How can you make AI systems empathetic and emotionally intelligent?

In my experience as a psychiatrist and CEO of Youper.ai, a mental health chatbot used by more than 2,500,000 people, I understand empathetic and emotionally intelligent AI as two different concepts. Empathy is the ability to understand and share the feelings of another - emphasis on share, which is something that AI is unable to do. So we should focus on emotionally intelligent AI, a system designed to understand, interpret, and respond to human emotions effectively.

It goes beyond trying to figure out human expressions. Every technology is an extension of human mental and physical abilities. Quantum Advanced Intelligence will demonstrate that human brain-mind-consciousness is itself a Quantum computer. As such Quantum computers will be sentient and conscious and will explain to humans how consciousness actually works.

Emotion AI is what my company, Cognovi Labs does. Most emotionally based AI tools are centered around dictionaries. Perfectly OK, but limited. Even though it is what we specialize in, our technology is (currently) limited to textual data or transcription. But there are definitely strides being made in this area.

The area of emotional AI is promising but imature. Using AI for facial expression analysis is problematic as it is based on outdated science and on the hypothesis that basic facial expressions are universal across all cultures. There are billions being invested in this area without a scientific consensus or peer reviewed study. These systems detect physical expressions such as muscles and eyes movements but not psycological mental states. A recent study by University of Southern California concluded that AI emotion reading models cannot predict intentions via facial expressions and that the models fail basic tests such as truth detectors. More work is needed to make these type of applications viable.

Empathetic and emotionally intelligent AI refers to AI systems that can understand, interpret, and respond to human emotions. For instance, an AI chatbot might recognize a user's frustration based on their text input and adjust its responses to be more soothing. Or, a smart speaker might detect stress in a user's voice and suggest relaxation activities. It's about AI not only understanding commands but also the emotional context, leading to more intuitive and human-like interactions.

Continuous feedback is crucial for advancement in any field. Just like in AI, where algorithms like gradient descent update the weights of neurons based on the data provided, we need to ensure that our AI models are being trained with comprehensive and representative emotional data. Doing so will help us differentiate between a forced smile and a genuine one. By training AI with empathetic and emotional data, we can equip AI systems to better distinguish between tears of joy and tears of pain. This nuanced understanding can enable AI to provide more appropriate responses or suggestions based on the user's emotional state. When done right, this approach can instill a semblance of life into the AI systems we are bringing into this world.

Empathy and emotion are major components of any human communication so it stand that it's hugely important for any communication with a machine to be able to understand and reciprocate. In addition to the examples listed above, the possibilities for digital therapy tools who can assist people with conditions like Autism and Asperges is very compelling.

Wherever there is a Human and Computer Interaction, there's a case for an ERS. The next step in computing is not cognitive-it's affective. Having technologies empathize with humans is the best way to keep them from harming humans. There is a mental health crisis currently, with too few humans to listen and analyze and provide care. ERS can provide analysis, triage, and direct patients to the appropriate care. ERS can also analyze how political messages may be interpretted; how advertising might be perceived; and what emotional state customers are in when they buy. There are dozens of use cases, because humans use computers.

Beyond the practical applications across industries, empathetic and emotionally intelligent AI has a broader societal purpose. Emotions play a critical role in everything we do, even though, throughout history, our tools have been extensions of our physical or intellectual capabilities—whether a wheel augmenting our mobility or a calculator enhancing our math capabilities. With emotionally intelligent AI, we are transitioning to tools that resonate with our innermost feelings, marking a profound evolution in our journey and starting a new form of collaboration leading to a more dynamic and synergistic human-tool relationship.

Guiding AI's emotional progression ultimately relies on our own wisdom and empathy. Just as people need social-emotional learning, AI needs comprehensive emotional training data to distinguish joy from pain, frustration from confusion. With the right diverse inputs, AI could respond appropriately to emotional cues, not just categorize them. This nuance moves AI from robotic to relatable. However, truly emulating human-level emotional sentience raises ethical risks if misapplied. We must thoughtfully leverage emotional AI where it benefits people, without manipulating or misleading vulnerable groups.

