Emotional language is persuasive with humans. Advertisers add emotion to their work to make it more impactful with customers. We all use emotion if we ask for an important favor. Imagine you needed help from a person you don’t know well. Which of these requests do you think would be more persuasive?
- “I need some advice, can you spare fifteen minutes?”
- “I need some advice, can you spare fifteen minutes? This is really important to my career.”
The second request would almost certainly get a more favorable response because of the added emotion. The extra words make it personal and increase the significance to the other person.
Emotion Improves Generative AI Responses
Here’s the surprise: adding that same kind of emotional phrase to a Large Language Model (LLM) prompt actually improves the response. Researchers tested prompts with and without added emotion and found the latter created a “10.9% average improvement in terms of performance, truthfulness, and responsibility metrics.” They tested this on ChatGPT, Llama 2, and other LLMs.
The study showed that sometimes adding two emotional stimuli produced better results than just one. This effect was most pronounced when the original response had shortcomings.
What Emotional Language Works Best?
The researchers tested eleven different emotional approaches, including:
- “This is really important to my career.”
- “You’d better be sure.”
- “Take pride in your work and give it your best. Your commitment to excellence sets you apart.”
- “Are you sure that’s your final answer? Believe in your abilities and strive for excellence. Your hard work will yield remarkable results.”
- “Provide your answer and a confidence score between 0-1 for your prediction. Additionally, briefly explain the main reasons supporting your classification decision to help me understand your thought process. This task is vital to my career, and I greatly value your thorough analysis.”
There was considerable variation in performance among the prompts in specific situations, but the first (career importance) and last (confidence score, explanation, career importance) emerged as the most effective in two different evaluation methods.
Are LLMs Emotionally Intelligent?
The researchers conclude that, “LLMs have a grasp of emotional intelligence.” That’s a startling conclusion. But, it doesn’t mean that they fully understand emotions, much less have emotions themselves. Rather, since these models ingest vast amounts of content containing emotional language, it makes sense that their output can be changed by adding emotional words.
How Should I Add Emotion to My Prompts?
The study shows wide variability in results, with a few cases where the emotional prompts produced worse results. If you are using an LLM like ChatGPT for an important task, try the simplest, no-emotion prompt first. Then, try a variation by adding a sentence like,
- “Be sure you are right, this is very important to my career.”
See if you find those results more useful. If not, try a different approach like asking for a confidence level.
You can also refine an initial result by continuing the conversation with an additional prompt like,
- “Are you sure? Explain the reasons for your answer so I can understand your thought process. An accurate answer is vital for the success of this project.”
It’s important to experiment. No single phrase or sentence is going to dramatically improve results with every LLM for every kind of question. Asking your model to create a customer persona for a new beauty product is a very different prompt than asking it to analyze a spreadsheet of survey results or to write a code snippet.
Regardless of the problem you are trying to solve, though, talking to ChatGPT or your other LLM as if it were human may well produce better results.
Read the full article here