For all the breathless coverage of artificial intelligence, the technology remains conspicuously bad at tasks a toddler performs without thinking. It cannot reliably count the windows in a photograph of a house. It struggles to follow a recipe that requires adjusting for what is already in the pan. It will confidently explain quantum mechanics while failing to notice that its arithmetic is wrong. These are not bugs awaiting patches; they are structural features of how current AI systems work, and understanding them matters more than memorising benchmark scores.
The gap between perception and reality has rarely been wider. Surveys consistently show that substantial portions of the public believe AI can already perform general reasoning, maintain genuine understanding of context across long conversations, and learn from individual interactions the way humans do. None of these things are true in the way people imagine them to be.
The reasoning mirage
Large language models predict the next plausible token in a sequence. They are extraordinarily good at this, good enough to produce text that reads like reasoning. But pattern-matching at scale is not the same as logical inference. When a model solves a maths problem, it is often recognising the shape of similar problems in its training data rather than deriving the answer from first principles. Change the numbers slightly or phrase the question unusually, and performance can collapse in ways that genuine understanding would not permit.
This explains the persistent fragility. Models that score impressively on standardised tests can fail spectacularly on slight variations. They lack what cognitive scientists call compositional generalisation—the ability to recombine learned concepts in novel configurations. A child who learns what "red" means and what "ball" means can immediately understand "red ball" in any context. Current AI systems approximate this ability without truly possessing it.
The embodiment problem
Much of human intelligence is grounded in physical experience. We understand "heavy" because we have lifted things. We grasp causality because we have pushed objects and watched them fall. Language models learn about the world entirely through text, which is a description of reality rather than reality itself. This creates systematic blind spots.
Ask a model to plan a simple physical task—packing a suitcase efficiently, rearranging furniture in a room—and it will produce plausible-sounding instructions that often fail on contact with the actual world. The model has no internal simulation of three-dimensional space, no intuitive physics, no sense of what fits inside what. Robotics researchers have spent decades on these problems and made real progress, but the integration of language understanding with physical competence remains genuinely hard.
Memory and continuity
Current systems do not learn from their conversations with you. Each interaction begins fresh, with the model having no recollection of previous exchanges beyond what is explicitly included in the context window. This is not a privacy feature; it is an architectural constraint. The model that helped you draft a document last week has no memory of doing so.
Various workarounds exist—retrieval systems that store and fetch relevant prior context, fine-tuning on specific data—but none replicate the continuous, automatic learning that characterises human cognition. You do not need to be reminded each morning who your colleagues are or what project you are working on. AI systems effectively do.
Our take
None of this diminishes what AI can accomplish. The technology is transforming industries, accelerating research, and enabling capabilities that seemed distant a decade ago. But the hype cycle serves no one well. Overpromising leads to misallocation of resources, inappropriate deployment in high-stakes contexts, and eventual backlash when reality intrudes. The most useful framing treats current AI as a powerful tool with specific, knowable limitations—not a nascent general intelligence temporarily constrained by hardware. Knowing what machines cannot do is the beginning of using them wisely.
