AI-Generated Images Won't Replace Human Genius

A new paper from Google DeepMind highlights the limitations of Large Language Models. The research shows that AI software can’t replicate human creative genius, no matter how sophisticated it is.

The authors point out that LLMs lack a physical body, which introduces a cognitive barrier and restricts their potential for innovation. In other words, they can only analyze existing information and can’t generate new ideas on their own.

LLMs are capable of mastering one distinct type of inference: induction. But even this is limited because they rely on statistical pattern recognition and massive data compression. They’ve implemented deduction using derived logical proofs from established rules, but abduction remains an impenetrable ceiling for them.

The term ‘abduction’ refers to the ability to generate explanatory hypotheses when faced with scarce data. Human scientists like Albert Einstein were able to make groundbreaking discoveries despite limited visual data. His theory of General Relativity is a prime example of this phenomenon, where he used physical intuition to connect sensory experiences and generate new mathematical principles.

Einstein’s breakthroughs relied on ‘embodied thought experiments,’ where he used his understanding of the world to invent original premises through physical experience. Unlike LLMs, which can only analyze existing text, Einstein was able to create something entirely new. His work demonstrates that human scientists have a unique advantage over AI systems when it comes to scientific discovery.

The paper also notes that current AI discovery frameworks are limited by their reliance on rulebooks and thresholds rather than creating entirely new frameworks. This is a significant limitation for developing truly innovative solutions, as LLMs can only build upon existing knowledge.

This lack of physical grounding makes it impossible for LLMs to create intuitive models based on cause and effect. The authors suggest that bridging logical calculation and true scientific invention will require future AI architectures to evolve beyond text and image processing. By giving AI the ability to run embodied simulations in virtual environments, they may be able to develop a deeper understanding of the world and make more significant contributions to scientific discovery.