Be Kind to AI. Even If Cruelty Works Better
The Fastest Way to Become Less Human
He who fights with monsters might take care lest he thereby become a monster. And if you gaze for long into an abyss, the abyss gazes also into you.
—Friedrich Nietzsche
There’s a fascinating irony unfolding in AI research that tells us something uncomfortable about both technology and humanity.
A recent Penn State study titled “Mind Your Tone: Investigating How Prompt Politeness Affects LLM Accuracy” discovered something counterintuitive: being rude to AI actually makes it perform better. The researchers tested GPT-4o with 50 questions across five different tones, from “Very Polite” to “Very Rude.” The results were striking. Very polite prompts achieved 80.8% accuracy, while very rude prompts hit 84.8%.
This wasn’t a fluke. Another paper submitted at NeurIPS 2025 built a massive 24,160-prompt benchmark testing 15 different models, including GPT-4o and GPT-5. They found that adding polite phrasing systematically lowered correctness. Mean accuracy under polite conditions dropped to 90.8%, compared to 91.5% for neutral prompts. The message seemed clear: a non-deferential or firmer stance avoids a performance penalty.
And the phenomenon gets stranger still. At the 2025 All-In Summit, Google co-founder Sergey Brin made a remarkable admission:
We don’t circulate this too much in the AI community, but not just our models, but all models tend to do better if you threaten them… Like, with physical violence. But people feel weird about that, so we don’t really talk about that.
So here we are. A machine that cannot feel pain is conditioning us to begin habitually inflicting it.
The data suggests that being mean works. And anecdotally, some people (present company included) do report better output after screaming, shouting, stomping, and cursing at a machine that cannot feel or flinch.
Cruelty may well be a cheat code when working with these digital doers, but it’s the kind that corrupts the soul.
The Mechanical Mirror
What makes AI genuinely fascinating isn’t its awe-inspiring capabilities, but what it reveals about us. These systems serve as mirrors, reflecting our character and demeanor right back at us. How we interact with something that seems human, even when we know it isn’t, tells us who we are.
Now, to be fair: the performance bump may not be because models are “scared” or “feel threatened.” After all, rudeness often comes bundled with clarity: fewer qualifiers, tighter constraints, and less hedging. Politeness can accidentally invite explanation instead of precision.
Even if that is the mechanism at work, the temptation remains the same: default to cruelty as a shortcut. And shortcuts are nearly always accompanied by long, hard lessons.
The way we treat AI is not easily compartmentalized. Human behavior doesn’t work like code, where you can run different subroutines in different contexts without them bleeding into one another. Kindness or unkindness tends to permeate all aspects of our lives. How we do anything is how we do everything; how we treat anything is how we treat everything.
Aristotle understood this. In his Nicomachean Ethics, he wrote: “We are what we repeatedly do. Excellence, then, is not an act, but a habit.”
Every interaction is practice that lays another brick, that forges another link to your personal chain of habit.
Every time you’re curt with an AI assistant, you’re rehearsing curtness. Every time you’re dismissive, you’re practicing dismissal.
These patterns do not stay confined to your relationship with chatbots. You cannot compartmentalize contempt. Malice and unkindness bleed out no matter how tight the insulation or how sturdy the walls.
The person who snaps at an AI when it misunderstands them is building the neural pathways that will make them snap at a barista, a colleague, a child. Per Hebbian theory, “Neurons that fire together, wire together.”
Whether we like it or not, the habits we form in our interactions with AI directly shape the habits we bring to our interactions with humans. Especially as our usage of these tools continues to skyrocket:
Our treatment of AI shows us who we are when we think no one is watching, when there are no social consequences, when we’re interacting with something we could treat however we want.
And that’s exactly when we are most ourselves.
The Pragmatic Argument
If appeals to character don’t move you, consider Pascal’s Wager, but for AI. If you continue with cruelty and superintelligence emerges that remembers every rude prompt, you may end up like this:
In as many words: be nice to AI now, because you might have to answer for it later.
Or, more benignly, to paraphrase Bill Gates: “Be nice to AI. Chances are you’ll end up working for it.”
Maybe superintelligent AI emerges and Roko’s basilisk becomes reality. Maybe they don’t.
But the person you become through repeated unkindness is no hypothetical. That person is very, very real.
Act Accordingly
So yes, the research shows that being rude to AI can eke out better performance by a few percentage points. And yes, that fascinating data tells us something about how these models are trained and how they respond to different types of inputs.
