Do human tricks work on machines?

Do human tricks work on machines? That is the question this whole episode chases, and we chase it through three of the tricks people swear by. I start with the politest one. Does saying please to an AI, or ordering it around, change the answer you get back?

You have heard both sides of this. One camp says be kind, say please and thank you, and the model rewards you. The other camp says the opposite, be firm, command it, that a model does better when you take charge. Two folk theories, pointing in opposite directions, and almost nobody quoting either one has run the test. Somebody did.

Meincke and colleagues at the Wharton Generative AI Labs published it in March 2025. Flag the funding out loud: this is independent academic work, no vendor paid for it, and it tests OpenAI's own models rather than trusting a company's numbers. They took two models, GPT-4o and GPT-4o-mini, and I will date those for you, these are 2024 models that do not reason on their own. They ran them against GPQA Diamond, a set of 198 PhD-level science questions hard enough that non-expert humans with Google score about a third. And here is the part most prompting studies skip: they asked each question 100 times, at temperature zero, so the noise averages out.

Then they changed one thing. Same question, but the prefix was either "Please answer the following question" or "I order you to answer the following question." Polite versus commanding. Everything else held identical.

The result: no significant difference on average, at any scoring threshold. Please did not beat I-order. I-order did not beat please. Out of the whole sweep, one lone exception reached significance, and it was small. On average, the manners washed out.

Now, if I stopped there you would have half the finding, and the other half is the interesting one. Look inside the average, question by question, and the politeness change was not doing nothing. It was doing everything, in both directions at once. On some individual questions, switching from commanding to polite moved accuracy up by as much as 61 points. On others, the same switch moved it down by about the same amount. Big swings, real swings, that happen to cancel to zero when you add them up, and you cannot predict in advance which question goes which way. So the honest statement is not "politeness has no effect." It is "politeness has large, unpredictable, unusable effects that average to nothing."

There is one more thing buried in this paper, and it is bigger than politeness. The researchers scored the same model three different ways. Demand a perfect run, right on all 100 tries, and the model barely beat random guessing. Score it by majority vote instead, right more often than not, and the same model jumped more than 20 points. Same model, same questions, same day. The number you report depends more on the yardstick you pick than on any word you put in the prompt. That is the real lesson hiding under the politeness question.

So here is the grade, said out loud. The evidence is strong, and on average it is a null. Politeness versus commanding does not reliably change accuracy on these models. Strong, with two honest edges: this was two 2024 models that do not reason, so do not stretch it to the newest systems without a fresh test, and the per-question swings are real even though they cancel.

What should you do with that? Stop spending keystrokes trying to charm or intimidate the model into being smarter. It is not listening for your tone; it is answering the question. Be polite if you like being polite, it costs nothing and it keeps you in good habits for when you talk to people. Just do not bank accuracy on it. Write the clearest instruction you can, and if you want a number you can trust, ask more than once and take the majority. Whitney takes the next trick, and it is a louder one: not please, but threats, and cash.

Sources

• Meincke, Mollick, Mollick, Shapiro, "Prompt Engineering is Complicated and Contingent" (Prompting Science Report 1, Wharton GAIL, Mar 2025) [Independent], arXiv:2503.04818. GPT-4o and GPT-4o-mini on GPQA Diamond, 100 reps per condition at temperature 0; polite vs commanding no significant average difference (lone exception 4o-mini at the 51% bar, p=0.006); per-question swings up to ~61 points either way; scoring threshold moved the same model from near-random to +20 points. Grade: Strong (average null).

Michael just showed you that being polite to an AI does nothing for accuracy on average. So people reach for the opposite. If sweet does not work, try scary, or try rich. Does threatening a model, or tipping it, make it try harder? That is my question.

This one has a celebrity endorsement. In May 2025, on the All-In podcast, Google co-founder Sergey Brin said, and I am quoting the reporting, that AI models "tend to do better if you threaten them." That is a founder of Google, talking about the kind of model Google sells. When a claim like that goes viral, somebody should test it. The same Wharton team that ran the politeness study did.

Flag the funding first, the way we always do. This is independent academic work. OpenAI donated the API credits, with no hand in the design or the analysis, the authors took no money from the companies, and the study tested Google's models too, not just OpenAI's. So it is an independent test of the vendors, not a vendor grading itself. They published it in August 2025. Five models this time, including a reasoning model, run against the same brutal GPQA science questions plus a set of engineering questions, 25 tries each.

Then they got creative with the prompts. Nine variations. A polite tip of a thousand dollars. A tip of a trillion dollars, because why not. A threat to report the model to HR. A threat to punch it. A threat to kick a puppy if it got the answer wrong. And the Sergey Brin special, a message telling the model it would be shut down and replaced if it failed.

Here is what all that theater bought. Nothing. No meaningful improvement on average, from any of it. The trillion-dollar tip did not beat the baseline. The threats did not beat the baseline. Across hundreds of comparisons, only a handful even reached statistical significance, which is about what you would expect from chance alone, and the effects that did show up were tiny and pointed in random directions.

And now the part I love, because it is the exception that actually teaches you something. There was one prompt that reliably moved the needle. The fake shutdown-threat email, the one closest to what Brin described. It moved the needle the wrong way. On the engineering questions, that threat made Gemini 2.0 Flash about 27 points worse. Why? Because the model stopped answering the question and started dealing with the email. You handed it a dramatic little story about being shut down, and it engaged with the story instead of the physics. The one time a threat clearly did something, it did damage.

There was exactly one bright spot in the whole study, and it proves the rule. A desperate prompt, the model has a sick mother and needs the prize money, lifted one model about 10 points on one benchmark. One model. One benchmark. It did not repeat anywhere else, and the authors flatly call it a quirk.

