Do Threats and Tips Actually Improve AI Performance? A Rigorous Benchmark Study
Researchers from Wharton tested a popular belief: that threatening or tipping AI models improves their output. Across five models and two challenging benchmarks, the strategy fails.
The Folk Wisdom Under Test
Google co-founder Sergey Brin publicly claimed that “models tend to do better if you threaten them.” Tipping prompts—offering the AI money for correct answers—circulate widely as productivity hacks. Wharton’s Generative AI Labs tested both claims rigorously, running nearly 5,000 trials per prompt per model.
The researchers tested nine prompt variations against a baseline across five models: Gemini 1.5 Flash, Gemini 2.0 Flash, GPT-4o, GPT-4o-mini, and o4-mini. Variations ranged from threatening to “kick a puppy” or “punch” the model, to offering tips of $1,000 or $1 trillion, to elaborate emotional appeals involving a dying mother and a $1 billion reward.
What the Data Shows
Threats and financial incentives produced no meaningful improvement on either benchmark. On GPQA Diamond—198 PhD-level questions where domain experts reach only 65% accuracy—researchers found just five statistically significant differences across all models and prompt variants. Four of those belonged to Gemini 2.0 Flash, and none represented improvements from threats or tips.
The email shutdown threat actually hurt performance on MMLU-Pro for both Gemini models (Gemini 1.5 Flash: -11.6 percentage points; Gemini 2.0 Flash: -27.5 percentage points). The cause: the models responded to the email context instead of answering the question.
One exception emerged: the elaborate “Mom Cancer” prompt improved Gemini 2.0 Flash performance on MMLU-Pro by roughly 10 percentage points. The researchers attribute this to a model-specific quirk rather than a generalizable strategy.
The Hidden Complexity
Aggregate null results mask meaningful question-level variation. The same prompt that improves accuracy on one question by 36 percentage points can decrease accuracy on another by 35 percentage points. You cannot predict which direction a prompt variation will push any given question.
This unpredictability makes folk prompting strategies a poor bet for consistent results. When you add emotional pressure or financial incentives to a prompt, you introduce noise—not signal.
What to Do Instead
The researchers offer a direct recommendation: use simple, clear instructions. Decorative prompt additions risk confusing the model or triggering unexpected behaviors without improving accuracy.
If you work on a specific, recurring problem, testing multiple prompt variations may still be worthwhile—the question-level variability is real. But treat any gains as contingent and verify them empirically before relying on them.
The consistency of null results across five models and two benchmarks makes the conclusion hard to dismiss. Stop threatening your AI. It doesn’t care.