How can I mimic ChatGPT?

Authoring General Design Discussions
1

In one of my WIPs, I want to poke fun at people who rely on LLMs to run their lives, by having a particular character talk exclusively in that bland, repetitive “AI voice”.

However, I also don’t want to actually have ChatGPT generate the lines, because even if the purpose is to mock it, I don’t think the IFComp audience would like seeing that some of the game text was AI-generated.

So instead—what are some “tells” of AI-generated text that I can imitate (by hand) while writing cheerfully banal dialogue? Em-dashes, “delve”, and “not just X but also Y” are famous ones, but I imagine there are people here with far more experience with LLM-generated prose than me.

8 Likes
2

The main ones are overuse of em-dashes, rule of three, bullet points, rewrite/summary/next steps offers, emoji, bold text, etc.

This has a ton more plus examples:

5 Likes
3

To add onto that, here’s a writeup from someone who looked at a bunch of AI-generated fiction and tried to analyze the style they write in. AI tends to write fiction in the same style no matter what, with rare exceptions.

There’s an accompanying Github repository with examples, if you can stand AI-generated short stories.

There are certain styles of metaphor they like to use, which are hard to describe concisely but can be seen in the examples linked above.

4 Likes
4

That’s not ChatGPT! That’s sparkling slop.

3 Likes
5

The tell I usually pick up is how a question response is presented in book-report form. There’s an introduction, an empathy statement, the actual answer, and then a concluding paragraph that wraps up and will sometimes attempt to globalize your question.

4 Likes
6

Three big ones I don’t see mentioned:

  1. Sycophancy. Compliment the player constantly. Every interaction is fascinating, insightful, “not just interesting—that’s profound.” When corrected, immediately apologize and adopt the player’s perspective with a summary like, “You’re right—my previous answer did not consider… [summarize the issue]”
  2. Summarizing the prompt. Begin every interaction by recapping what you’re being asked to do (with appropriate compliments).
  3. Confidence. Explain in detail how every answer is correct :white_check_mark:, comprehensive :white_check_mark:, and passes all unit tests (confirmed) :white_check_mark:, even while completely misunderstanding the point or hallucinating nonsense.

One could also play with the idea of the character hallucinating factual information.

12 Likes
7

I think the overall unifying principle is that your goal should be to write an entire paragraph that has the same informational content as a single-sentence prompt. Specific tells like “it’s not just x — it’s y” and restating the prompt can arguably be understood as special cases of “ways to use a lot of words that add nothing of value.”

10 Likes
8

I suspect they will be different six months from now, and six months after that.

7 Likes
9

One of my earliest experiences with ChatGPT was to type the prompt “write something that sounds like it was written by ChatGPT.” The system locked up and didn’t respond, as if it was in a self referential loop. I swear this is true.

2 Likes
10

It’s true that it’s not ChatGPT specifically, but pretty much all AIs write the same. They’re all trained on each others’ data, so they homogenize into the same voice and the same cliches. People will talk about their personalities, but in my view it doesn’t matter because you can easily prompt an AI to adopt a different personality and it will. If an AI seems to have a default personality, that’s just what the default prompt tells it to be. But the fiction writing style doesn’t change no matter which AI you’re using.

Also, there are dark, dark places online where you can see ChatGPT fans post screenshots of their chats, or read ChatGPT-generated fanfiction. Not sure how useful this would be as a reference.

I rather feel that as more and more of the internet consists of AI-generated text, and AIs continuously scrape that text to use for training, they’ll solidify their existing writing patterns. I’m no expert, though, so we’ll see. Could be wishful thinking on my part.

2 Likes
11

The surface-level tells will change for sure (I hardly use ChatGPT, but I gather it has been taught to correct its overuse of em-dashes and the word “delve”). But the structural hallmarks of LLM-generated writing (putting everything into the most general context, restating the prompt, spinning out single facts into multiple sentences) seem like they’ll be harder to bury.

One that I think is very obvious with certain applications of LLMs and probably very easy to parody is the voicing of internal narrative. Since all an LLM can do is generate text, all of its “thought processes” have to be represented in the text.
So you get a lot of “but wait, let’s consider …” and the like while the LLM develops its thought processes.

5 Likes
12

Maybe. I’m no expert, but my understanding is that even if all LLMs are pre-trained on the same corpus, a lot of the “special sauce” that goes into training the model is in the supervised fine-tuning and reinforcement learning from human feedback, and that different methodologies here give the different models their “personalities.”

1 Like
13

I still don’t know how to feel about it picking Delve so frequently, when I published a mobile game called “The Delve”, which I had locked in the name for long ago, but published when LLMs were already widespread. I was honestly just referring to “dungeon delving”.

Anyway, I have played around a bunch with LLMs, and its usual speech pattern is absolutely grating to read. It does differ slightly between models though, but there are obvious clichés. In this case I suspect you’d want ChatGPTs pattern though.

There’s two modes I’d like to consider: LLM as a conversation partner, and LLM as a writer. The former is default mode, and the latter happens when you ask it to write a story.

