07.02.2024
19 min
1311
A viral trend in which people repeat the same phrase using different tones of voice could unexpectedly benefit major technology companies. Experts believe such videos may help train artificial intelligence systems to better recognize and understand human emotions.
A trend in which users repeat the same phrase using different tones of voice has quickly gained popularity across social media platforms. The exact same words can sound friendly, sarcastic, irritated, or even aggressive depending on the emotional delivery.
Experts believe these videos could help train emotion recognition systems. For artificial intelligence models, they offer an opportunity to better understand the difference between what is being said and how it is being said. At the same time, researchers note that this type of data has limitations. Due to the humorous nature of the trend and the exaggerated emotions often displayed, such videos do not always reflect real human behavior, making their value for AI training somewhat limited.
Participants in the trend repeat the same short phrase while expressing different emotions, ranging from support and excitement to sarcasm, frustration, or anger. This contrast between words and delivery is what made the format popular and attracted millions of viewers.
However, according to digital ethics expert and North Star Strategies CEO Clara Fulks, there may be more to the trend than simple entertainment.
“That is going viral because big tech companies need emotional training data to train their emotion recognition models.”
According to Fulks, this type of content can be seen as a win for companies developing or selling emotion-analysis technologies.
Olga Kokhan, founder and CEO of data services company Tinkogroup, says the trend highlights one of the biggest challenges in emotion recognition. The same words can carry completely different meanings depending on tone of voice, speaking pace, accent, and context.
“Thousands of users saying the same phrase with different emotional expressions create a dataset that teaches AI systems to distinguish between what is being said and how it is being said.”
Such data could be used to improve speech analysis systems, sentiment detection, customer support tools, and accessibility technologies. For example, some AI-powered call center solutions already attempt to assess a caller’s emotional state and suggest appropriate responses to agents. However, Kokhan and other experts emphasize that data collected from viral trends has significant limitations. Social media content is often created for humor, engagement, or virality and does not always reflect real-life situations.
John Licato, PhD, Associate Professor at the Bellini College of Artificial Intelligence, Cybersecurity and Computing, notes that participants in viral trends frequently exaggerate their emotions on purpose.
“This does not necessarily provide the most useful training data.”
Licato also points out that modern large language models are already fairly good at identifying broad emotional categories, but they still fall short of human-level understanding when it comes to complex emotional states.
“When we start using them to detect emotions at a more detailed level, they perform much worse. The nuances of human emotions are extremely difficult to capture correctly because they can vary significantly from person to person.”
Kokhan adds that emotions are heavily influenced by culture, personal experiences, context, and circumstances. As a result, even advanced AI systems can misinterpret sarcasm, mixed feelings, or subtle emotional cues.
Despite the technology’s limitations, experts already see potential risks. A 2025 report from the UK’s Institute for the Future of Work warns that emotional AI shifts surveillance away from what people do and toward how they feel.
For example, employers could use such systems to monitor employee engagement, mood, or loyalty.
“Over time, this kind of surveillance may pressure people to display emotions that machines expect to see, smiling when they do not feel like it, hiding frustration, or adjusting their facial expressions to avoid triggering alerts.”
The trend of repeating the same phrase is far from the first example of user-generated content being used to improve AI models. In late 2025, the “Hug Your Younger Self” trend became popular, with users generating AI images of themselves standing next to childhood versions of themselves. At the time, experts warned that participants were effectively handing over biometric data that could be used to train facial recognition systems or create deepfakes.
Similar concerns were raised about the “Back to 2016” trend, in which users shared decade-old photos of themselves. Jonathan Drake Steele, founder of cybersecurity consultancy Steele Fortress LLC, described such trends as a “goldmine of data” for AI developers. According to Steele, facial recognition systems require images of the same person taken over long periods of time. Collecting such datasets is usually expensive, but social media users often provide them for free when participating in viral trends.
