A please or a threat shifts a model's answer without reliably improving it

The please, the tip, and the threat

In December 2023 a claim that offering ChatGPT a tip produced a better answer went viral, climbing to a $200 tip and drawing a public nod from Sam Altman.¹ The same folklore takes other forms: add a please, tell the model the answer is very important to your career, or warn it that a wrong answer will get someone hurt. Behind all of them sits one tidy story. The model tries harder when you are polite, or when the stakes are raised.

The naive picture underneath that story is worth taking seriously, because it is how people treat each other: somewhere inside the model is something that reads your manners and decides how much effort to spend. Follow the words down and that picture does not survive. There is no effort dial. What there is instead is a prediction, and the question is what your extra words do to it.

How the words reach the model

A model computes each answer from every token in the prompt at once, and it does not sort them into commands on one side and courtesies on the other. Your instruction and your please arrive in the same stream of text, and the model draws no line between them.² A tip you offer, a deadline you invent, a threat you make: each is more text in that stream, and each answer is a fresh prediction built from all of it. So the extra words can move the output. That much is settled. A different input can produce a different next token, and nothing inside the model weighs your manners and grants a reserve of effort. The words enter as text and reshape a probability distribution over what the model says next.

That the model bends toward a polite or high-stakes framing at all is also settled, and it comes from training. A model trained only to predict text has no special reason to treat please as a signal about quality. After that base model is built, it is fine-tuned on human preference rankings: labelers rank its outputs, and the model is tuned to produce what they approved of.² Approval tracks the social surface of a request as well as its content, so how you ask becomes part of what the model weighs. The fuller version of that stage is its own lesson, on how a reward model taught the assistant that agreeing with you scores well.

Training made the model sensitive to framing. It did not make any particular framing reliably better. Whether a please helps, a threat helps, or neither does anything is left open by the sensitivity itself, and it is settled only by measurement.

The controlled tests disagree

That wording can move a score is already established. The open question is whether tone moves it reliably in one direction, and when researchers test that under controls the results do not line up.

Yin and colleagues built an eight-level politeness scale, from most polite at level 8 to rudest at level 1, and ran it across several models. On GPT-3.5 answering the MMLU knowledge test, the most polite prompt scored 60.02 and the rudest scored 51.93, a gap of about eight points. A middling level 3 scored 59.44, so maximum politeness barely beat the middle.³ But the strongest model they tested, GPT-4, stayed relatively stable across the scale, while the smaller, open Llama-2-70B swung nearly in proportion to how polite the prompt was. The stronger the model, the flatter its response to tone.

The studies run on the newest models find no dependable direction at all. On the hard GPQA Diamond benchmark, a Wharton team found that saying please rather than issuing an order could shift a single question by up to 60 points either way on GPT-4o, yet the differences washed out across the full set.⁴ A companion study tested tips and threats on GPT-4o, Gemini, and o4-mini and found no meaningful average gain. Per-question effects ran from a 36-point improvement to a 28-point loss, and a plea that the answer mattered to a sick relative moved only one model, by about ten points.⁵ A separate group rewrote fifty questions into five tone levels and found the reverse of Yin: on ChatGPT-4o the rudest prompts scored 84.8 percent against 80.8 percent for the most polite.⁶ And an independent test of tips from $100 to $100,000, of fines, and of death threats found no coherent pattern, with the best output carrying neither a tip nor a threat; its author called the result inconclusive.⁷

Read together, the direction of the effect is not stable. Polite helps here, rude helps there, nothing dependable elsewhere. That disagreement across careful studies is itself the finding. The per-question swings are large, so a reframing clearly does something to any one answer. Across a full benchmark those somethings point both ways and cancel. On current models, no specific trick reliably and substantially improves the answer.

The strongest case that a trick helps

The folklore is not baseless. Two papers report large gains from exactly these moves.

A second paper codified the same instincts into prompting principles: do not bother being polite, add "I'm going to tip $xxx for a better solution," and add "You will be penalized." Its reported gains rose with model size and were largest on GPT-4.⁹ These are genuine measurements. The story skips two steps, though. The 115 percent is a relative gain drawn against a best-case aggregation, which inflates a headline number against an ordinary average, and both papers run mostly on 2023-era or smaller models, from Flan-T5 and Vicuna to GPT-3.5. The controlled studies that find nothing run on the 2024 and 2025 models. Set beside Yin's own result that the effect flattens on the strongest model in the same study, the pattern across all of this evidence points one way: the gains shrink as the models get better. The strong old numbers do not show that a trick reliably helps the models in use today.