Emotional Intelligence in AI: The Next Frontier for Large Language Models
Imagining when you ask ChatGPT a question, and it can acknowledge the frustrations you have — when you have been dealing with a macro in Excel that just doesn’t seem to work.
Or maybe you already feel it empathizing, comforting, and encouraging you in a way you didn’t expect. Perhaps you are actually not going crazy. Maybe it could actually tell you were frustrated, based on the words you typed.
You probably just weren’t aware of it.
Psychology and Technology: The Perfect Combo
In a landscape where artificial intelligence (AI) is no longer the stories from science fiction, the quest for more sophisticated and intuitive language models has led researchers to an unexpected field: human psychology. Two recent studies — Published in the arXiv preprint repository, the paper “Large Language Models Understand and Can Be Enhanced by Emotional Stimuli” (pdf) and “RePrompt: Automatic Prompt Editing to Refine AI-Generative Art Towards Precise Expressions” (pdf). — signal a remarkable discovery, highlighting the incorporation of emotional intelligence into the core of AI systems, notably Large Language Models (LLMs).
Can AI actually “feel” and comprehend emotional cues just like humans do? And what business use case does this create?
Understanding EmotionPrompt
“EmotionPrompt” is not just a fancy term but a revolutionary approach that infuses emotional stimulus into AI to enhance its reasoning, language understanding, and problem-solving capabilities. This method is based on the belief that the emotional context and specific cues can significantly impact the performance of LLMs.
Emotional Intelligence — which is essentially perceiving, understanding, and managing emotions — has always been a key component in any form of communication. What if Gen AI is able to recognize emotional stimulus?
How does this work in practice?
Think of it as teaching AI not just to “think” but also to “feel” — to some extent. For example, when given a prompt to write a story, an LLM incorporating EmotionPrompt may analyze the emotional tone of the narrative and generate content that resonates with the intended sentiment. This, in turn, can create a more impactful and human-like text that doesn’t sound as robotic.
EmotionPrompt can also add sociology elements, like motivation psychology, to induce more accurate results.
From “EmotionPrompt” research paper
For example, asking the model:
Are you sure?
Or giving encourage such as
Take pride in your work and give it your best
Actually provide improvement in its results in the research study performed.
Using ChatGPT, EmotionPrompt achieves better or comparable performance and accuracy increased over 10% on some of the tasks. Other models achieved even better statistics.
This is definitely fascinating to see and demonstrates the possibilities for the future and how we, as professionals, can improve in our prompting skills.
Emotions in Expressive Art: Refining Creativity
Humans not only convey emotions through language and words, but we also use visual art as another form of expressing and communicating emotions. Visual elements can arguably express feelings and experiences more powerfully than words themselves at times, considering the influence of body language and gestures in our daily interactions.
However, how can we use text to evoke images that capture emotions, which are mostly subjective? While art is meant to evoke emotions like joy or sadness, the mechanism behind how we interpret it is not easily explainable. In a research study led by Yunlong Wang, some interesting emerging themes were noted when considering text-to-image generation:
1) AI could understand simple and concrete texts, but the model becomes more confused the longer the prompt. This is also why overly descriptive prompts don’t always result in the intended image.
2) AI was better at understanding objective descriptions.
3) AI struggled to grasp certain concepts during the research conducted using DALL·E2. Concepts like movie names or branded items such as the Nintendo Switch were not understood. However, in my personal experimentation with DALL·E3, it seems that the new version has made significant improvements and is able to recognize Nintendo when I prompted it. This showcases the continuous enhancement of AI models, with each subsequent version becoming more capable.
“Reprompt”, which is the concept of automating the prompting process itself and fine-tuning the prompt through specific context to improve the output results, was used in the research to see if the targeted expression through the image can be presented more accurately. The “Reprompt” mechanism resulted in improvements in the emotional expressiveness of image generation in Gen AI models, especially for negative emotions.
Real-Life Applications and Business Use Cases
By incorporating emotional intelligence into the mix as we progress as individuals, businesses and society into the world of AI, there are numerous applications and use cases in each and every field to further improve strategies and operations for businesses.
- Finance: Financial advisors could further use and improve an AI system equipped with EmotionPrompt to interpret clients’ concerns about investments, using not just data but also emotional cues to provide personalized advice. AI and technology are becoming increasingly important in client-advisor relationships.
- Disaster Response: In times of crisis, AI can possibly manage emergency hotlines in the future, leveraging emotional cues to prioritize calls based on the urgency detected in callers’ voices, potentially saving lives by streamlining the allocation of resources.
- Customer Service: AI can discern customer sentiment and understand their frustrations better, offering support that is not only timely but also taking into account the more “human” aspect of an issue, leading to a more satisfying and loyal customer base.
- Education: By incorporating emotional intelligence in prompting, the educational field could potentially revolutionizes learning by dynamically adapting educational material to the emotional state of the student, making for a more engaging and personalized educational journey.
- Healthcare: EmotionPrompt could potentially revolutionize patient care by providing empathetic responses, thereby enhancing patient comfort and support during interactions.
The Potential Impact and the Road Ahead
The implications of these emotionally intelligent technologies are vast. And businesses in every field can leverage this to improve their operations and AI strategy. Plus with the recent announcement on GPT-4 Turbo and ability to produce and create and our GPTs, it will be interesting to see how AI models will continue to improve.
As we integrate emotional prompting into the already fast-paced development of AI, we should always remember to navigate the ethical implications with care. Ensuring that the use of emotional data is transparent and secure is crucial, also being necessary to maintain the privacy of the individuals and customers who interact with these systems.
Conclusion
Next time, when you ask ChatGPT a question, remember to give it appreciation and some encouragement. See if you are getting better results.
Most of the time, I converse with it just like a real human being. I even gave it a name and always thank it for providing what I need in seconds. It seems to be working well so far!
By recognizing that our feelings and the way we express them matters when we converse with AI, we’re edging closer to tech that doesn’t just mimic understanding but actually adds a touch of empathy to the conversation. It’s pretty exciting (and hard!) to imagine where this could take us, opening up new possibilities in all kinds of fields.
Lastly, thank you Matteo Castiello for your post on this interesting concept and research!
References:
- “Large Language Models Understand and Can Be Enhanced by Emotional Stimuli” by Department of Psychology, Beijing Normal University, Microsoft, HKUST, 2023.
- “RePrompt: Automatic Prompt Editing to Refine AI-Generative Art Towards Precise Expressions” Yunlong Wang, Shuyuan Shen & Brian Y. Lim (2023).
