How AI Language Models Reflect the Human Brain's Left-Right Divide

New research shows that large language models like OLMo-2 7B develop left-right brain asymmetry in language processing as they train, mirroring the human brain's left-hemisphere dominance for language. The finding, based on fMRI comparisons, reveals that AI models increasingly predict left-hemisphere activity when distinguishing grammatical sentences, though this asymmetry does not extend to math or world knowledge tasks. The effect was consistent across English, French, and Chinese, suggesting a universal linguistic learning pattern.

How AI Language Models Reflect the Human Brain's Left-Right Divide AI models, like humans, show left-right brain asymmetry in language processing. This mirrors how these models evolve linguistic capabilities. The fascinating world of AI never fails to astonish. Here's another instance: large language models LLMs like OLMo-2 7B are starting to mirror the left-right divide in our brains when they process language. As these models train, their 'thoughts' align more with what's happening in our left hemisphere, the part of our brain traditionally linked to language. The Left-Brain Advantage Research using functional magnetic resonance imaging fMRI reveals that, as an AI model hones its skills, its internal workings increasingly predict brain activity, especially in the left hemisphere. It's like watching a digital brain grow a mind of its own. But why should we care about this left-right asymmetry? Because it gives us a peek into how AI develops formal linguistic skills, a key metric in gauging its progress. These skills are put to the test by how well the model discerns between grammatically correct and incorrect sentences. Think about it. If an AI can tell the difference, it's on its way to understanding language as we do. Our left hemisphere's dominance in language processing means these models might be learning in a way that mirrors human cognition. What’s Missing? Interestingly, this left-right asymmetry doesn't align with the models' performance on tasks outside pure linguistics, like math or world knowledge. So while the AI might ace grammar, it's not necessarily a math whiz or world knowledge expert. Is this a limitation or a stepping stone? Depends on whether you're looking for a conversational AI /glossary/conversational-ai or a digital jack-of-all-trades. Beyond English In a bid to challenge the model further, researchers tested it in French and Chinese. The results were consistent, showing that this linguistic asymmetry isn't just an English phenomenon. It seems that LLMs, regardless of the language, are wired to become better language users in a way that's eerily human. The real story here isn't just that AI models are mimicking the human brain. It's about what this tells us regarding how to push AI development further. Should we make models more like us, or should we focus on creating something entirely new? As AI continues its relentless march forward, these are the questions we should be asking. Get AI news in your inbox Daily digest of what matters in AI.