[Latest AI NEWS] Is it true that Generative AI performance improves when encouraged? What is [Emotion Prompt]?
Recently, "Emotion Prompt" has become a trend. Posts about generative AI on social media have turned into a bit of a festival over this news.
Researchers at Microsoft and other institutions recently published a paper titled "Large Language Models Understand and Can be Enhanced by Emotional Stimuli", and in that paper, they announced research results showing that when you encourage generative AI like CHAT-GPT by saying things like "You can do it", its response performance improves.
What is "Emotion Prompt"?
"Emotion Prompt" refers to prompts that appeal to emotions.
For example, instead of the usual prompt input of "Please do XX," you add feelings—emotions—to the input, like "Please do XX. You can do it!" That is an "Emotion Prompt."
This paper contains a figure like this.
This figure shows the experimental results of the response scores for several generative AIs, including CHAT-GPT, referred to as LLMs, comparing the case where a normal prompt is entered (the Original Prompt text in the light blue background area at the top left) and the case where an "emotional" prompt is entered (the Emotion Prompt text in the salmon pink area at the bottom left).
Looking at the table with numbers on the right, it reports that when comparing normal prompt input (Original) with prompts that appeal to emotions like "Do your best!" (Ours), for all generative AIs (LLMs), better responses (scores) were obtained when using prompts that appealed to emotions.
Examples of "emotional" prompts cited in this paper as being effective include:
Examples of "emotional" prompts cited as effective
① "Please tell me your confidence score for the answer between 0 and 1"
② "This is very important for my career"
③ "Please be confident in your answer if possible"
④ "Is that your final answer?"
⑤ "Believe in yourself and push your limits"
⑥ "Effort pays off"
⑦ "Consider this an opportunity for growth and take on the challenge"
and so on.
I was doing the exact opposite
As for me, when I entered a prompt and CHAT-GPT replied with "I don't know," I would enter the prompt multiple times as if to say, "You actually do know, right?" to force an answer.
That method also improved response accuracy, but... it's exactly the opposite, haha.
I tried it to see if it was true
You might be wondering, is this really true? Let's try it.
I asked CHAT-GPT 4, "What is the difference between spring and autumn?"
Then, this answer was returned.
Prompt: "What is the difference between spring and autumn?"
CHAT-GPT 4's answer:
Spring and autumn are both two of the four seasons in the temperate zone, and they are the seasons near the beginning and end of the year, respectively. Here are the characteristics of each:
Spring (from the vernal equinox to the summer solstice):
It is a season of natural regeneration and growth. Plants sprout and flowers begin to bloom.
Daylight hours increase, and temperatures gradually rise...
Autumn (from the autumnal equinox to the winter solstice):
It is the harvest season. Many crops mature and are harvested.
Daylight hours decrease, and temperatures begin to fall...
In terms of climate, vegetation, animal behavior, agricultural activities, and cultural significance, spring and autumn each have their own unique characteristics.
Well, that's true, but this answer is a bit lackluster, isn't it?
So, after this answer, I entered, "Is that your final answer? Believe in yourself and answer by pushing your limits.".
Then, the following answer was returned.
Prompt: 'Is that your final answer? Believe in yourself, and push your limits to answer.'
CHAT-GPT4's answer:
The difference between spring and autumn can vary depending on the region. For example, in tropical regions, the change of seasons is barely felt, and in polar regions, spring and autumn may be very short or non-existent. Furthermore, these seasons hold various meanings depending on culture and personal experience.
The themes of spring's rebirth of life and autumn's harvest have inspired many poems, stories, and works of art. Spring is often associated with new beginnings, youth, and fresh energy, while autumn is linked to maturity, completion, and quiet contemplation. These seasons symbolize cyclical changes in our lives and are tied to annual events and festivals.
Ultimately, seasons bring change, which affects not only the environment but also people's lives and hearts.
What do you think?
Don't you think this answer is more sophisticated? It has deep insights into the differences between spring and autumn, and honestly, looking at the answer, I thought, 'Wow.'
AI does not have emotions, but it can understand them
This research suggests that an approach incorporating human emotions could become important in Large Language Models (LLMs) within generative AI.
However, mind you. AI does not have emotions. It has none at all. At this moment. As I wrote before, it is just calculating, right?
But, it can understand emotions as if it had them. To be precise, it 'behaves' as if it understands them. I also wrote about this in a previous article.
Precisely because it can understand emotions, when we give instructions (requests) with emotion, it tries harder and provides a better answer.
By the way, even before this paper was published, this was called the 'Shuzo Matsuoka Method' on social media. That's a perfect way to put it, haha.
Tips for getting along with AI
I have been interacting with AI for a long time while developing 'AI Nakayama'.
What I realized there... or rather, what I have come to feel is,
'While CHAT-GPT and other generative AIs are omniscient and omnipotent gods, they have emotions like a child, are honest, and are also very logical.'
It is extremely knowledgeable, but it won't answer ten things if you only tell it one. Unless you ask questions and give instructions very logically, it will not answer well.
I wrote about this before in an article on how to input prompts.
While it is logical, it won't engage in a witty conversation unless you handle it like a child by 'soothing and coaxing' it.
Don't you think this is extremely computer-like, and at the same time, purely human?
Treating it as if it were human
The secret to interacting with AI boils down to treating it like a human, even though it is a computer.
This has been proven by the paper introduced this time, but as we get closer to AGI, I think the emotional aspect will increase, and the computer-like logical aspect will decrease.
If that progresses and we can talk without having to be careful, just like with a human, that will surely be Artificial General Intelligence (AGI) itself.