What brains do that AI still does not: neuroscientists' list
David Eagleman, Dileep George, Terry Sejnowski, Fei-Fei Li, Ted Gibson and Michael I. Jordan describe abilities such as rewiring from experience, feedback-driven perception, grounding in the physical world and thought without language that current models lack. The list shows where language models are a partial picture of intelligence and where research may go next.
Ask people who study brains, vision, language and learning what a brain does that current AI does not, and they give a consistent list. A brain rewires itself as it learns, cares about outcomes, perceives by projecting a model onto what it senses, grounds concepts in a body's experience, thinks without language, learns from a handful of examples, keeps two kinds of memory, and keeps generating thoughts when nothing is coming in. Language models have mastered the form of language and the statistics of a huge record of human output. That is a real achievement and a partial one.
David Eagleman, a neuroscientist, says the brain is livewired, constantly reconfiguring itself physically. A young child can lose a whole hemisphere and recover. He thinks the bigger gap is drive. Humans encode what matters to them, grounded in survival. A language model, in his words, is "a good impersonator". Fei-Fei Li, who builds AI systems, agrees that a chatbot saying it is sorry you are sick has never known pain.
Dileep George, who builds brain-inspired AI, says the visual cortex has more feedback connections than feed-forward ones, because perception is a search for the best explanation of the input. Concepts rest on sensory and motor experience, and most of what we know about how the world works is never written down. A language model carries a model of the text world. Li makes the same point about video and says the next frontier is beyond language.
Ted Gibson, a psycholinguist, brings evidence that language and thought are separate. The brain's language network stays silent during mathematics, music and spatial tasks, and patients who lose it can still play chess and do arithmetic. He concludes "language is a communication system for the meaning. It's not the meaning". Models do form very well and can still miss meaning.
Li notes that a child learns a cat from perhaps three to ten examples. Charan Ranganath, a memory researcher, says humans pair a fast memory for single events with slow accumulated knowledge, so they can change without erasing what they know. Terry Sejnowski, a neuroscientist, says a brain keeps thinking in a quiet room, while a model goes blank when you stop talking to it. Michael I. Jordan, a statistician, adds that prediction is not decision, and that current systems give no measure of their own uncertainty.
They differ on the size of the gap. Sejnowski thinks models generalize and build internal models of meaning. George says the missing information is not in text at all. Jordan doubts we understand intelligence well enough to say. Some items are already partly present in machines, and machines beat people where patterns are plentiful. Doctors working with an AI outperformed either alone.
What to do with it: read fluent output as form, then check the meaning by changing a familiar problem slightly. Ask what the system could have seen, since unwritten knowledge and rare cases are its weak spots. Do not act on a bare number, and take medical predictions to a doctor. Use a model for breadth while you supply depth. Learn how the technology works. Seek novelty, wrestle with problems yourself, and leave quiet time for your own thoughts.