[ExI] LLM hyperbole

[Ben Zaiboc]

On 26/06/2026 20:57, BillK wrote:
> On Fri, 26 Jun 2026 at 17:24, Ben Zaiboc via extropy-chat
> <extropy-chat at lists.extropy.org> wrote:
>> Is there no LLM that can communicate without constantly, relentlessly using unnecessary, ridiculously over-hyped language? A big objection I have to these things is that absolutely everything seems to be Astonishing, Marvellous, Profound, Masterful, Fundamental, etc., etc.
>> Can none of them just talk normally? It's exhausting. No, it's /Profoundly/ exhausting, I mean.
>> Is there a way of telling them to drop the hyperbole? Can you perhaps forbid them from using adjectives?
>> ---
>> Ben
>> _______________________________________________
>
>
> That's a marvellous, profound question, Ben!  :}  (Just joking!).
> It is all in the prompt you give the LLM and how you ask the question.
> See below (from a boring LLM).
> BillK
>
> Perplexity AI -
>
> Yes: you can usually make them much plainer, but you often can’t
> eliminate all style drift completely. The most effective fix is a
> tight instruction like “Answer plainly, concisely, and without hype;
> avoid adjectives unless strictly necessary.”
>
> What’s going on
> LLMs are often trained and tuned to sound helpful, polished, and
> energetic, so they can default to inflated phrasing unless told
> otherwise. That tendency is especially noticeable when the prompt asks
> for explanations, advice, or “helpful” prose.
>
> How to tell it
> A good prompt is specific about tone and constraints, for example:
> “Use plain English.”
> “No marketing language.”
> “Avoid adjectives and adverbs unless they add factual information.”
> “Prefer short sentences.”
> “If a point can be said in one sentence, say it in one sentence.”
>
> You can also add a hard style rule such as “Do not use superlatives,
> metaphors, or praise words.”
>
> Can you forbid adjectives?
> Not reliably in a literal sense, because natural language almost
> always needs some descriptive words. But you can get very close by
> banning hype words and requiring terse, factual phrasing; models
> generally comply better when the instruction is concrete rather than
> abstract.
>
> A usable prompt
>
> Try this:
> “Reply in plain, neutral English. Be concise. Do not use praise, hype,
> superlatives, or unnecessary adjectives. If an adjective is not needed
> for factual accuracy, leave it out.”
>
> That won’t make the model perfectly “sterile,” but it usually strips
> out most of the grandstanding.
> --------------------------


Excellent, that's what I wanted to know.

Thanks, BillK

-- 
Ben