AI Emotion Tracking: Not Just in Science Fiction

The idea of computers being able to recognize emotions was first introduced by Rosalind Picard in her 1995 book Affective Computing. The computer scientist stated that if we want genuinely intelligent computers, we must give them the ability to recognize and understand emotions. Today, that describes the ability of emotion AI.

But how does AI learn to read your emotions? In technical terms, it can fit in one sentence: by training on datasets where emotions are represented as statistical, measurable units. However, emotions are complex and rarely fit predefined labels. A smile doesn’t always mean we are happy.

In this article, I explore AI’s ability to read emotions, its limitations, and how it affects our lives.

Key takeaways

AI emotion tracking gained popularity amid the COVID-19 pandemic, as it was used to help measure passengers’ emotions at Athens International Airport, while a significant number of employers in the US started using emotion AI for productivity-monitoring of remote workers.
Emotion recognition systems continue to be used in airports to track passenger satisfaction and offer tailored services.
Emotion AI faces restrictions and even bans in certain contexts due to its limited emotion-tracking capabilities and biases against some demographic groups.
AI emotion monitoring systems analyze highly sensitive biometric data, including facial expressions and voice, which can be easily exploited.

What is AI emotion tracking and how it works

Emotional AI (EAI) combines artificial intelligence techniques and natural language processing (NLP) to evaluate emotional signals. This technology allows interpretation of facial expressions, vocal tone, and even spoken or written feedback.

This means that EAI machines can listen to voices and pick up cues that correlate with stress or anger. Also, they can analyze images and, based on subtle micro-expressions, recognize other emotions.

From airports to offices: where emotion AI is already watching

The main accelerator for AI emotion tracking was the COVID-19 pandemic. When our living rooms became offices, some employers opted for Workplace AI, including emotion monitoring tools, to track employee productivity. Meanwhile, similar technology was applied at airports to monitor passengers' emotions. While airports continue to successfully apply emotion AI to improve the user experience, employees in other industries report a negative impact on their well-being.

AI emotional recognition systems at the airports

In 2021, Athens International Airport launched its Passengers’ Emotional Engagement measuring initiative. The goal of the initiative was to measure passengers' emotions and stress during the pandemic. Now, you might be imagining a dozen cameras scanning and analyzing every movement of your facial muscles at once. However, in practice, this technology was applied to analyze data gathered from various channels, including call centers, social media, comment management, and customer care.

Using AI emotion-tracking helps airports understand which factors increase consumer spending. For example, it aids in identifying what drives passengers' emotions and, based on that, airports can create better experiences.

Emotional monitoring in the workplace

The pandemic also led to a surge in workplace monitoring tools as offices transitioned to remote work. As of today, about 67% of employers continue to use various online productivity-tracking tools, including emotion-monitoring tools.

The workplace AI claims to improve performance, but the research has found the opposite. In fact, a study finds that employees under emotional surveillance identified negative impacts on their well-being, including loss of privacy and autonomy, and decreased performance.

Ironically, what emotion AI technology claims to improve, such as productivity and well-being, workers report having decreased. Emotional surveillance caused anxiety and distraction, leading to reduced performance. Their well-being was also affected negatively, as the workers engaged in emotional labor, changing how they feel and display their emotions.

Emotion AI may hold racial and gender bias when applied in high-stakes decision making processes

While there’s no recent data available, as of 2023, over 50% of large employers in the US used emotion AI to infer with their employees internal states. For example, call centers use AI emotion tracking to monitor their operator's tone of voice and determine the quality of their performance. The emotion recognition systems evaluate whether they sound pleasant enough or not. Based on this evaluation, they may receive a compensation bonus.

AI voice analytics technologies are also considered to be used in finance with loan default prediction and in the medical field, with mental health screenings. However, researchers discourage AI emotion recognition systems implementation in high-stakes contexts. There’s still no consensus on whether AI emotion tracking can be done in a consistent, reliable, and accurate way. Also, it demonstrates inconsistent results across different demographic groups. Therefore, using these systems to make life-altering decisions for people may come with harmful consequences.

When applied in a high-stakes context, the limitations of emotion AI become truly concerning. Emotion AI agents are pretrained on data sets that can include bias. Also, facial analysis systems can be substantially inaccurate identifying emotions of darker-skinned individuals. Moreover, researchers have identified racial and gender biases and unequal performance across various demographic groups.

Other limitations

While the capabilities of AI emotion tracking seem fascinating, they remain quite limited. One of the biggest drawbacks of emotion recognition systems is their simplification of emotions in the dataset. Commonly, the dataset is constructed based on the beliefs of the creators and the actors, hired to portray emotions (which are often simplified and exaggerated). As a result, they’re neither very reliable nor accurate. Yet that doesn’t stop people from using it as an effective form of surveillance.

Another limiting factor of AI emotion monitoring is cultural bias. For example, some gestures in one culture may have a completely different meaning in another culture.

When mood becomes data

For AI to process human emotions, it has to turn the input, whether it’s face capture, voice, or posture, into data that can be further analyzed. But the main question persists: how is this sensitive data being stored and used later?

Emotional AI uses highly sensitive data, including images, videos, voice recordings, and even biometric data. Naturally, this is an area of concern, as once emotions are quantified, they can be exploited.

Where AI emotion monitoring is no longer welcome

In response to the growing risks of AI emotion tracking, the technology is increasingly restricted or even banned in various areas. AI emotion recognition systems have weak scientific grounding, frequent bias, and a high risk of misinterpretation, which can cause serious consequences, including undermining individual dignity, autonomy, and mental integrity.

As a result, the EU Artificial Intelligence Act introduced a list of prohibited AI practices, effective February 2, 2025. This means that emotion recognition systems are prohibited from use in the workplace and schools. It’s also forbidden to use these systems for profiling-based criminal risk prediction. The exceptions apply only in a handful of cases, including medical or safety reasons.

While the US is not rushing to introduce any AI-related bans, as of today, four states, including Illinois, New York, Nevada, and Utah, have restricted AI use in therapy. Washington state lawmakers are also working on introducing stricter regulations on AI. A newly proposed House bill, 2225, identifies emotional recognition algorithms as raising new concerns about psychological safety, transparency, and accountability.

Conclusion

Emotion-tracking systems are already shaping our experiences in various sectors. From airports to offices, we encounter some form of emotional AI that decodes our inner feelings.

However, emotions don’t follow fixed rules. In some instances, we humans can’t tell what emotion a person in front of us is feeling. This only reveals the flaws in the emotion recognition systems. They are pretrained on biased and oversimplified datasets that increase the risk of incorrect interpretations, which can have negative consequences for people’s lives and well-being if not applied carefully.