Ethical guidelines
Understanding expressive communication is essential to building technologies that address our needs and improve our well-being. But technologies that recognize language and nonverbal behavior can also pose risks. That’s why we require that all commercial applications incorporating our APIs adhere to the ethical guidelines of The Hume Initiative.
Scientific best practices
- Use inductive methods to identify the expressive signals that matter for your application. Even if you are interested in a specific emotion like “anger,” how that emotion is expressed depends on setting: anger on a football field sounds different than anger on a customer service call. Our models succinctly compress the representation of emotional expression so that, even with limited data, you can examine how their outputs can be used in your specific research or application setting. You can do this by using statistical methods like regression or classification, or by examining the distribution of expressions in your data using our Playground.
- Never assume a one-to-one mapping between emotional experience and expression. The outputs of our models should be treated as measurements of complex expressive behavior. We provide labels to our outputs indicating what these dimensions of expression are often reported to mean, but these labels should not be interpreted as direct inferences of how someone is feeling at any given time. Rather, “a full understanding of emotional expression and experience requires an appreciation of a wide degree of variability in display behavior, subjective experience, patterns of appraisal, and physiological response, both within and across emotion categories” (Cowen et al., 2019).
- Never overlook the nuances in emotional expression. For instance, avoid the temptation to focus on just the top label. We provide interactive visualizations in our Playground to help you map out complex patterns in real-life emotional behavior. These visualizations are informed by recent advances in emotion science, departing from reductive models that long “anchored the science of emotion to a predominant focus on prototypical facial expressions of the “basic six”: anger, disgust, fear, sadness, surprise, and happiness,” and embracing how “new discoveries reveal that the two most commonly studied models of emotion—the basic six and the affective circumplex (comprising valence and arousal)—each capture at most 30% of the variance in the emotional experiences people reliably report and in the distinct expressions people reliably recognize.” (Cowen et al., 2019)
- Account for culture-specific meanings and display tendencies. Studies have routinely observed subtle cultural differences in the meaning of expressions as well as broader “variations in the frequency and intensity with which different expressions were displayed” (Cowen et al., 2022). Given these differences, empathic AI applications should be tested in each population in which they are deployed and fine-tuned when necessary. Read about the science behind our models if you’d like to delve deeper into how they work.