Bridging the AI-human language gap: Study proposes neologisms for better understanding

At the heart of the study is the assertion that humans and AI conceptualize the world differently. Humans rely on a shared language with words that encapsulate familiar experiences, while AI operates based on abstract patterns and data structures that are fundamentally different. This gap creates a communication problem: humans struggle to understand how AI reaches its conclusions, and AI cannot precisely grasp the nuances of human thought.

Artificial intelligence (AI) has grown to perform complex tasks, yet our ability to understand and control it remains limited. As AI systems become increasingly sophisticated, their decision-making processes and reasoning mechanisms remain largely opaque. Traditional methods of AI interpretability attempt to map human concepts onto machine behavior, but they often fall short in explaining how AI truly “thinks.” This lack of clarity raises concerns about AI safety, trust, and reliability in applications ranging from content moderation to autonomous systems.

A recent study, "We Can’t Understand AI Using Our Existing Vocabulary," by John Hewitt, Robert Geirhos, and Been Kim from Google DeepMind, argues that existing human vocabulary is insufficient to fully comprehend AI’s internal processes. The paper, currently under review, proposes a new approach: developing neologisms - new words that define human concepts in ways that machines can process and, conversely, introduce machine concepts into human understanding.

The communication barrier between humans and AI

At the heart of the study is the assertion that humans and AI conceptualize the world differently. Humans rely on a shared language with words that encapsulate familiar experiences, while AI operates based on abstract patterns and data structures that are fundamentally different. This gap creates a communication problem: humans struggle to understand how AI reaches its conclusions, and AI cannot precisely grasp the nuances of human thought.

The researchers frame AI interpretability as a communication challenge, emphasizing the need to create new words - neologisms - that bridge this conceptual gap. A machine’s understanding of “sentiment” in text, for instance, may differ drastically from human perception, leading to misunderstandings in applications such as content moderation or sentiment analysis. Without a shared vocabulary, efforts to control or predict AI behavior become increasingly difficult as models grow more advanced.

Neologisms: A new approach to AI interpretability

The study suggests that creating new vocabulary elements for AI-related concepts could provide better clarity and control. Unlike traditional interpretability methods that rely on human-labeled data or behavioral testing, neologisms would offer a systematic way to establish shared meanings between humans and machines. For example, introducing a word specifically designed to denote AI’s perception of response diversity could allow users to precisely adjust variability in generated text.

A proof-of-concept experiment included in the study showcases how defining a neologism for "length" enabled improved control over AI-generated response length. Similarly, another experiment demonstrated how defining a “diversity” neologism allowed models to generate more varied responses when prompted. These results suggest that neologisms can be used as control mechanisms to guide AI outputs in more predictable and human-aligned ways.

The implications of AI-specific vocabulary

Developing a shared human-AI vocabulary could have significant implications across multiple domains. In areas such as AI safety, governance, and ethical decision-making, being able to accurately specify and interpret AI behavior would reduce risks associated with opaque decision-making processes. Additionally, machine-learning researchers could use neologisms to create more transparent models that are easier to debug and refine.

Moreover, neologisms could help address biases and anthropomorphic interpretations of AI behavior. By defining concepts in a way that acknowledges the fundamental differences between human and machine cognition, researchers can prevent overgeneralized assumptions about AI’s capabilities and limitations. This approach aligns with broader efforts to make AI more explainable and controllable without relying solely on human-centric analogies.

Future directions: Building a human-AI language

The research highlights the need for continued exploration into the development of neologisms for AI interpretability. Future work could focus on systematically identifying key areas where new terms are needed and testing their effectiveness in real-world applications. Additionally, researchers might explore methods to standardize these new words, ensuring they are broadly understood across different AI systems and research communities.

By fostering a structured approach to communication between humans and AI, neologisms could pave the way for more effective collaboration between people and intelligent systems. As AI continues to evolve, expanding our vocabulary to encompass its unique modes of reasoning may be essential for unlocking its full potential while maintaining transparency and control.