Rethinking AI's Human-Like Traits: Reality Check for LLMs

The debate around large language models (LLMs) often spirals into questions about their alleged human-like traits. Are these models capable of understanding language or possessing morality? The reality is, researchers suggest, that this line of thinking might be misleading.

Unpacking the Myth

The study dives into a fascinating area where LLMs are compared to anything from LEGO sets to urban landscapes like the Greater Boston Area. The argument is simple yet profound: any system, when analyzed deeply enough, could exhibit similar perceived human-like traits. Strip away the marketing and you get a clearer picture.

The researchers trained a basic neural network on the strategy video game Age of Empires II and found that the anthropomorphic traits ascribed to LLMs aren't unique. This suggests that what we often interpret as 'understanding' or 'intent' might just be our own projections onto these systems.

The Substrate Debate

Here's what the benchmarks actually show: the substrate plays a essential role in determining the attributes we see in LLMs. The substrate, whether it's a digital game or physical LEGO bricks, can shape the behaviors we interpret. So, are we seeing genuine AI understanding, or is it just a mirage caused by the substrate?

Consider this: if two systems on different substrates can display similar traits, does it make sense to label these traits as inherently human-like? The numbers tell a different story. An empirical discussion about LLMs demands clear-cut measurement criteria, not vague human comparisons.

A New Approach

The researchers propose a 'null' assumption, advocating for experiments set up without assuming LLMs have human-like traits. It's a radical shift from the norm, challenging us to rethink how we measure AI capabilities.

Why should this matter? Because assuming LLMs possess human-like attributes without evidence leads to circular conclusions. If we want meaningful insights into AI, we need to ground our assumptions in the system's substrate and behavior.

So, next time someone claims an AI model 'understands' you, ask this: Is it really understanding, or just a sophisticated mimicry shaped by its environment?