What’s behind AI text understanding?

Since the release of “Her” (2013), a film directed by Spike Jonze, the idea of a humanized AI assistant capable of interacting naturally through voice has been in many people’s imagination. In the film, a man falls in love with an AI named Samantha, who, despite sounding incredibly real, will never truly be human.

Twelve years later, this is no longer just science fiction. Generative AI tools like ChatGPT and virtual assistants such as Apple’s Siri and Amazon’s Alexa help people navigate traffic, create shopping lists, and complete tasks. However, beyond text processing, automatic speech recognition systems—just like Samantha—are still far from achieving the level of human cognition.

What are these limitations?

According to the study See What I’m Saying? Comparing Intelligent Personal Assistant Use for Native and Non-native Language Speakers (2020), written by linguists and computer scientists, the response of AI varies depending on the user. Research shows that, for English speakers, error rates increase if the person has a foreign accent, is Black, speaks in African American Vernacular English (AAVE), switches between languages (code-switching), is a woman, elderly, very young, or has a speech impairment.

Why do AIs struggle to understand certain people?

Unlike humans, speech recognition systems are not empathetic listeners. While we interpret words by considering intonation, facial expressions, and context, making an effort to understand, AI relies on probabilistic guesses—and often gets it wrong.

What’s behind? AI Training Data

But what behind these errors? It lies in the data used to train AI models. To learn how to understand and replicate human speech, AI systems are trained on vast amounts of text and audio data. The problem is: Whose voices are being used?

If an AI system achieves high accuracy when interacting with white, middle-class Americans men in their 30s, it’s likely that most of its training data comes from them. This means that people from different backgrounds, ages, and social classes may experience more difficulty using these technologies.

To improve AI accuracy, training data must be more diverse in terms of gender, age, race, languages, and accents. Also, systems should be designed to recognize uncertainty and ask for clarification, just as a human would.

Linguistic in AI

For non-native English speakers—that is, most of the world’s population—the challenges are even greater. The most advanced AI models have been primarily developed in English, performing better in English than in other languages. While AI has the potential to improve translation capabilities and expand access to information, most languages suffer from limited data, making them harder to integrate into large language models (LLMs).

Even within widely spoken languages like English and Spanish, AI can make mistakes depending on the dialect. This happens because these systems consider the biases embedded in their training data.

The Value of Human Connection

AI is expected to gradually improve its ability to understand different accents, linguistic variations, and language switching.

Miscommunication happen even in human-to-human interactions. However, when talking to another person, there is always a chance of bumping into an empathetic listener, capable of get context, emotions, and nuances—something AI still struggles to replicate.

QA Booster: Natural Language in Software Testing

The challenges of AI speech comprehension go beyond virtual assistants and impact other critical areas of technology, such as software development and automated testing.

At NextAge, we know the importance of natural language in making AI-driven interactions more intuitive and accessible. We know this technology can help in everyday tasks, including QA processes in software development.

That’s why we created QA Booster, the new NextAge vertical dedicated to AI-powered test automation. What makes this solution unique is its use of natural language, enabling tests to be more intuitive, comprehensive, and adaptable to different communication styles.

Beyond improving user experience, QA Booster makes the QA process faster, more efficient, and accurate, reducing testing time while delivering reliable results displayed in an intuitive dashboard.

With advancements like these, AI can not only overcome its limitations but also become a powerful ally in innovation and automation, ensuring smarter and more inclusive systems.