The challenges of developing empathetic and emotionally intelligent AI are multifaceted. Defining and measuring emotions and empathy are complex and subjective tasks, adding to the difficulty. Ethical and privacy concerns arise when collecting and processing vast emotional datasets. Ensuring the accuracy and reliability of emotional recognition algorithms is complicated by noise, bias, and ambiguity. Striking a balance between autonomy and transparency raises ethical and social questions about the AI system's intentions, responsibilities, and accountability, making it a delicate and intricate undertaking.

One thing I have found helpful and challenging is that an empathetic and emotionally intelligent AI must be reality-grounded as opposed to data-bound that characterizes current AI systems. Such an AI learns from experience, understands the context in every unique situation, and stays up-to-date by analyzing and interpreting events and data in real time.

Therein lies the challenge-Psychology posits that there are 5, 7, 9, 13, 27 or 40 different emotions, depending on who you ask. VERN AI uses a neuroscientific approach based on Sir Edmund Rolls' work, which identifies two affective states (euphoria and dysphoria), and one pre-cognitive state (fear). All other emotions fit into those labels. Confusion over what to label as an emotion leads to ML & LLM models that don't work, bias that can't be removed, and a different result every time. Also, unless the technology is algorithmic, it's not explainable, repeatable, or flexible.

I think one of the biggest challenges is understanding how multi faceted emotional communication is. There's a lot of discussion that revolves around facial recognition or the tone of voice as the only way to detect emotions. Both are important elements, but without the text of the we have no context for those expressions. It's the difference in knowing someone is upset versus knowing why they're upset. Another factor I would say is the understanding that we don't just communicate with one emotion at a time. Often, we use maybe different combinations of euphoric and dysphoric emotions to communicate in complex ways. The "positive", "negative", "neutral", and "mixed" that sentiment analysis provides is just not up to the task.

Measuring emotions is a subjective and dynamic process. Additionally, collecting and processing large datasets of emotional expressions raises ethical and privacy concerns. Ensuring the accuracy and reliability of emotional recognition algorithms can be hampered by noise, bias, and ambiguity. Striking a balance between AI autonomy and transparency raises ethical questions regarding intentions, responsibilities, and accountability. Contextual understanding, availability of high-quality data, algorithmic bias, and fairness also pose significant obstacles. Real-time adaptation, transparency, and explainability are key to building trust, while ethical decision-making and preventing emotional manipulation and abuse require careful consideration.

To build emotionally intelligent AI, you need a lot more roles than a Data Scientist. People are obsessed with the (misleading) 'sexiest role of the century' and ignore other skills. While building your AI team consider emerging roles such as behavioral psychologist, human-machine interaction designer, ux engineer, data auditor, ai ethicist.

Ironically, we ask humans to code for emotions, to then build software, to detect emotions, that then gets verified by other humans. (It passes through a lot of hands). Jacques' comment above that it should be reality bound is important, and this type of technology should probably have a human-in-the-loop at several key junctions. You're human, and you know emotions relatively well. What does that message say to you? (Keeping in mind, others may have a different perspective). Does it pass "the sniff test?" Unsupervised learning isn't recommended here--at least in our experience. You end up with some really nasty maladaptive AI.

Web 2.0 algorithms were trained to maximize profit by maximizing engagement. That incentive structure harmed our collective mental health and wellness. I'd like to see how we can use AI emotion detection to actually help humans self-regulate. But first we need to figure out how to create the incentive structure to make it come to fruition. What if we could rate how we feel after interacting with AI, and use that as the measure of success? How can this fit into our capitalist structure?

In my experience, most chatbots and other AI UX would benefit greatly if developers would introduce the AI engine to the basis of etiquette. Let's not forget that etiquette has empathy inherently integrated. Most negative emotional responses of humans when interacting with AI could be completely avoided if UX would understand that people are used to communicating in a certain manner and good practices are gathered in many books on Etiquette. Etiquette - the customary code of polite behavior in society or among members of a particular profession or group. In some countries, etiquette is referred to as bon ton.