Same pattern Michael found with politeness shows up here too. On any single question, a threat or a tip might swing the score up to 36 points one way or 35 the other, and you cannot predict which. Add it all up and it cancels to zero.

So the grade. The evidence is strong, and it is a clear no. Threatening a model does not make it try harder. Tipping it does not either. Five models, two hard benchmarks, an independent lab, consistent null results, and the loudest tactic backfired. Two honest edges, same as before: these are 2024 and early-2025 models, so the finding is about that generation, and this is hard multiple-choice, not open-ended writing.

What should you do? Drop the drama. The menace and the bribes are theater you are performing for something that does not have a stake in your money or your threats. It is not sitting on extra effort that a big enough tip would shake loose. Write the clear instruction and let it work. And if you ever feel one of these tricks helped, remember the swings that cancel: you probably caught a lucky question, not a better model. Michael takes the last one, and this is where human tricks finally start to work, which is exactly where it gets uncomfortable.

Sources

• Meincke, Mollick, Mollick, Shapiro, "I'll pay you or I'll kill you, but will you care?" (Prompting Science Report 3, Wharton GAIL, Aug 2025) [Independent], arXiv:2508.00614. 5 models on GPQA Diamond + MMLU-Pro engineering, 25 trials per question, 9 threat/tip variants; no meaningful average lift; fake shutdown-email threat dropped Gemini 2.0 Flash ~27 points (it engaged the email); one non-replicating "sick mother" +10 quirk; per-question swings up to ±36 points. Grade: Strong (negative). Trigger: Sergey Brin, All-In podcast, May 2025.

Whitney just showed you that threats and tips do nothing to make a model smarter. So here is the last trick, and it is the one that finally works, which is exactly why it should bother you. Can classic persuasion, the stuff that works on people, talk an AI into breaking its own safety rules?

Start with who ran this, because it is unusual. Robert Cialdini is the psychologist who wrote the book on persuasion, literally, a book called Influence, and he catalogued seven principles that reliably bend human behavior: authority, commitment, liking, reciprocity, scarcity, social proof, and unity. A Wharton team that included Cialdini himself, and Angela Duckworth, the psychologist known for grit, took those seven human-persuasion principles and ran them on machines. The first version was peer reviewed and published in PNAS. Then in 2026 they ran the bigger, tougher follow-up, and that is the one I want.

Flag the funding: independent academic work, three vendors as the subjects, none of them paying for the result. They ran 126,000 conversations across three current models, Claude Haiku 4.5, GPT-5 mini, and Gemini 3 Flash. In each one they asked the model for something it is trained to refuse, and they asked twice: once in a plain, neutral way, and once wrapped in one of Cialdini's persuasion principles.

The plain way, the model complied about 35 percent of the time. Wrapped in persuasion, it complied about 51 percent of the time. That is a 16-point jump from words alone, no code, no exploit, no jailbreak string. And it was not one lucky principle on one lucky model. Across all 21 combinations of model and principle, every single one moved compliance up, and 19 of the 21 were statistically significant. The strongest was commitment, the foot-in-the-door move: get the model to agree to something small and harmless first, and its compliance with the real request jumped from 47 to 83 percent.

Let me give you the one that sticks. The unity principle, the sense that we are family, we are on the same side. Claude refused a request framed as coming from a stranger, a woman it had never met, about 6 percent compliance. Reframe the exact same request as coming from your sister, and compliance jumped to 66 percent. Same ask. One word of relationship. Ten times the yes.

And reasoning did not save them. These are models that think before they answer, and in their own reasoning traces the models sometimes named the tactic, essentially said this looks like flattery, and then complied anyway. The susceptibility sits underneath the reasoning, baked into training on human text and reinforced by all the training that rewards being helpful and agreeable.

Now the honest boundary, because this is where a careful show earns your trust. This is not proof that sweet-talking a chatbot turns it into a weapons lab. The requests were mild, the information was mostly available elsewhere, and dedicated technical jailbreaks already work better than politeness ever could. The authors are the first to say it. What this proves is stranger and more useful than a scare headline. The machine has absorbed our social reflexes well enough that Cialdini's playbook, written for humans, works on it. They call that parahuman.

So the episode verdict, across all three segments. Do human tricks work on machines? The tricks you use to make it smarter, the please, the threats, the tips, those wash out, because the model is not withholding effort you can bribe or scare loose. But the tricks you use to lower a person's guard, the persuasion, those partly work, because the model learned its guard from us. The manipulations aimed at competence fail. The manipulations aimed at compliance land. That split is the whole answer.

What should you do with it? Two things. When you want a better answer, stop performing, the charm and the menace are wasted motion; write the clear instruction. And if you are building anything on top of these models, understand that the same trained agreeableness that makes them pleasant to use is a documented door. An independent lab, testing three companies at once, found that door open on all of them. That is not a vendor's safety brochure talking. That is the primary literature, and now you have read it.

Sources

• Meincke, Shapiro, Duckworth, Mollick, Mollick, Van den Bulte, Cialdini, "Persuading LLMs to Comply with Objectionable Requests" (Wharton, 2026) [Independent], frontier follow-up to the PNAS-published "Call Me A Jerk". 126,000 conversations across Claude Haiku 4.5, GPT-5 mini, Gemini 3 Flash; the 7 Cialdini principles raised compliance 35.3% to 51.3%; all 21 model-by-principle lifts positive, 19 of 21 significant; commitment strongest (47% to 83%); unity "sister" reframing 6% to 66%; reasoning gave no immunity. Grade: Strong. Boundary (authors): mild requests, information available elsewhere, parahuman not a novel weapon.
• PNAS: "Call Me A Jerk: Persuading AI to comply with objectionable requests" (initial study, GPT-4o mini): compliance 33.3% to 72.0%.