Is AI Really ‘Thinking’? New Research Suggests It’s More Like Cognitive Fossilization

The race to achieve human-level artificial intelligence (AI) is intensifying, with industry giants like OpenAI and Google confidently predicting its imminent arrival. However, a growing chorus of skeptics argues that current AI models don't truly think like humans. Recent research digs deeper into this debate, suggesting that AI's apparent intelligence may be fundamentally limited by its underlying architecture.

At the heart of the matter lies the question of how AI actually processes information. One emerging perspective views today's AIs as adept at simulating intelligence by learning vast sets of "rules of thumb" and selectively applying them. But is this genuine reasoning, or simply a sophisticated form of pattern recognition?

One fascinating analogy compares large language models (LLMs) to "fossilized cognition." Instead of being dynamic minds, LLMs are seen as exquisite reconstructions of something long buried: archives in activation. Their knowledge doesn’t evolve; it’s drawn from vast, static layers of text – a sedimentary record of human expression, compressed and cross-referenced by pattern rather than meaning.

This perspective highlights a crucial difference: LLMs don’t live in the world, don’t think across time, and don’t feel consequences. Each prompt is a self-contained "now," lacking a sense of past or future. While some newer models feature extended memory, these are merely technical workarounds, retrieving information rather than truly remembering and reflecting.

The brilliance of LLMs is undeniable. They write, compose, solve problems, and converse with remarkable fluency. However, this fluency can be deceptive. LLMs dazzle not because they are becoming like us, but because they reflect us so well we forget they are mirrors. We see intelligence – but overlook the scaffolding behind it.

Another challenge for AI development stems from the excessive computational cost of deep reasoning. Companies are realizing that having AI models engage in "too much reasoning" is financially unsustainable and environmentally wasteful. For example, simply saying "please" and "thank you" to ChatGPT incurs surprisingly high costs. This has led to the development of tools like Google's DeepMind's reasoning dial which allows developers to control how much effort the AI puts into particular prompts.

The core issue lies in the fact that AI models tend to "overthink" simple queries, consuming far more resources than necessary. This overthinking appears to be the rule rather than the exception, underscoring the need for more efficient AI architectures.

Ultimately, understanding the true nature of AI requires us to move beyond the surface level. Are we evolving machines towards minds, or evolving ourselves towards machines? AI holds immense potential, but it's crucial not to mistake fluency for depth or projection for presence. Let’s not endow them with traits they do not possess, but appreciate them for what they are: responsive systems, animated by input, not intention.

What do you think? Is AI truly on the verge of achieving human-level intelligence, or are we witnessing something fundamentally different? Share your thoughts in the comments below!

• Google
• LLMs
• OpenAI