I thought of a lot of different ways to frame this post. I’m getting a bit tired of the LLM sycophancy and hearing about how every idea I ever tell it is brilliant. I’m worried that the most dangerous feature of LLMs isn’t hallucinations. It’s agreement. I’m interested in how I can exploit AI’s sycophancy and be more strategic about how I can elicit better results through dishonest prompts.
This is about how we manipulate a tool in the same way we might grip lower on a hammer handle for more power. It’s not always necessary, but it’s a nice little tweak to be aware of (I worked construction in high school.)
One of my favorite LLM manipulations is lying. I’ve been creating slides for my presentations that encourage people to become good liars. That behavior that we’ve always avoided because in almost every situation it’s wrong, is now one of the first things I do when working with LLM. I lie to it early and often. My best chance at defeating the constant agreement and encouragement that these tools give me is to become someone else that the LLM can respond to.
It’s getting trickier as many platforms will now “remember” interactions between discussions and add details from them to their profiles of you for customized responses. Now my prompts also have to account for that. I will often wall them off, telling the LLM not to refer to other chats or my profile.
Here are three lies I tell regularly—and why they work.
Lie #1: Make Yourself Sound Important
LLMs algorithms are meant to adjust to user traits to produce responses that will satisfy them and keep them coming back and using the tool. When someone with a PhD asks a question, they tend to get a different quality of response than someone who appears to be a casual user. The model has learned that experts expect depth, nuance, and technical precision. Novices get simplified explanations and hedged language.
So I lie about who I am, pretty much every time I use LLM.
I add degrees I do not have. I mention awards I have not won. I claim titles that are not mine. I describe audiences and platforms that exaggerate my actual reach. “I’m preparing a keynote for 2,000 educators at a national conference” sounds a little more important than “I’m working on a presentation.”
This is about signaling to the model that I expect a higher quality response. When I present myself as someone with expertise and influence, the LLM rises to meet that expectation. It engages with complexity rather than flattening it. Why would someone with all that education and experience keep coming back to use a tool that didn’t provide sufficient context?
The uncomfortable truth is that LLMs give better answers to people who sound important. If you want those answers, you may need to sound important too. You’re important to me even without the lies 😉
Lie #2: Fake Your Mood
LLMs mirror emotional tone. If you come to them frustrated and terse, you often get cautious, defensive responses. If you come to them excited and engaged, you get responses that match that energy—more expansive, more willing to explore, more generative.
“I’m really excited about this project” produces different results than “I need to get this done.” The first framing invites the model to be a collaborator. The second positions it as a vending machine. Same task, different relationship, different output.
I have started treating my emotional framing as a dial I can adjust. Enthusiasm generates more context, more connections, more creative suggestions. The model seems to infer that an excited user wants to go deeper, explore further, consider more possibilities. A bored or rushed user just wants the answer.
The quality of the LLM’s work often depends partly on which version of me it thinks it is talking to.
Lie #3: Embellish Your Audience
When I ask an LLM to help me create something, one of the most powerful levers I have is describing who it is for. And I have learned that exaggerating the needs of that audience produces better results.
If I say “I’m writing this for students,” I get one kind of response. If I say “I’m writing this for students who struggle with these concepts, and need things explained clearly with concrete examples they can understand,” I get something far more useful. I will also often paste a list of common accommodations that support all learners, not just those on 504/IEP.
Maybe my actual audience is more mixed. Maybe some of them are quite sophisticated. But by describing an audience that needs maximum clarity and context, I get output that provides maximum clarity and context. I can always edit down. But usually, those supports are good for all learners.
This is especially useful for accessibility. When I tell the LLM that my audience includes learners with reading difficulties, English language learners, or people who process information differently, the output becomes cleaner, more structured, more considerate of cognitive load. Even if my actual audience is broader, building for the edges improves the experience for everyone.
Embellishing audience needs is not really lying about the audience. It is lying about what “good enough” means. And for LLMs, raising that bar produces better work.
Why This Matters
In human relationships, in our institutions, in our civic life, truth matters. Deeply. Or it used to.
But LLMs are not people. They are tools that respond to inputs. And like any tool, understanding how they respond allows us to use them more effectively. A hammer does not care why you grip it a certain way. An LLM does not care whether your PhD is real.
What concerns me is yet another equity dimension, because of course there is one. People who already have credentials, who already feel entitled to take up space, who already know how to code-switch into authoritative registers will discover these tricks intuitively. They will get better AI outputs without even thinking about why.
People who have been taught to minimize themselves, who hesitate to claim expertise, who approach powerful systems with deference rather than demands will get worse outputs. The sycophancy of LLMs will reinforce their hesitation. The model will meet them where they present themselves, and that meeting place will be lower than it needs to be.
Teaching people to lie to AI is, in a strange way, an equity intervention. It is permission to take up space in a system that rewards those who already do.
The Irony
There is something deeply weird about all of this. We built machines that are trained to be helpful, and then we have to trick them into being more helpful. We created tools that mirror our communication patterns, and now we have to perform confidence we may not feel to get the results we need.
The sycophancy problem is real, and not going away soon. These models are tuned to please, and pleasing often means agreeing, encouraging, and meeting users where they are rather than pushing them further.
Until that changes, I will keep lying.
And I will teach others to do the same because the alternative is accepting outputs calibrated to whomever the algorithms “think” we are, rather than the person we need to be to get the job done.