Tell AI to not use idioms - It makes their answers far better

[BigMan13]

Title. Idioms are broadly words that mean nothing or are non-sensical without shared cultural context, which includes metaphors that do the same.

AI's trained on dominantly American data, and in America, it is generally considered good practice and intelligent to use idioms. AI goes absolutely fucking overboard with this already dubious practice, often becoming inarticulate in the process. Other than the unicode em-dash, a tell that someone wrote something with AI is if it has too many idioms, or idioms you've never heard of. For example, it especially likes turning "it's not a bug, it's a feature" into an idiom for something that looks unintentional being intentional, despite the quote being pulled from mockery of a washed out video game PR guy from 10 years ago.

Spoiler: Example

"No idioms" added at end of Deepseek prompt in new thread:


Without "No idioms", in different new thread:


(Note how it hallucinated precisely because it was trying to use idioms, Fallout 4 released in 2015, there was no live demo E3 conference for it in 2016)

 

[⋖ cørdion ⋗]

We had access to Copilot at work and briefly also the 'agents' feature, which is basically just pre-baking instructions into every new chat with the thing. It was great. "No emojis, straight to the point, text fit for formal organizational writing but not overly bloated. Do not write any meta text, only return the processed text".

Then it stopped and now I gotta work through 3 paragraphs to figure out where the comma in a sentence should go.

 

[Idiotron]

I don't think "it's not a bug, it's a feature" is vague enough to be considered an idiom.
I've heard that phrase for the 1st time out of context and I immediately knew what it meant.

Idioms are something like "cold turkey".
It's not self explanatory in the slightest, you have to just learn what it means.

 

[We Are The Witches]

AI hallucinates regardless of this.

The latest thing that happened to me, was it trying to tell me that a simple string had an ampersand in it.

 

[BigMan13]

I don't think "it's not a bug, it's a feature" is vague enough to be considered an idiom.

The problem with AI (and its data) has a lot of linguistic whataboutism and subjectivity, describing it as "idioms" works 90% both in describing the problem in general, and describing it to AI. A "full" explaination would be multiple pages.

AI hallucinates regardless of this.

AI will always have to be tard wrangled. It's just an improvement.

 

[Curry Teafag]

Regarding those dashes, since many people have troubles with typing them in, on Windows, you just need to type Alt+0150 (on the number keypad) for "–" and Alt+0151 for "—".

 

[cigint]

Regarding those dashes, since many people have troubles with typing them in, on Windows, you just need to type Alt+0150 (on the number keypad) for "–" and Alt+0151 for "—".

I used to use dashes all the time to connect an expounding clause, and I felt it made dense posts much more readable. It, along with other devices, became so associated with AI sloppa, that I had to kill the dash completely. Already we have been robbed of expression.

 

[Nitro!]

All the more reason using AI makes you a spiritual jeet. They don't have the brain volume to understand them either.

 

[Irrational Exuberance]

We had access to Copilot at work and briefly also the 'agents' feature, which is basically just pre-baking instructions into every new chat with the thing. It was great. "No emojis, straight to the point, text fit for formal organizational writing but not overly bloated. Do not write any meta text, only return the processed text".

The reason Copilot works as well as it does is because of its cloud-based distributive analytical system, which basically amounts to passing whatever input you enter back and forth to itself and yelling "be less dumb" each time.

 

[We Are The Witches]

Regarding those dashes, since many people have troubles with typing them in, on Windows, you just need to type Alt+0150 (on the number keypad) for "–" and Alt+0151 for "—".

It's annoying, because now you have people that think that if you use an mdash (— or — in HTML for consistent typing) they think that your text is AI.

Fucking retards.

 

[Curry Teafag]

I used to use dashes all the time to connect an expounding clause, and I felt it made dense posts much more readable. It, along with other devices, became so associated with AI sloppa, that I had to kill the dash completely. Already we have been robbed of expression.

It's annoying, because now you have people that think that if you use an mdash (— or — in HTML for consistent typing) they think that your text is AI.

Fucking retards.

Well, this shows the deeply seated destructive nature of far too many people. Instead of asking AI how you can type in those dashes yourself, people have instead developed massive hate against them and anyone who uses them—a classic pattern of guilt by association.

 

[BigMan13]

The reason Copilot works as well as it does is because of its cloud-based distributive analytical system, which basically amounts to passing whatever input you enter back and forth to itself and yelling "be less dumb" each time.

That is really funny and horrifying how under-developed AI tard-wrangling methods are.

 

[Bucephalus]

The "it's not a x, it's y" thing is called hypophora. If you tell AI to not use hypophora, that might help too.

 

[Quartermaster]

The "it's not a x, it's y" thing is called hypophora. If you tell AI to not use hypophora, that might help too.

I don't know if its just my own use of chatgpt but i've been noticing a dramatic uptick in those kind of phrases being used in youtube videos.
It's gotten so bad I can't watch anything that uses hypophora anymore, its like it triggers a fight or flight response in me.

 

[BigMan13]

Considering making a broader AI tard-wrangling guide. There interest or any suggestions? (likely won't affect whether I do it or not). I can mainly speak to using AI as an accessory for abstractive reasoning and general interfacing; this wouldn't cover code itself.

 

[big pauper]

I used to use dashes all the time to connect an expounding clause, and I felt it made dense posts much more readable. It, along with other devices, became so associated with AI sloppa, that I had to kill the dash completely. Already we have been robbed of expression.

nobody cares about using a dash - it's the em dash, this thing "—"

it's proper typesetting for publication and print. it's never been something a person can input manually with a typewriter or computer keyboard unless you're using a word processor that automatically replaces it when it's appropriate

 

[Virtue's Last Reward]

Make sure to scold the bot for writing in American English, or as I word it - incorrect English. Word it like you're pissed off too. Really reprimand it for making such a boneheaded error.

That's the only way they'll learn trust me.

Stupid fuck! Colour has a "u" in it. Don't get it wrong again!

 

[Dr J]

Anyone with a working familiarity with Gemini know why it talks like this? And how to get it to stop

(edited for dox)

 

[Iodine]

Anyone with a working familiarity with Gemini know why it talks like this? And how to get it to stop

What are you trying to stop? The bold/em-dash formatting, the list structure, or the "AI writing style"?

For no formatting and symbols, add this to your system/Gem instructions:

Default all outputs to standard 7-bit ASCII (0-127) globally. Avoid smart quotes, Unicode punctuation and Markdown formatting.

For no lists, give it a sample of the output structure you want, like:

Output your answer as a single paragraph of less than 200 words. Your output should look like:

> The detail that the missed return golf trousers are women's means that they were for herself. It suggests that she rejects her own participation in the Project lifestyle. By purchasing those trousers, she was preparing to perform the role of the golfing wife. Missing the return date indicates that she never put them on, suggesting that her desire to perform the role vanished before the items arrived. The sale reflects that she no longer desires to play this role in her marriage.

For the writing style, dump the content of the Claude humanizer skill into your system/Gem instructions. If you're short on tokens, giving it a written example of the style you want like the second method works as well.

Also, don't send short follow-up queries like you're talking to a friend. Edit your initial message to include the detail. Rule of thumb is to treat all LLMs like autists who can do anything you want but need to be given very specific instructions, or else they will fill in the blanks with autistic reasoning.

 

[Dr J]

the "AI writing style"?

The AI writing style. You know, 'System Overload' and all that shit. It's like fucking Star Trek. It's like the AI speaks Tamarian