In conversation mode, it will in fact praise you into high heavens. So, when I ask a question like “I see you did X like Y. Wouldn’t it work better if Z?”

The response will often be something like:

You are absolutely right :white_check_mark:, doing it like Z wouldn’t just be easier, it would be a professional grade solution to the problem!

It will:

  • Make the X cleaner :white_check_mark:
  • Do away with the needless Y :white_check_mark:
  • Make your producttruly stand out

Your plan is a master stroke

Of course, it will then completely drop the ball and based on prior context get it completely wrong.

Now, asking it to write will do away with the verbose sycophancy, but it will still use the rule of three (or five), and most gratingly, it will use every comparison as an opportunity to use the “It’s not X, it’s Y” structure.

Also, unless instructed otherwise, it can feel a bit like an ad copy. Much of the writing I am personally most familiar with uses nice, medium length paragraphs which contains a story beat or piece of information per paragraph.

It’s not as bad as LinkedIn influencers who will write a single thought per sentence, like so:

Going to work is the most important thing you can do.

How did you respond?

I’m sure most of you didn’t agree.

I understand.

Hearing that can be shocking.

But let me explain.

…following this is a newline seperated, poorly argued shrine to toxic productivity

But, ChatGPT will still find a way to give every paragraph a header, and cut up the text not as actual informational text, but the type of ad copy you would find on those suspicious sales websites.

So, I suppose… Write like an ad-copy praising the other person’s ego?

3 Likes
14

I find Ken Cheng’s infamous parody Linkedin posts quite funny, since he gets the tone of Linkedin influencers down pat.

From I’m buying my nephews and nieces dogshit presents for Christmas this year:

I’m buying my nephews and nieces dogshit presents for Christmas this year.

Why?

Because they didn’t sign up to my new subscription service.

It’s called the “Good Uncle” app.

Family members pay monthly to ensure I’m consistently a fun and generous uncle.

Last year I bought them amazing presents: games consoles, rare pokemon cards, stock options.

I told them if they want this to continue, they have to “go Premium” on Good Uncle.

It only costs £9.99pm and guarantees presents all year round. The parents refused.

Well, I now have no choice but to enact the Enshittification of Christmas.

Last year I gave Sean a Switch 2. This year I’m giving him Steven Bartlett’s book.

Then I’m going to get drunk and ruin Christmas.

I can’t wait to watch Good Uncle subscribers skyrocket.

He has a lot more of these.

5 Likes
15

That is genuinely hilarious. That could be the sole reason I can’t bear to even open that accursed website, weren’t it for the constant (self-)congratulating going on being almost equally as grating.

16

Years ago, I had written a story about an alien world that was tidally locked to its sun. That means, like our own moon in relation to Earth, it always faced the same side toward the sun. This baked one side in eternal sunlight, and froze the opposite side in eternal darkness. I even had a giant sand dune that got so hot at the equator closest to the sun that it stayed constantly molten in a lake of unnaturally pure liquid glass, as it had geological time to differentiate, with heavier solid particulates settling to the bottom. The only truly habitable area was the fringe of land in perpetual Twilight, the sun just peeking over the horizon, permanent night only a few miles away. This caused the human colonists who landed here to spread out on this ring, but heavily impacted the politics, as, on a ring, you only ever have two neighbors, those to either side of you. This made it more culturally important to honor diplomatic immunity for passage of goods and information. If you violated it, it was extremely easy to become very isolated as a result.

Anyway, sorry, I’m getting distracted. There’s much more than worldbuilding to that, but the world of the story was relevant to the point.

The point is, I had a full draft of this (naively) going out to publishers in late 2009, early 2010. The title, given the nature of the world, was naturally Twilight.

Unfortunately, at the same time this was happening:

4143d720-6ff9-4c59-940a-592055c36e67.sized-1000x1000 (1)
4143d720-6ff9-4c59-940a-592055c36e67.sized-1000x1000 (1)2000×2963 460 KB

My story lacked vampires and werewolves, sparkly or otherwise.

5 Likes
17

As proof that all great minds (see what I did there) think alike, might I direct you at my Sun Keld RPG, set on a one-face world where there are molten lakes at the pole, and all the cities are located on the equator in the realm known as “Twilight” …

Of course, mine has an equatorial river, a secondary sun, and furry lizard people ruling a magical Sumerian empire, so …

1 Like
18

Didn’t Necroscope have something like that involving an alternate Earth where the eternal night side was controlled by vampires? I’m afraid my knowledge of that series is second hand.

1 Like
19

I was writing lore snippets for a Sci-Fi LARP I designed, to really encourage players to be creative with their homeworld, and I actually very much casually dropped a tidally locked world on the lawless outer rim ruled by tech vampires who use advanced gene editing and revitalisation technologies to drain people of their biological essentials to rejuvenate themselves.

Interestingly, that type of setup very much works to reinforce the idea of a vampire as a metaphor for social inequality. I believe the original concept of the vampire also had something about that? The LARP is all about the inequality inherent to the gilded age, as its set on a planet that is one big corporate town, so I felt it was appropriate.

3 Likes
20

I did this for my Ecto entry. My inspiration was this bluesky post:

3 Likes