Hume CEO Alan Cowen on Creating Emotionally Aware AI

Cognitive Revolution "How AI Changes Everything"
9 Dec 202377:20

TLDRIn this episode of the Cognitive Revolution podcast, host Nathan Leens interviews Alan Cowan, CEO and Chief Scientist at Hume AI, a company focused on developing AI that prioritizes human happiness and well-being. Cowan discusses the challenges and strategies involved in training AI to understand and predict emotional responses accurately across diverse populations. He emphasizes the importance of measuring and optimizing for human well-being, rather than just short-term outcomes like financial gain. The conversation delves into the nuances of emotional measurement, the potential applications of AI in various fields, and the ethical considerations that come with building emotionally aware AI systems.


  • 🧠 AI's potential to surpass human understanding of emotional importances is vast, given its access to more data than any human has seen.
  • 📈 The goal of AI should be to optimize for human well-being and happiness, which involves predicting and understanding emotional responses better than any human could.
  • 🤖 The development of emotionally aware AI is a high-stakes part of the cognitive revolution, with significant implications for how AI technology will transform work, life, and society.
  • 🌐 AI systems should be designed to prioritize human well-being over other objectives, such as increasing bank account balances.
  • 🔄 There's a need for AI to continuously learn and adapt in the environment while avoiding actions that could lead to negative emotional outcomes for humans.
  • 🌟 The key to AI alignment lies in optimizing longitudinally for well-being, ensuring that AI actions are beneficial in the long term.
  • 🚀 Despite concerns about the pace of AI development, there's optimism that the problem of aligning AI with human interests can be solved through careful design and ongoing learning.
  • 🧬 The AI race may have slowed down due to less sharing of core ideas, which could inadvertently give us more time to figure out how to control and align AI properly.
  • 📊 AI's ability to predict human well-being should be as accurate as its ability to predict other outcomes, and the data for such predictions exists and can be utilized.
  • 🌍 The future of AI involves building systems that not only understand emotional affordances but also act upon them to enhance human well-being and happiness.

Q & A

  • What is the main goal of Hume AI in developing its platform?

    -The main goal of Hume AI is to teach AI to make people happy by providing tools to measure and optimize for human well-being, focusing on emotional expressions and understanding how actions affect human happiness and satisfaction in the long term.

  • How does Alan Cowan define happiness and well-being?

    -Happiness is defined as the in-the-moment experience of positive emotions, while well-being is a broader concept that includes happiness, satisfaction with life, and the reflection on one's experiences of being happy.

  • What are the challenges in collecting emotional response data?

    -Challenges in collecting emotional response data include accurately measuring human emotions across diverse populations and cultural contexts, understanding which emotional responses to optimize for, and maintaining high-level control as AI systems begin to access not just verbal and facial responses but also brain states directly.

  • How does Hume AI approach the development of emotional measurement models?

    -Hume AI approaches the development of emotional measurement models by building platforms that capture multifaceted emotional experiences, including text, image, audio, and video data, and using techniques like semantic space theory to better understand and measure emotions in everyday interactions.

  • What is the significance of understanding emotional affordances in AI?

    -Understanding emotional affordances in AI is significant because it allows the AI to predict how its decisions will affect human well-being, enabling it to prioritize actions that lead to positive emotions and outcomes over those that might result in negative emotions or states.

  • How does Hume AI ensure the ethical use of its technology?

    -Hume AI ensures the ethical use of its technology by establishing an Ethics Committee, developing guiding principles for empathic AI, publishing best practices for measuring well-being across various risk use cases, and articulating a list of supported and unsupported use cases.

  • What are some of the applications of Hume AI's technology?

    -Some applications of Hume AI's technology include improving user experience, mental health monitoring, customer support outcomes, and driver monitoring for signs of drowsiness or distraction.

  • How does Hume AI handle the potential for bias in its models?

    -Hume AI handles the potential for bias by using a scientifically controlled survey platform to collect diverse data, randomizing tasks, and training models to extract objective measures of emotional expressions that are independent of demographic factors such as age, gender, and ethnicity.

  • What is the role of reinforcement learning in Hume AI's approach?

    -Reinforcement learning plays a role in Hume AI's approach by allowing the AI to learn from people's reactions in videos, using those reactions as feedback to improve the AI's language generation and its ability to evoke desired emotional responses.

  • How does Hume AI's custom model API work?

    -Hume AI's custom model API works by allowing users to upload their own labeled data, which the API then uses to train a model that predicts the desired outcomes. The API provides an endpoint for users to deploy the model and receive predictions based on their specific data and use case.



🤖 The Future of AI and Emotional Intelligence

The paragraph discusses the development of AI with pre-existing knowledge that surpasses human understanding of emotional importances in everyday interactions. The ideal AI system is envisioned to predict human well-being better than material gains, such as increasing bank account balances. The speaker expresses concern but also optimism about the potential of AI to solve complex problems related to human emotions and well-being.


🌟 Defining Happiness and Well-being in AI

This section delves into the distinction between happiness and well-being. Happiness is described as the momentary experience of positive emotions, while well-being encompasses life satisfaction. The speaker emphasizes the multifaceted nature of measuring well-being, including the richness of emotional experiences. The conversation highlights the importance of emotion as a fundamental component of human interaction and the goal of aligning AI technology with human well-being.


📊 Understanding Emotional Expression through Data

The paragraph discusses the challenges of assembling data sets for emotional AI, focusing on self-reported emotions and the importance of controlling for biases such as gender, age, and ethnicity. The speaker describes a large survey platform used to collect diverse data, ensuring that the AI model is trained to understand facial expressions and vocal tones independently of identity and context.


