I'm sure we all have that feeling sometimes when we're reading something, that it seems off. Like robotic sounding, too well-structured. Like you don’t know whether or not to trust it. Psychologists call this 'Good Instincts'. AI content has made its way everywhere, and it's getting much harder to identify what's human-written and what's AI generated. But does it even matter? If something was written by AI and you didn't know until after, would your opinion on the content change? Or would you feel betrayed, or maybe embarrassed that you couldn't tell?

Just like you and I, AI models have a certain vernacular and writing pattern that it follows when producing answers. It's not obvious, but it's there. What this means is that these patterns could find their way into how we write. We pick up language and syntax fairly easy via repetition. And when AI is so easy to use and replace our own writing with, we see the patterns often. As much as I hate to admit it, I've noticed some of these patterns in my own writing (but luckily, I don't know how to type an em-dash).

---

How we absorb vernacular

Absorbing vernacular works like learning language, though a lot of the time it is involuntary in comparison. Living in a new city (especially in the UK), you could pick up their accent and slang through immersion. This is when you spend so much time in a specific community, that common words and phrases in that area just spring to mind because you've heard it so often. We either absorb vernacular naturally or intentionally. Other than immersion, we have:

• Active listening - Observing phrasing, cadence and vocabulary while paying attention to the tone and context to grasp subtext.

• Natural Acquisition - Reading works and material from that area, allowing words to enter your mind naturally.

• Contextual Repetition - Keeping a list of vocabulary to familiarise yourself with new words through repeated exposure.

In our case of AI vernacular, I should hope that we aren't learning it intentionally. So we will stick to natural absorption of vernacular in this article. Regardless of the method, we learn through repetition just like with anything.

---

AI Content and Writing Patterns

To absorb language naturally, it must be repetitive enough to create a pattern. So what kind of patterns do AI models follow? Of course, these may depend on the specific model or brand (Claude, ChatGPT etc.). But they usually stick to similar structural patterns mainly due to the underlying architecture. To start, we'll look into structural habits.

• Heavy use of bullet points and headers, especially when prose would read more naturally.
• Responses that are verbose.
• A weird rhythm where the question is restated, answered, and then summarised.

The list goes on, and is exhaustive, but common structure patterns are usually some form of being highly structured, and thorough. This could be due to human feedback during model training, people often rating more structured, verbose answers highly. Along with this, there are some overused words that are almost universal across models. Onto the vocabulary.

• Niche words that humans rarely use, like delve, nuanced, additionally, crucial (though I like using this), pivotal.
• Uncommon phrases like "It's worth noting that...", "In the realm of...", "At its core..."
• Hedging phrases used often, "it's important to remember" or "this is a complex topic"

Authoritative and thoughtful words tend to be rated higher in human feedback. More on this at the end of this section. What about tone now?

• Relentlessly agreeable and validating.
• Diplomatic - reluctant to take a firm stance.

By design, AI models could be agreeable and diplomatic to keep us engaged with them. Lastly, let's briefly touch on rhetorics.

• Listing in threes.
• "It's not this — it's that."

To make their point, AI models tend to use the rule of three, especially in lists. A more interesting rhetoric pattern is the second point listed. That contrastive stucture of "It's not X - it's Y" is used to reframe a concept in order to create a sense of depth. AI generally cannot tell what is significant, so giving weight to certain points frequently by using that phrase makes it seem out of place and overgeneralised to us.

Now we know what some patterns are, of course we haven't covered them all but it's enough to get some ideas flowing. We need to explore why AI talks this way.

It's difficult to pinpoint exactly what the causes are, but there are real convincing ones when we get into the guts of how AI has been trained. A key part of the training process is RLHF (Reinforcement Learning Human Feedback). This is to align AI models to human preferences, by taking feedback and guiding AI to produce answers that are more engaging. The result is an alignment to helpful, honest and safe answers, it just so happens that these answers are agreeable, and diplomatic. A peaceful, non-confrontational alignment to producing answers was deemed the most helpful during human feedback stages.

That's mostly good, we wouldn't want to speak to someone that disagrees with us all the time. But overly-agreeable models prove to be slightly less useful. They invalidate bad ideas and soften necessary criticism. It creates a kind of comfort zone that isn't beneficial to be in. So a lot of modern models are being made to be a smidge more disagreeable, a healthy amount. More human? It's hard to say. Seems like that's the direction. But I think that it's better to be able to distinguish between AI and human content, don’t you?

---

Flaws with AI detectors

On that note, I'm sure that you've seen, or maybe even used, AI detectors. Always marketed as 'not guaranteed to be correct'. They could even be correct 99% of the time. But that 1%. Would you place trust in a system that could be wrong? Obviously, you wouldn't stake everything on an AI detector's assumption. But when false positives do happen, the implications could be huge. Schools, for example, are not adapted for AI. Students are going to use AI as much as they can. Call it laziness or something, but I would have used it when I was in school so I don't blame them. What can the teachers do about it? They'll most likely use an AI detector. Imagine being a student that has not used AI, but gets marked down because an AI detector made a false positive assumption.

It's clear that we can't rely on only one source for a definitive answer. But the fact that AI detectors do exist and work proves that there are distinguishable patterns. Patterns that we can use to decide for ourselves what we do with the information. Do we disregard it because it is AI generated? Or do we evaluate it on its own merits? Even if you still aren't confident in your own abilities of telling apart AI content, that's okay. It's getting much harder to tell. All you can do is focus on your own writing. Do you want to sound less like AI? Sure, but the better goal is to sound more authentic, more like yourself.

---

Avoiding sounding like AI - social stigma

There's been a kind of mini revolt against using AI for generating text or media. I'm sure you've heard the term, AI slop. A lot of people are bored of reading something, just to find out that it wasn't even a human that made it. There's a stigma to using AI for making content. Because it is seen as lazy, inauthentic and impersonal. So do we avoid trying to sound like AI? I reckon that the more attention we bring to trying to avoid sounding like AI, the more we may end up actually sounding like AI. Don't think of a white bear. Not exactly the same, but you get the idea.

Even though there is a social stigma to using AI content, it's prevalent and not something we can avoid. There will be vernacular that we pick up from AI, I'm almost certain. But instead of trying to sound less like AI, just focus on sounding more like yourself. That's all we can do, in a world filled with bots, AI content, and people trying to be someone they aren't. Authenticity should be what we strive to achieve.

---

TL;DR

AI models follow recognisable writing patterns - overly structured, agreeable, and full of certain tell-tale phrases. Largely shaped by how they were trained on human feedback. These patterns are everywhere now, and like any language we're exposed to repeatedly, they could start creeping into how we write. AI detectors aren't reliable enough to be the answer, and obsessing over not sounding like AI might make it worse. The better goal is simply to sound more like yourself.