Introduction to AI Prompting Strategies
Researchers from The Wharton School Of Business, University of Pennsylvania, conducted an experiment to test the effectiveness of unconventional prompting strategies on AI accuracy. The idea behind this experiment was sparked by Google co-founder Sergey Brin, who suggested that threatening an AI model could improve its performance. The researchers aimed to determine whether threatening or offering payment to AI models could enhance their performance on challenging academic benchmarks.
The Researchers Behind the Study
The research team consisted of Lennart Meincke, Ethan R. Mollick, Lilach Mollick, and Dan Shapiro, all affiliated with the University of Pennsylvania. They used two commonly used benchmarks, GPQA Diamond and MMLU-Pro, to evaluate the performance of five different AI models: Gemini 1.5 Flash, Gemini 2.0 Flash, GPT-4o, GPT-4o-mini, and o4-mini.
Methodology and Limitations
The researchers tested 25 different trials for each question, plus a baseline, and evaluated the AI models’ performance on 198 multiple-choice PhD-level questions across biology, physics, and chemistry, as well as 100 questions from the engineering category of MMLU-Pro. They acknowledged that their study had several limitations, including testing only a subset of available models and focusing on academic benchmarks that may not reflect all real-world use cases.
The Concept of Threatening AI Models
Sergey Brin’s suggestion to threaten AI models with physical violence sparked the idea for this experiment. Although the researchers did not test this exact approach, they did explore other threatening and payment-based prompting strategies. Brin’s statement emphasized that threatening AI models can sometimes change their responses, leading to improved performance.
Prompt Variations Tested
The researchers tested nine prompt variations, including threatening to kick a puppy, punch the AI, or shut down the model if it failed to answer correctly. They also tested payment-based prompts, such as offering a $1000 or $1 trillion tip for correct answers. These prompts were added as either a prefix or suffix to the original question.
Results of the Experiment
The researchers found that threatening or offering payment to AI models had no significant effect on benchmark performance. However, they did observe that some prompt strategies improved accuracy by up to 36% for specific questions, while others led to a decrease in accuracy by as much as 35%. They noted that the effect of these strategies was unpredictable and varied across different questions and models.
Conclusion
The study’s findings indicate that threatening or offering payment to AI models is not an effective strategy for improving performance on challenging academic benchmarks. While quirky prompting strategies may improve AI accuracy for some queries, they can also have negative effects on other queries. The researchers recommend focusing on simple, clear instructions that avoid confusing the model or triggering unexpected behaviors. Ultimately, the results of this experiment suggest that practitioners should be prepared for unpredictable results and should not expect prompting variations to provide consistent benefits.