How Neuroscience is Making us Rethink the Role of Language in AI and Localization

Large Language Model (LLM) research and recent neuroscientific discoveries are prompting us to reconsider the essence of language and its relationship with thought. This exploration is not just a philosophical pondering but a journey that could redefine our understanding of the human brain itself.

Traditionally, language has been seen as a fundamental pillar of human cognition, intertwined with our capacity for complex thought. However, recent explorations into the neural underpinnings of AI language processing, exemplified by the work of cognitive neuroscientists like Anna Ivanova at MIT, challenge this view. She is using her training in cognitive science to examine language models to see in how far they behave in human-like ways? Do their internals resemble the internals of human brains?

In a recent podcast she explained: "We can ask humans to do a bunch of different tasks, answer questions, see how well they respond, how accurate, whether the responses are consistent with each other. We can do the same thing with models. So, if we want to measure a particular ability in a model, if we come up with a way to ask our questions in a way that the model understands, we can then evaluate their behavior. And these models are generative text models, so they're very good at generating sentences. GPT models generate them word by word.

We can, for example, see how likely a particular sentence is in response to a general prompt. Or, for example, you can look at sentences like 'Yesterday I went to the beach' versus 'Yesterday I go to the beach'. One of them is grammatical, one of them is ungrammatical. We know that 'yesterday' should be accompanied by a verb in the past tense, so 'went' and not 'go'. And so we can just measure how likely a model is to generate one sentence versus another. We can systematically see whether these models prefer grammatical outputs over ungrammatical outputs."

One of the fundamental questions she is looking to answer is "Can we think without language?"

Disentangling Language from Cognition

In some instances, strokes can affect areas of the brain responsible for language, causing a condition known as aphasia. This condition impairs the ability to understand and produce language. While people with aphasia may have difficulty speaking and conveying their ideas, their thinking abilities are usually unaffected.

Studies show that the brain's language processing areas are distinct from those involved in other cognitive functions like planning, memory, social reasoning, and empathy. This further suggests that our thinking is not solely dependent on language.

Patients with global aphasia often retain the ability to perform mathematical calculations, understand social intentions, and engage in complex reasoning. Some continue to participate in creative and strategic activities, indicating that language is not necessary for these thought processes.

Some readers may be taken aback by these findings, especially those who feel their thinking is deeply intertwined with language. If you're among this group, here's another intriguing discovery recently made: not everyone experiences thought as inner speech. This suggests that the presence of inner monologue is not a prerequisite for the ability to think.

However, the relationship between language and thought is not entirely separate. While language loss in adults doesn't completely hinder other cognitive functions, the development of these functions may be influenced by language, particularly in childhood. For example, language can be crucial for learning numbers, as seen in cultures with limited numerical vocabulary. Furthermore, delayed language exposure in deaf children can affect their social and cognitive development.

Nevertheless, through the lens of patients with global aphasia, who retain cognitive functions despite severe language impairments, we're compelled to question the degree to which our thinking is reliant on language.

Scientists now believe that AI models, like humans, can differentiate between language and thought. This involves distinguishing between formal language skills, practical language use, and how these skills are somewhat connected to achieving artificial general intelligence (AGI).

Language: A Compression Algorithm for Thought?

The metaphor of language as a "compression algorithm" for transferring thought encapsulates the crux of ongoing debates in the world of AI. This perspective not only reframes our understanding of language's role but also sheds light on the limitations of current LLMs.

These models, designed to mimic human language capabilities, confront the challenge of interpreting and generating language that carries the weight of subjective interpretation and cultural context. Thus, they inadvertently contribute to the discourse on whether thought and language are as synonymous as previously assumed.

Neuroscientific research supports the notion that the brain's language processing centers are distinct from those involved in other cognitive functions. This distinction reveals that much of our cognitive experience—planning, memory, social reasoning, empathy, and moral decision-making—operates independently of language. Such insights not only illuminate the resilience of human cognition in the absence of language but also raise intriguing questions about the evolution and function of language itself.

Based on our knowledge of the human brain's specialized cognitive processes, we now also think that a language model that operates in a human-like manner would likely need to be modular. However that modularity might not have to be pre-built, it could emerge or be induced in an end-to-end system, possibly through multiple objective functions, curated data, or specific architectural decisions

AI and the Origin of Language: Redefining Thought Beyond Verbal Language

As LLMs evolve, their development trajectory offers a parallel inquiry into the origins and nature of language. By achieving linguistic capabilities, albeit through different means than human intelligence, these models highlight the multifaceted role of language in cognitive processes. The question then arises: Can true language abilities exist without the rich tapestry of non-linguistic cognitive components?

The debate extends beyond the binary of language presence or absence. Not everyone experiences inner speech, and non-verbal languages—such as pictorial representations and body language—underscore the diversity of thought processes. This leads to a broader understanding of cognition as a domain that transcends verbal language, encompassing a spectrum of sensory, emotional, and conceptual experiences.

Implications for AI and Localization

For professionals in AI and localization, these insights challenge us to rethink how we approach the development of language technologies. They invite us to consider the nuanced ways in which language shapes and is shaped by cognition, and how AI can bridge the gap between linguistic expression and the underlying complexity of thought.

As we explore the frontiers of AI's capabilities in language and cognition, we're reminded of the rich interplay between these domains. The journey of understanding the essence of language and thought continues to evolve, offering fertile ground for innovation and philosophical inquiry in the field of AI and localization.

How can we foster a deeper appreciation for the intricate relationship between language, thought, and technology, you think?

As we advance in our exploration of AI's role in language processing, let us remain mindful of the profound questions at the heart of this journey:

What does it mean to think, and how does language enrich this fundamental human experience?