🌐 Cultural Nuances in Emotional Expression

This section explores the cultural universality of certain emotional expressions and the importance of training custom AI models to understand cultural nuances. The speaker mentions studies that show commonalities across cultures in expressing emotions like anger and fear, and the need to account for cultural variations in the expression of emotions like positivity and arousal.


📈 Measuring and Predicting Emotional Outcomes

The paragraph discusses the process of measuring and predicting emotional outcomes using AI, including the use of various models for speech prosody, vocal bursts, and facial expressions. The speaker explains how these models capture hundreds of dimensions per second to make predictions, and how the addition of non-linguistic information improves the accuracy of predictions related to human interaction, preferences, and well-being.


🧠 The Science of Emotion Measurement

This section delves into the scientific challenges and methodologies of measuring emotions, including the development of a new theory called semantic space theory. The speaker discusses the complexity of emotional expression and how it's captured in different dimensions, which are then used to predict outcomes such as driver drowsiness or user experience in product studies.


🌟 The Role of AI in Enhancing Positive Human Experiences

The paragraph discusses the potential of AI to enhance positive human experiences by teaching it to make people happy, thereby paving the way for pro-social AI. The speaker talks about the importance of building platforms that measure and optimize for human well-being and the ethical considerations involved in developing empathic AI.


💡 The Intersection of AI and Emotional Intelligence

This section explores the intersection of AI and emotional intelligence, discussing the development of AI toolkits to understand emotional expressions and align technology with human well-being. The speaker emphasizes the importance of measuring emotions accurately and the potential applications of AI in improving user experience, mental health, and customer support outcomes.



💡Artificial Intelligence (AI)

Artificial Intelligence refers to the simulation of human intelligence in machines that are programmed to think and learn like humans. In the context of the video, AI is being developed to understand and predict human emotions and well-being, aiming to improve human life by making AI systems that are empathetic and pro-social.

💡Emotional Intelligence

Emotional intelligence is the ability to recognize, understand, and manage our own emotions and the emotions of others. In the video, emotional intelligence is a crucial aspect of AI development, where AI systems are being trained to detect emotional cues from human expressions and language to better interact and enhance human experiences.

💡Cognitive Revolution

The Cognitive Revolution refers to the significant changes and advancements in the field of artificial intelligence and cognitive science, leading to the development of AI systems that can perform tasks requiring human-like cognition. In the video, the term is used to describe the transformative impact of AI on various aspects of work, life, and society.


Well-being refers to a state of being comfortable, healthy, and happy. In the context of the video, well-being is a central goal for AI systems, as the technology aims to optimize and improve human emotions and life satisfaction beyond just momentary happiness.

💡Emotion Recognition

Emotion recognition is the process of identifying and understanding emotions based on facial expressions, vocal tones, and other non-verbal cues. In the video, emotion recognition is a key technology that AI systems are being developed with, to accurately detect and respond to human emotional states.

💡Semantic Space Theory

Semantic Space Theory is a concept in cognitive science that proposes a mathematical model where the meaning of words is represented in a high-dimensional space based on their relationships and associations. In the video, Alan Cowan mentions introducing a new theory about measuring emotion using semantic space theory to better understand and categorize human emotions.

💡Pro-social AI

Pro-social AI refers to artificial intelligence systems designed to promote positive social behaviors and outcomes. In the video, pro-social AI is the ultimate goal for Hume AI's research and development, aiming to create AI that contributes to the happiness and well-being of individuals and society.

💡Ethics Committee

An Ethics Committee is a group of individuals who review and guide ethical standards and practices, often within a specific field or organization. In the context of the video, the Ethics Committee at Hume AI is responsible for developing guiding principles for empathic AI and ensuring that AI technologies are used ethically and responsibly.

💡Multimodal System

A multimodal system refers to a technology or method that involves multiple modes of communication or data input, such as text, speech, images, and facial expressions. In the video, the development of a multimodal system for AI enables it to understand and respond to a broader range of human emotions and expressions more accurately.

💡Longitudinal Data

Longitudinal data refers to data that is collected from the same subjects over a long period of time, allowing for the analysis of changes and trends over time. In the context of AI, longitudinal data is essential for training models to understand the long-term effects of actions on human well-being.


The development of emotional aware AI is one of the highest stakes parts of the cognitive revolution.

Hume AI is focused on teaching AI to make people happy, paving the way for pro-social AI.

The ideal of AI is to be better at predicting human well-being than increasing one's bank account balance.

Emotional importance in everyday interactions is informed by data that AI has learned from.

Hume AI has established a 10-member Ethics Committee and developed six guiding principles for empathic AI.

The company has published best practices for measuring well-being across various risk use cases.

Hume AI's emotion recognition technology works by overlaying an emotional recognition layer on conversations.

The platform measures and optimizes for human well-being in a multifaceted way, capturing the richness of emotional experiences.

Emotion is an essential component of all human interaction, and AI should understand this to improve user experience and mental health outcomes.

Hume AI's technology can measure speech prosody, vocal bursts, and facial expressions to understand emotions better.

Adding non-linguistic information like prosody and facial expressions can significantly improve predictions about human interaction outcomes.

The company's technology can accurately predict emotions from text, image, audio, and video inputs.

Hume AI's custom model API allows users to upload their data and receive predictions based on their specific needs.

The platform can predict a wide range of outcomes, from detecting drowsiness in drivers to assessing mental health in video diaries.

Hume AI's approach to emotional measurement involves understanding the relationship between short-term happiness and long-term well-being.

The technology can be used to improve customer support outcomes by understanding the emotional states of participants in conversations.