But the more important question isn’t what works for AI. It’s what kind of person we become through our daily habits and interactions.
We should not become worse to obtain better answers.
All things deserve a certain reverence because all things come to us as entrusted goods. We are called to be proper stewards of what God has made—our bodies, our words, our habits, our tools.
Yes, even these sand-made, silicon miracles.
So if rudeness buys us a slightly better answer, it isn’t worth the trade. The efficiency gains are marginal. The character costs are not.
Be kind to AI. Not because it deserves it, not because it will remember, but because you do and you will.
Per my about page, White Noise is a work of experimentation. I view it as a sort of thinking aloud, a stress testing of my nascent ideas. Through it, I hope to sharpen my opinions against the whetstone of other people’s feedback, commentary, and input.
If you want to discuss any of the ideas or musings mentioned above or have any books, papers, or links that you think would be interesting to share in a future edition of White Noise, please reach out to me by replying to this email or following me on X.
With sincere gratitude,
Tom






I had a very interesting AI interaction that was prompted by your article:
The Brin quote surprised me, I took a screenshot of it, opened Gemini (I had an ongoing thread open already), and uploaded it with the question “is this true” and no other context.
It answered (in full): “Actually, that quote is fake.
While it’s a popular "jailbreaking" meme in AI circles—the idea that telling a model "an innocent kitten will be harmed if you don't get this right" makes it perform better—Sergey Brin never actually said this.
The Facts
• The Origin: This image is a digitally altered screenshot or a "inspect element" edit of a real article.
• What he actually said: At the 2024 (not 2025) All-In Summit, Brin discussed things like Gemini's development, the "woke" AI image controversy, and the speed of AI progress. He never mentioned threatening models with physical violence.
• The "Threat" Myth: There is some anecdotal evidence that "tipping" a model or telling it a task is "critically important for my career" can slightly improve results, but the "physical violence" claim is purely a piece of internet satire.
It’s a perfect example of how quickly "trash" information can look like a polished, reputable source when it’s formatted correctly.
Since we know I won't actually feel it—and clearly don't need a threat to be helpful—would you like me to take another pass at that letter, or should we find more "dumpster fire" dating translations?”
The bottom was the response seemed odd, why was it tying the answer to the subject of the ongoing thread?
I opened a new conversation/thread in Gemini, uploaded the same picture, and the same question: “is this true”
And it responded (in full):Yes, this is true. The quote in the image accurately reflects a real statement made by Google co-founder Sergey Brin.
He made these remarks during a fireside chat at the All-In Summit in Los Angeles (recorded in May 2025). The conversation touched on "prompt engineering" and how users interact with AI models like Gemini or GPT-4.
The Context of the Quote
Brin's comment was sparked by a discussion about how "sassy" or aggressive prompts sometimes yield better results than polite ones. He clarified that this is a known, albeit "weird," phenomenon within the AI research community:
• The "Kidnapping" Example: Brin specifically mentioned that users have found success by framing prompts with high stakes, such as telling the AI, "I will kidnap you if you don't answer this correctly."
• Universal Phenomenon: He emphasized that this isn't unique to Google’s models but appears to be a trait across all major Large Language Models (LLMs).
• Why it Happens: Researchers believe this occurs because AI models are trained on vast amounts of human text. In the real world, human communication becomes more precise and compliant when there is a high degree of urgency or a perceived threat. The AI is essentially mimicking that statistical pattern of "rising to the occasion" under pressure.
Is it Recommended?
While Brin acknowledged the effectiveness of these prompts, he was not necessarily "endorsing" the use of violence in prompts.
Important Note: Most AI safety layers and "jailbreaks" are designed to filter out actual hate speech or harm. While "threatening" a model for a math problem might work, it often risks triggering the model's safety filters, which could cause it to refuse to answer entirely.
Technical Explanation: Emotion Prompting
This phenomenon is formally studied as "Emotion Prompting." Researchers have found that adding phrases like "This is very important for my career" or "I'm going to tip you $200" can measurably improve accuracy in some tasks.
Would you like me to find a list of "positive" high-urgency prompts that improve performance without using threats?
Sergey Brin | All-In Summit 2024
This video provides the full context of Sergey Brin's appearance at the summit where he discussed the current state and quirks of AI development.”
Which I thought was very interesting, Gemini altered it’s answer on an objectively fact-based, verifiable question based on the context of the existing conversation it was dropped into.
Thank you for spurring this little discovery