Why LLMs are not the AI that many hoped and some feared
Note: This essay was originally published on Medium. You can read it here without a paywall.
When it comes to the topic of artificial intelligence (AI), it can be very difficult to separate facts from hype — even hallucination (both human and machine).
In a recent essay about the alleged OpenAI chatbot “hack” of a server owned by another AI company, author Cory Doctorow poked holes in the claim that the bots had “gone rogue” to independently attack a rival.
“The Hugging Face hack isn’t a mysterious, supernatural occurrence,” he wrote. “It’s a Python loop and a chatbot. The people responsible didn’t accidentally create god: they created autonomous malicious software and then failed to closely monitor it, resulting in it doing something both foreseeable and bad.”
A useful way of looking at it, he posited, is that “LLMs [large language models] are real, but AI — which he then defined as ‘chatbots that wake up, “set their own goals,” and “spontaneously” start hacking servers’— is fake.”
READ
Medium: LLMs Are Real, AI is Fake by Cory Doctorow
Coming to terms
Is AI fake?
I wrote before about the problem of the public conflating the large frontier-model LLMs with all of artificial intelligence. But a bigger problem may be that we lack a common, universally accepted definition what AI is.
Merriam Webster defines artificial intelligence as: 1) the capability of computer systems or algorithms to imitate intelligent human behavior; also, plural, artificial intelligences : a computer, computer system, or set of algorithms having this capability; 2) a branch of computer science dealing with the simulation of intelligent human behavior by computers.
In its 2026 Guide to AI, the IBM Corporation defines it this way:
“Artificial intelligence (AI) is technology that enables computers and machines to simulate human learning, comprehension, problem solving, decision making, creativity and autonomy.”
Any technology that simulates one or more aspects of human brain function could be said to be artificial intelligence. Facial recognition technology processes visual data from cameras the way that our brains process visual information from our eyes in order to recognize specific places and people.
The online translation app I use evaluates the patterns and characters in the words I type and siphons it through the known vocabulary and grammar of my target language in the same way a human translator would.
But just because an application can simulate one or even several of these functions as well or better than a human, does not mean it is intelligent in the way that humans are.
Their core function is extremely fast and incredibly detailed pattern recognition and reproduction. The chatbots and similar products do not “think” in any meaningful sense of the word — they process.
The IBM guide categorizes both current and future AI-assisted technologies as either “weak” or “strong” AI.
Weak AI: Also known as “narrow AI,” defines AI systems designed to perform a specific task or a set of tasks. Examples might include “smart” voice assistant apps, such as Amazon’s Alexa, Apple’s Siri, a social media chatbot or the autonomous vehicles promised by Tesla.
Strong AI: Also known as “artificial general intelligence” (AGI) or “general AI,” possess the ability to understand, learn and apply knowledge across a wide range of tasks at a level equal to or surpassing human intelligence. This level of AI is currently theoretical and no known AI systems approach this level of sophistication. [Emphasis added] Researchers argue that if AGI is even possible, it requires major increases in computing power. Despite recent advances in AI development, self-aware AI systems of science fiction remain firmly in that realm.
There is no sentient AI (yet)
To understand, it can be helpful to look at the different ways that humans and LLMs learn new information.
People can learn new facts and concepts after just a few introductions, and sometimes just a single instance. Think learning a new vocabulary word or the name of something.
This is because human brains posess neuroplasticity, the ability to reorganize their neural networks after exposure to new information. Artificial neural networks don’t have neuroplasticity, and have to undergo a training process called backpropagation in order to incorporate new knowledge.
As IBM Master Inventor Martin Keen explains, “Backpropagation is used during training, when an LLM model processes millions of text examples and adjusts its internal weights to minimize the difference between its predicted outputs and the actual text in the training data.”´
An LLMs knowledge base is also largely static, meaning it does not change based on how it is used. New information is incorporated in discrete training phases, not through interactions with its end users.
“It requires huge numbers of training examples and many forward and backward passes to refine those predictions,” he continues. “So, while you and I may be able to pick up a new word after hearing it once or twice, and LLM may have to see it several thousand times in its training corpus before it can be used reliably in context.”
This is important to keep in mind when interacting with the large, commercially available AI chatbots that have been trained to exquisitely mimic human interaction.
There is a big difference between being able to sound and look like a person, and actually having an individual person’s sense of self, personal motivations and goals.
When it comes to LLMs, at least, just because it looks like a duck, quacks like a duck and swims like a duck — doesn’t make it a duck.
LLMs, Machine Learning, and AI
This is why some people argue that Machine Learning (ML), of which LLMs are a subset, is not truly artificial intelligence.
In order to be true AI, they argue, a technology must be able to process information from different sources, make inferences based on what it learns — without the need for something to be explicitly stated — and then develop original ideas from those inferences.
Machine learning involves predictions or decisions based on data — very sophisticated statistical analysis. It is considered to be “learning” because it is able to make new predictions and change its outputs based on the data rather than having all of its actions completely programmed.
The simplest machine learning programs may have just one artificial neural network or decision tree that information is passed through before the result is obtained.
Sophisticated “deep learning” models may have hundreds of layers of neural networks that interact with each other and parse the training data before getting to the prediction output or “answer” layer. These interactions occur without human oversight or intervention.
LLMs are examples of such deep Machine Learning models.
Other people argue that this represents a form of AI and that Machine Learning is one part of the overall artificial intelligence (AI) framework — that ML is part of AI.
Even AI isn’t *AI*
So, whether you consider LLMs to be AI or not depends on the definition you agree with.
If you think the term artificial intelligence only refers to artificial general intelligence (AGI), then the answer is no. If you use the broader IBM definition, then, arguably, yes.
But one thing is certain, the chatbot or agent you are using does not have and will not develop sentience, personal agency, and intrinsic motivation.
Even the AI experts who think AGI— the ‘strong AI’ — can emerge spontaneously from current technology think it will require huge amounts of computing power harnessed to even larger data sets and using more advanced hardware.
Other experts believe that even the large transformer model LLMs still lack the capacity to lead to AGI, no matter how much compute and data is added.
That doesn’t make the current models useless or some kind of scam. In fact, they are powerful tools that have the potential to lead to amazing advances in science, medicine, and technology. Unfortunately, we have seen that they also have the potential to be misused in devastating ways.
Ensuring that these technologies will do more good than harm will require thoughtful and informed regulation.
And the key to progress in this area will be for everyone to be able to see beyond the myths and hyperbole and understand them for what they are.
Read More
I am a writer, not a computer scientist or engineer. So I wrote this article for people like me who are curious and want to stay informed. Linked below are some articles with more detailed and in-depth information.
IBM: What Are Large Language Models (LLMs)?
IBM: What Is Artificial Intelligence?
IBM: AIs vs. Machine Learning vs. Deep Learning vs. Neural Networks: What’s the Difference?
Medium: “LLMs are a Dead End.” What Yann LeCunn is Really Arguing for Instead.
Tech Policy Press: Most Researchers Do Not Believe AI to be Imminent. Why Do Policymakers Act Otherwise.
Comment Below