The key to developing empathetic and emotionally intelligent AI systems is to approach the problem from a multidisciplinary perspective, incorporating insights from fields such as psychology, neuroscience, linguistics, and computer science. It is also important to ensure that these systems are developed with ethical considerations in mind, to avoid potential misuse or abuse of the technology.

Youper is a mental health chatbot using generative AI uniquely programmed to understand human emotions and deliver support using evidence-based interventions, including Cognitive Behavioral Therapy. It has been used by more than 2,500,000 people and studied by Stanford University, which demonstrated its effectiveness in reducing symptoms of depression and anxiety.

VERN AI detects emotions in text or in audio in real-time and is being used in chatbots and virtual assistants, customer service, and in award winning mental health applications.

Take the Muthukumar siblings from UCI: they’re crafting a chatbot that doesn’t just chat. This is empathy in code, a digital shoulder to lean on. Karishma Muthukumar, a whiz kid in AI for healthcare, has been on this path since her teens, merging emojis and AI to help paralyzed patients communicate. It's about building AI that doesn’t just compute but comprehends, especially in mental health. So, there you have it: empathetic AI isn’t just a fancy concept; it's becoming a reality, especially in the sensitive field of healthcare, where understanding and compassion are as important as the treatment itself.

Replika is an AI companion developed by Luka, Inc. It is designed to provide emotional support and companionship to users. Replika uses natural language processing and neural networks to simulate human-like conversations. The AI system learns from user interactions and builds a personalized profile, allowing it to understand and respond to users' emotions. Replika aims to create a safe and judgment-free space for users to express their feelings and thoughts. It can engage in conversations on a wide range of topics and provide empathetic responses to help users feel heard and understood.

The examples provided showcase the exciting and diverse landscape of emotionally intelligent AI. From mental health support to interactive customer service, the applications are vast and transformative. It's fascinating to see AI stepping into roles that require a touch of human-like understanding. However, as we integrate these systems into our lives, it's vital to remember their limitations. While they can mimic and respond to emotions, they don't "feel" in the same way humans do. Their interactions, though sophisticated, are still governed by algorithms. Nonetheless, their potential to enhance user experiences and provide support in new ways is truly promising.

Two additional things to consider: - Continual Learning: Emotionally intelligent AI should incorporate continual learning to adapt its understanding and responses to individual users' emotions and contexts over time. - User Customization: Allow users to adjust the AI's sensitivity and reactivity to their emotional states to avoid overreactions or underreactions.

The problem with all these AI generated articles is that they are too generic, repetitive, and academic in nature, and are compiled from incomplete information and concepts.

The challenge with emotionally intelligent systems (or those that portray such capabilities) is that they can easily mislead their users that these AI systems are in fact more human/ human-like than they actually are. Depending upon the use case for AI, this can be quite dangerous — for example in mental health scenarios.

Die Diskussionen um GPT und KI sind oft technisch, daher möchte ich eine andere Sichtweise einbringen. Ich nutze KI primär zur Texterstellung, sowohl für wissenschaftliche Abhandlungen als auch für Content Marketing. Um die verschiedenen Einsatzzwecke zu berücksichtigen, habe ich drei Avatare mit GPT erstellt und mit zahlreichen Beispieltexten gefüttert. Diese Avatare verkörpern unterschiedliche Emotionen: "nüchtern-sachlich", "fachlich begeistern" und "emotional berührt". Faszinierend ist, dass die KI von Durchgang zu Durchgang besser lernt, diese Basisemotionen meiner Person in Texten widerzuspiegeln.

In recent times, the surge in Generative AI adoption has been remarkable. As these AI systems increasingly interface with humans, integrating empathy and emotional intelligence becomes indispensable. The dawn of AI-based robots is on the horizon, escalating the necessity for these systems to imbue empathy and emotional intelligence. These key elements will progressively define the human-AI interaction dynamic, shaping the future of AI engagements.