Last Updated on 29 August 2026

Like every major technological development that becomes an industry and starts reshaping humanity, Artificial Intelligence has attracted extraordinary optimism and capital. Hundreds of billions of dollars now flow each year into models, chips, data centers and companies built around them. As bubbles often do, AI really took off when it stopped being only a technological revolution and began exciting the masses. Since ChatGPT launched in November 2022, Large Language Models (LLMs) have brought AI into almost every home. With mass, prolonged excitement comes mysticism: predictions, rumors and beliefs repeated until they begin to look like facts, and this is the fuel of bubbles. AI itself is not new, as its neural-network roots reach back to Frank Rosenblatt’s Perceptron in the 1950s. Today’s LLMs are all still built from the same neurons, weights, biases and activation functions. Their job is to learn patterns from a large, messy and incomplete slice of human knowledge while generalizing to new inputs. Transformers, introduced in 2017, enabled the current LLM wave by giving models a better way to capture relationships across context, igniting much of the excitement we see today. These days, billions of prompts are sent daily, agents are everywhere and trillions of tokens are used. The possibilities seem endless, accompanied by doomsday predictions that mirror the scale of the imagined upside. As always, the future is unknown and its interpretation is a clash between optimists and pessimists. Will AI take our jobs? Will it destroy humanity? And perhaps the most consequential question: will it achieve recursive self-improvement by building AI models that improve themselves indefinitely?

As of today, no one can answer these questions with certainty, but we can examine AI’s specific weaknesses and how difficult they are to solve. AI models are basically pattern-recognition machines: data enters a trained model, which maps the input through patterns and relationships learned during training, producing an output that can look remarkably like meaning. LLMs are extraordinarily good at processing text and producing relevant-looking responses by repeatedly predicting what token should come next, but they depend heavily on context. They often hallucinate, confuse levels of abstraction, compound errors, misclassify things and sound convincing while being wrong. Context, the information available to the model when it produces an answer, is at the core of the problem. The more precise and relevant it is, the fewer plausible interpretations remain. Generating high-quality context, however, is incredibly difficult for machines and humans alike. Exacerbating the problem is the fact that small changes can materially alter the output, making LLMs highly context-sensitive and often brittle. Who has not received a decisive answer from an LLM, pushed back slightly and watched it completely change its “mind”? This is why there is always some hesitation when reading an LLM’s output: however convincing it sounds, we can never be certain that it is true.

The problem becomes more serious when AI-generated output is fed back into the model as context. A complex response is rarely perfectly accurate, complete or aligned with reality. Once that imperfect output becomes part of the next prompt, the model is no longer working from the original situation alone, but from a partially distorted version of it. Repeated without human correction or external verification, small inaccuracies can quickly compound while the conversation remains fluent and convincing. A model can infer which details appear relevant within the information it has, but cannot reliably determine from that information alone whether the frame itself is correct, whether something important is missing or whether a plausible premise is actually true. It has no independent grip on reality inside the context window. Modern AI can search the web, query databases, run code and use other tools, but the same basic problem remains: the system must know what to look for, which source to trust and when its assumptions need to be challenged. Humans still do much of the work of keeping the context grounded in truth. Where the stakes are high, LLMs need experienced supervision to distinguish real from imaginary, relevant from irrelevant and useful from merely fluent. This problem significantly hampers their ability to recursively self-improve without a human holding their “hand”.

AI is also expensive to provide. It depends on chips, data centers, electricity, cooling, cloud infrastructure, data pipelines, research, engineering and constant maintenance. This makes the question practical as well as technical. It may often be cheaper to keep humans in the loop to focus the machines, choose the right context, catch distortions and stop when the answer is good enough. Human supervision is not a sentimental attachment to old ways. It is a significant context management layer. Humans decide what the machine should know, what it should ignore, when it is hallucinating, when it is technically right but practically useless and when the work is finished. Without that boundary, automation can become an expensive way to manufacture confidence without truth. The more complex the subject, the harder the output is to verify, the more valuable an experienced professional becomes.

This is why AI is unlikely to outright replace most jobs in one clean sweep. It will replace tasks: drafting, summarizing, boilerplate coding, searching, formatting, translating, classifying and producing first versions of work. Jobs, however, bundle tasks with judgment, context, responsibility and quality output. LLMs enhance the work of capable humans and remove technical boundaries. Capable humans are the ones defining the problem, understanding the stakes, choosing the relevant context, checking whether the answer is true and taking responsibility for the result, and this will remain a major limitation as long as machines cannot reliably build, challenge and ground their own context. AI will therefore change the center of gravity in many professions. The most valuable workers will be those who are already capable and experienced, who can build quality context, frame problems, use AI to accelerate execution and apply judgment where the model is weakest. The machine can help with the work. The worker still has to grasp what the work is, and as always, quality output starts with a sharp human mind.