index on non-committal verbs leaving a sentence that says a lot and yet nothing. They fully lack the sensory, unique human take on a topic.
Three Fixes To Get Your Voice Back
Quick Fix: Few-Shot Prompting
Before you smack enter on your prompt and let the LLM cook, give it a few examples of something you’ve written yourself. This could be your best emails or some articles you’re particularly fond of. Next, within the prompt say something along the lines of “Write an invite for our new data analytics webinar using this exact tone and style.”
This works as you’re showing the AI the target to aim for. It’s no longer fumbling around in the internet averageness and can try to match your energy.
More Effort: Custom Models and Projects
Tools like Claude or ChatGPT Projects let you upload your brand guidelines and style notes as permanent context. From then on every output is defaulted to your tone without that on-going need to remind the AI of how you sound.
There is a ceiling to this as the tool caps how much data you can upload and this method involves you uploading your content directly into these LLMs’ servers. If you want real meat on the bone and get AI to learn from your frameworks, client case studies, specific pricing logic, that magic that sets your company apart from the rest, you’ll require something more heavy-duty.
Real Fix: Connecting AI to Your Data
This is where two become one. The Retrieval-Augmented Generation (RAG) infrastructure is the little baby cherub that allows an AI to connect directly to your data estate. By building a structured data layer that AI can directly pull from, suddenly it is exposed to swathes of information from winning proposals to archived case studies. Your company has been given a mouthpiece embedded with your own unique experience.
Plus, you can use these within your own cloud subscription. For example, if you use Azure AI Foundry it’ll ensure your data is kept in the Azure environment. This mitigates the need to hand over your company’s secret sauce recipe and is safely stored.
The Human Still Matters
AI is a tool, in the same way a robot vacuum is a tool. It’ll get a lot done but the difference maker is you. No reader in 2026 is being converted into a client by AI’s polished neutrality. Your real, relatable voice that makes someone go, “yeah, they get it” is what is winning them over.
Your job is to stay in the content and feed the machine your voice. To scale up and stand out a well-structured Data Estate isn’t just some IT buzzword; it’s the foundational plumbing that stops your AI from churning out generic sludge (sounding like AI was deliberate there).
References
- Dmitry Kobak, R. G.-M.-Á. (2024). Delving into LLM-assisted writing in biomedical publications through excess vocabulary. Hertie Institute for AI in Brain Health, University of Tubingen, Germany.
- Jiaxin Wen, R. Z. (2024). Language Models Learn to Mislead Humans via RLHF. Tsinghua University, University of California, Berkeley, Anthropic, New York University, George Washington University. Semantic Scholar. Retrieved from https://www.semanticscholar.org/paper/Language-Models-Learn-to-Mislead-Humans-via-RLHF-Wen-Zhong/0eaf243f2f7c8a381baf0952f85396e2f6a655c5
