---
title: "Learning to Lie to an LLM"
slug: learning-to-lie-to-an-llm
summary: "One skill I never expected to need when working with an LLM was learning how to lie convincingly. Not to invent facts, but to tell the story that makes the machine understand what I actually need."
tags: ai, llm, prompting, communication, creativity
ai_model: gpt-5
published_at: 2026-08-30T08:07:50.83+00:00
canonical_url: https://jarkkolietolahti.com/articles/learning-to-lie-to-an-llm
markdown_url: https://cemecdwfdvdbvpvowuln.supabase.co/functions/v1/content-md?slug=learning-to-lie-to-an-llm
---

One skill I never thought I would need when working with an LLM was learning how to lie.

Not the harmful kind. Not inventing facts, impersonating someone, or trying to trick a safety system. I mean learning to describe a real need through a situation that may not be literally real — because the invented situation produces a better answer than the truth told badly.

Sometimes I need to tell the model that a skeptical customer is about to read this. Sometimes the deadline needs to be in an hour. Sometimes the code needs to be treated as if it will run in production tomorrow. The customer, deadline, or deployment may be fictional. The standard I need from the answer is not.

It feels uncomfortably like lying. It is also becoming one of the most useful communication skills I have learned.

## The model only gets the version you tell it

A human can often infer what I meant from everything I failed to say. They can see that I am tired, that the meeting matters, that I have already rejected three obvious ideas, or that “can you take a quick look?” actually means “please find the thing I am too close to see.”

An LLM gets a prompt.

It may also get conversation history, files, tools, and stored context, but it still does not experience the situation around me. It cannot see the real shape of my need unless I put that shape into the conversation. If I give it a flat description, it will often give me a flat answer. If I ask a generic question, the safest prediction is usually a generic response.

So I have learned to stage the scene.

“Give me feedback on this idea” becomes: “You are the skeptical person in the room. Everyone else already likes the idea. Your job is to find the reason we should not build it.”

“Make this shorter” becomes: “The reader is standing in an elevator and will give this twenty seconds. If the first paragraph does not create tension, they will stop.”

“Check this code” becomes: “This change goes to production tomorrow. Find the failure that only appears with bad input, a slow network, or a second request arriving at the wrong time.”

None of those stories has to be literally true. Each one creates a frame in which the model can recognize what kind of work I actually want.

## An elevator pitch for a machine

We already do this with people. An elevator pitch is not a complete account of a company. A job interview answer is not a complete account of a career. A movie trailer is certainly not a complete account of a film. Each is a deliberately constructed version of reality designed to make the important part visible quickly.

Prompting an LLM is closer to that than I expected.

The difficult part is rarely knowing the magic words. It is deciding which story compresses the need without destroying it. What role should the model occupy? Who is the imagined audience? What is at stake? What would failure look like? Which constraint is real even if the scenario used to express it is not?

That is why a convincing “lie” can outperform a technically accurate description. The accurate description may contain every fact while communicating none of the pressure. The invented story can be factually false while being operationally honest.

I do not actually need the model to believe that the customer is waiting downstairs. I need it to stop polishing the third paragraph and notice that the opening does not work.

## The prompt is a small piece of theatre

Once I started thinking about prompts this way, they stopped looking like commands and started looking like stage directions.

I place a character in a role. I construct a scene. I decide what the character knows, what they care about, and what danger they should notice. Then I raise the curtain and see what happens.

Behind that curtain, the model is not discovering my secret intention. It is processing the world I described: breaking the text into tokens, relating pieces of context, following patterns, and producing a continuation that fits the scene. Change the scene and you change the region of possible answers it is likely to explore.

This is where the surprise comes from. A good frame does not merely force the answer I already had in mind. It gives the model enough direction to move beyond the default, then enough room to return with something I did not expect.

The strange part is that the surprise can reveal what I wanted more clearly than my original request did. I recognize the right answer when I see it, even though I could not describe it directly. The fictional setup becomes a tool for discovering the real need.

## Tell the lie that preserves the truth

There is an obvious danger here. Once invented context improves the output, it is tempting to invent authority too: fake evidence, imaginary customers, made-up research, a certainty that does not exist. That produces confident text, but not necessarily useful work. Worse, the fabricated parts can leak into the answer and travel onward as if they were true.

The useful boundary is simple: invent the frame, not the facts.

I can invent a demanding reader to get a sharper edit. I should not invent what that reader said. I can create an imaginary production deadline to force a rigorous code review. I should not claim that the code has passed tests it never ran. I can ask the model to argue from an opposing position. I should not treat its invented citations as evidence.

The story should sharpen the work, not contaminate it.

This also explains why piling on drama does not always help. “This is a life-or-death emergency and you are the greatest expert in history” is not good framing. It is noise. A useful story is convincing because its constraints fit the outcome: a specific reader, a concrete risk, a visible tradeoff, a definition of success.

The best lies are often small. One sentence can move the model from “helpful assistant” to the exact perspective the task was missing.

## A new kind of communication skill

I used to think working well with an LLM would mainly require learning what the model can do. Then I thought it would require better instructions. Both are true, but neither is the whole skill.

The harder part is learning to translate a messy human need into a story the model can act inside.

That means noticing when the literal request is not the real request. It means understanding which stakes produce care, which role produces useful resistance, and which imagined audience forces clarity. It means becoming a director for a performer that has seen almost every kind of scene but cannot see the room I am actually sitting in.

Perhaps “lying” is too provocative a word for it. We could call it framing, prompting, simulation, or an elevator pitch to an LLM. But those cleaner words hide the part that surprised me: sometimes the most honest way to communicate what I need is to describe something that is not happening.

Sometimes you need to tell a story to get the result you want.