The language bottleneck: How linguistic constraints shape AI and its applications

Artificial intelligence (AI) has advanced rapidly, particularly in the development of large language models (LLMs) that can generate human-like text. While these models excel at producing coherent and complex sentences, there may be deeper limitations at play—specifically, the language these models use to “think”.

Blog by Dennis Hulsebos

Language as a Framework for Thought

To understand these limitations, consider the Sapir-Whorf hypothesis, a theory suggesting that language shapes thought. This hypothesis posits that the structure and vocabulary of a language influence how its speakers perceive the world. For example, languages with multiple words for different shades of colour might lead speakers to notice those shades more distinctly. When applied to AI, particularly LLMs, this suggests that the language models are trained on might limit their ability to reason and conceptualise. An illustrative example is how LLMs, despite their fluency in human language, often struggle with simple arithmetic. This suggests that language alone isn’t always enough for precise or detailed reasoning, requiring AI to use more specialised tools like a code interpreter.

Jargon: The Language of Specialised Thinking

Professionals often use jargon—specialised language specific to their field—to communicate complex ideas efficiently. But jargon does more than just facilitate communication; it also shapes how experts think about their work. For instance, medical jargon allows healthcare professionals to describe and understand complex medical conditions quickly and accurately. This insight has practical implications for AI. If AI models are trained heavily on jargon-rich data, these models could become more adept at specialised reasoning within specific fields. The use of jargon doesn’t just help experts convey ideas; it also refines their cognitive approach, enabling more nuanced and detailed thinking. Similarly, an AI trained extensively in jargon could develop a more specialised form of reasoning, making it a more effective tool in fields like law, medicine, or engineering.

A Thought Experiment: Hieroglyphics and AI

Consider a thought experiment: What if we trained an AI model on hieroglyphics, an ancient symbolic language? Hieroglyphics convey ideas visually and symbolically, rather than through abstract words. Training a model on such a system could enable it to better interpret visual information, potentially bridging the gap between text and images. This experiment highlights the possibility that alternative languages or symbolic systems could enhance AI’s capabilities, especially in fields where interpreting complex visual data is crucial.

Implications for Professionals

These insights offer a new perspective on how professionals might interact with AI. Recognising that the language we use—and the language AI is trained on—can shape the way both humans and machines think. Professionals can tailor their use of AI tools to their specific needs. For instance, experts in highly specialised fields might seek out or advocate for AI systems trained on the specific jargon of their profession—leading to more precise and relevant outputs. Additionally, exploring AI applications that integrate visual or symbolic languages could be particularly valuable in areas like design, architecture, or data visualization, where these forms of communication are critical. By understanding the intricate relationship between language and thought, professionals can better leverage AI in their day-to-day work, pushing the boundaries of what these technologies can achieve.

Conclusion

While AI models like LLMs have made significant advancements, their abilities are shaped—and potentially limited—by the language they process. By understanding and leveraging this, professionals can enhance their interaction with AI by using jargon, and by encouraging the development of systems that not only communicate effectively but also think in ways that are more aligned with the specialised demands of various fields. Whether through the use of jargon, symbolic languages, or more innovative AI tools, the future of professional practice lies in embracing the complex interplay between language and technology.