There are three key similarities in how AI, especially LLMs, and humans perceive information. These are three foundational aspects of perception: memory, learning, and conceptual understanding. People often say that AI is unable to truly understand things, but if you examine how AI and humans filter information, they are more similar than expected.
Memory
Memory refers to the amount of storage of and ability to retrieve information. In AI, memory is whatever is present on the hardware, which could come from GPU memory, CPU memory, or in-formation retrieved online through API calls or URL request fetching. A common problem in chatbots is that the ability to process memory degrades over context length; if you give an LLM too much information, it won't be able to retrieve information from 1 million tokens ago. With humans, the mechanism is the same, with memories presumably stored somewhere in our brains; if you overload a human with information, we will forget the oldest pieces too.
There's also a distinction between short-term and long-term memory. Humans famously exhibit this distinction: sometimes, we can remember information encountered in a split-second for a few seconds, but not longer, while other memories, such as how to ride a bike, are stored forever. LLMs can also share this pattern. If the model does not have a mechanism for storing or saving context, then the model simply forgets it entirely, because it went the equivalent of in one ear and out the other. More recent LLMs, such as those with agentic capabilities, will save old context to special memory files so they can reference and understand old information without having to be prompted.
It is fashionable to draw a distinction between actual memory in the sense of recall/being able to remember something without prompting or looking at it, and fetching information that has been previously stored by looking at it again. For humans, there is a clear distinction between the two: if you memorize a deck of cards, you may be able to recall that the 37th card was the 6 of clubs without having to look at it. With AI, there is also such a distinction. If you give an LLM a certain pattern, image, or words, it will convert it (encode) into an embedding, which then gets passed through attention mechanisms. Asking it to recall the original words is a matter of restoring the original encoding. The LLM, if not equipped with agentic tools, is unable to "peek" back at the context region where the answer was specified; it can only generate words forward through time, one of the fundamental limits of transformers/autoregression. Instead, it gets the answer because it has been trained on these types of memory tasks. However, if that LLM saves the context to a memory file and calls a tool to read that file, it can literally just copy the answer because it saw it.
Learning
The next big concept is the process of learning. The reason "artificial intelligence" is named the way it is, and is based on what are called "neural networks", is because in the 1950s, scientists believed that the way machines detected objects or assessed conditions would be similar to how neurons in the brain fired. There would be similarities in the biological structure of the brain and how information propagated from neuron to neuron, just like how information propagated from earlier nodes to later nodes through activation functions and linear matrix combinations.
Although I enjoy calling AI "magic matrices" instead, because its mathematics does not involve terribly sophisticated functions, AI does exhibit fundamental "learning" traits that match those of the human brain. AI can perform supervised or unsupervised learning, finding its own patterns by minimizing a loss function. If exposed to enough samples of dogs vs cats, the AI can learn to distinguish between dogs and cats due to having learned representations through higher-level (later in the network) or lower-level (earlier) features which correspond to likelihood of a certain category. Isn't this exactly the same as humans though? How do we determine whether it's a cat or a dog? It's because we look at the ears, eyes, etc. and think whether this is more associated with dogs or cats. And we see that there is a strong correlation between AI misclassified labels and human-misclassified labels, bar pathological examples like adversarial image attacks, e.g. slightly twisting the pixels of a cat to make the AI call it an avocado while to a human it still looks like a cat.
Another form of learning, reinforcement learning, is how we can reward correct answers through rewards. For AI, this is by optimizing a rewards function, which oftentimes must take place dynamically or online (you don't know how well you've performed or the optimal path until the very end when you've finished the task; while you're playing the game, there's nothing you can know). This is how AlphaGo learned to play weiqi/baduk/go: a series of rewards being optimized. Similarly, humans learn the same way: they assess the benefits of a chess move by its consequences, e.g. did it produce a victory or a loss? This can also be generalized to human experiences: avoiding lions makes one less likely to be eaten. The fundamental structures that govern how humans and AI learn are identical.
Conceptual Understanding
The last fundamental similarity in learning is how AI and humans can both make higher-order judgments. Low-order judgments would be at the five-sense-level, like color and taste. This is the mode that Sensing people operate primarily on in the 16 Personalities. On the contrary, higher-order thought would be more abstract or semantic qualities, such as facial expressions, emotions, and demeanor.
People like to say that AI can't make higher-order qualitative judgments, but we often see AI being able to evaluate movies and emotions in depth. AI services are increasingly popular for counseling. The rebuttal is that AI can only achieve these judgments because a human trained it, but don't humans also work like so? After all, we were also trained by someone else to recognize certain emotions. When dealing with unfamiliar territories, we, like AI, draw on previous experiences to make a best guess for what we should do in the current situation. Our own ability to abstract think is no different than AI's: drawing from previous experience and using previous knowledge to inform the present.
Conclusion
So given how similar AI is to humans, and how our cognitive capabilities are much more similar than different, is there anything that can differentiate us? They say that the Turing Test is what separates AI from humans. Informally, if a human cannot distinguish between responses given by a human or an AI, then AI will have succeeded in being indistinguishable from humans, and can be marked as human. At the present moment, people can often distinguish between AI responses and human responses if you ask sufficiently many questions which reveal certain incriminating qualities (AI loves the em dash – and ellipsis …). But can't I argue that we simply need to train AI for longer to match with the rubric and benchmarks that humans are producing? AI can do anything humans can if you give it enough compute resources and time, because AI operates on the same thinking patterns as humans!