Does Saying “Please” to ChatGPT Actually Get You Better Answers?

This brief looks at whether polite prompts actually improve AI output, drawing on three studies that found genuinely mixed results: older models reward moderate politeness, but newer models like GPT-4o performed slightly better with blunt prompts, and in every study extreme tone in either direction (fawning or rude) hurt performance. It explains the mechanism (politeness nudges a model toward the more formal register common in its training data, which helps for complex drafting but can introduce hedging that muddies simple factual answers), notes the real infrastructure cost of padded prompts at scale, and argues banks shouldn't write policies mandating or banning politeness. Instead, it recommends training employees on what actually improves output: specificity, giving the model a role, providing an example, and naming the desired format.

Roughly 70% of US users say they reflexively type “please” and “thank you” into ChatGPT and Copilot. Almost nobody stops to ask whether it actually helps. Now that three peer-reviewed studies have tested the question directly, the answer turns out to be more interesting than a simple yes or no.

Why This Question Matters for a Bank

For an institution rolling out Copilot or Claude across hundreds of seats, prompt quality is the difference between a tool that saves a loan officer ten minutes on a credit memo and one that produces something that has to be rewritten from scratch. If politeness genuinely improves output, it belongs in acceptable-use guidance. If it doesn’t, employees should know that too, and stop spending effort on padding that doesn’t move the needle.

There’s also a cost angle worth naming. OpenAI’s Sam Altman acknowledged that polite phrasing has added tens of millions of dollars to the company’s annual electricity bill across its user base, and called it money well spent. For a bank running a few hundred seats the dollar cost is trivial, but the underlying question, whether courtesy is functional or just ornamental, is worth answering properly.

What the Research Actually Found

Three studies have approached this seriously, using different models, languages, and task types, which is exactly why their findings diverge:

• A cross-lingual study from Waseda University tested eight prompt templates per language, from highly polite to outright rude, across English, Chinese, and Japanese prompts on GPT-3.5, GPT-4, and Llama-2. Impolite prompts consistently hurt performance, but heavily formal or flattering phrasing didn’t help either. Moderate politeness performed best, and the ideal level varied by language, suggesting tone sensitivity is tied to a model’s training data.
• A Penn State study, “Mind Your Tone,” rewrote 50 multiple-choice questions into five tone variants and tested all 250 prompts on ChatGPT-4o. The surprising result: rudeness won. Very polite prompts scored 80.8% accuracy; very rude prompts scored 84.8%. The likely explanation is that polite framings introduce hedging that adds ambiguity, while blunt prompts strip out padding and state the request more clearly.
• A follow-up cross-model study extended this work to more models and task types. The takeaway: tone sensitivity varies by model generation. Older, GPT-3.5-era models reward politeness, while newer models like GPT-4o and o1 show flatter response curves and are less affected by tone overall, with effects sometimes running in the opposite direction.

Why Does Tone Affect a Machine at All?

The model has no feelings; it isn’t flattered, offended, or motivated. The mechanism is statistical, not emotional. Large language models predict text based on patterns in their training data, and polite, formal language tends to appear in more carefully written sources like academic papers and professional correspondence, while curt language shows up in shorter, more direct contexts. Prompting politely nudges the model toward that richer, more carefully reasoned register; prompting bluntly nudges it toward direct, declarative output.

That distinction matters for task type. For complex drafting, like a credit memo or policy summary, the richer register that politeness invokes can help. For binary factual recall, like looking up a regulatory timeline, that same richness can introduce hedging that obscures the answer. There’s no single rule because the effect depends on what you’re asking the model to do.

The Cost Nobody Mentions

Every word in a prompt is a token, and every token consumes compute. At platform scale, politeness phrases like “please,” “thank you,” and “if you don’t mind” add up to real infrastructure cost. At a single institution running a couple hundred seats, that marginal cost is invisible, just pennies. But the underlying principle scales: words that don’t change the answer are tokens you’re paying for and time your employees are spending. In high-volume workflows like drafting, summarization, and batch analysis, trimming that padding can compound into genuine productivity gains.

What This Means for Your Bank

Banks rolling out Claude Team or Copilot shouldn’t write acceptable-use policies that mandate or forbid politeness; that’s not where the real productivity lives. What actually improves AI output:

• Be specific about the task. “Summarize this credit memo in 150 words for the loan committee” beats “Could you please summarize this?”
• Give the model a role. “Act as a BSA analyst” anchors the response far more than any tone ever will.
• Provide an example. One short example of the format you want outperforms any amount of polite phrasing.
• Specify the output format. Naming a table, numbered list, or executive summary changes results dramatically.
• Avoid extremes either way. Excessive flattery and outright rudeness both degrade performance. Neutral and clear wins.

Looking Ahead

This research will keep evolving as models do. The Penn State finding that rude prompts outperformed polite ones was itself a reversal of earlier guidance, and as models get better at parsing intent, tone effects will likely shrink further. What won’t change is the underlying principle: clear, specific, well-structured prompts beat any amount of social padding.

For a bank, the practical work isn’t policing whether employees say please. It’s giving them a short, plain-language guide to writing good prompts, building it into onboarding for any AI tool, and making sure the model is doing real work rather than generating filler someone still has to rewrite. The most important finding across all three studies is the one that gets the least attention: excessive politeness consistently hurt performance. Flowery preambles add tokens, introduce ambiguity, and signal to the model that the request itself is uncertain. That’s the habit worth training out of your team.

The Bottom Line

Be polite to AI if you want to. There’s a reasonable workplace-culture argument that habits of rudeness toward machines can bleed into rudeness toward colleagues. But the productivity gain isn’t coming from the “please.” It’s coming from specificity, role, examples, and format. Train your team on what actually matters.