When I hear companies like Anthropic make claims that they don’t know why AI comes to some conclusion, or they suggest that Claude somehow “resorted to blackmail” to prevent itself from being shut down as featured on 60 Minutes, it makes me both angry and a bit fearful all at once. AI marketing, at least as it exists at the time of this article's writing, is filled with disingenuous assertions that create marketing leverage and product branding. It invokes fear, confusion, and hype. There's an important distinction being blurred here. We may not be able to trace why AI produces a specific response to a given prompt at any given time because of the sheer magnitude of steps but that is quite different than saying we don't know how it works. IT systems can suffer from the same problem, but thankfully they are much easier to solve because they are much smaller. Conflating the two is, at best, imprecise and at worst, deliberate. For example, it may be true that Anthropic's latest model Mythos, announced in April 2026, can find and exploit system vulnerabilities but introducing it with doomsday theater benefits them.
I’ll state upfront that I’m not an AI researcher. I haven’t designed and built frontier models, nor am I anti-AI. But I’m AI-adjacent, having been in the tech industry for quite a while and having a pretty general idea of what’s going on. I’ve also written my own document classification tool (github:J8k3/document-sorter) back in the pre-AI coding era, taking a much smaller-scale approach to model training. I've revisited that tool recently using AI assisted development and though I wasn't incredibly far off at the time, refined that tool to be more effective.
AI is not a sentient entity, full stop. It has no continuous consciousness. It is not sitting there thinking about you between requests. It is a request and response between a client and server, the inference layer, that invokes a model on the backend. In between, there is idle capacity, not a self-aware or self-thinking model.
However, it can be run iteratively to provide what seems like reasoned output as it processes its own prior responses. Mix that with a strategy to break down input into relevant pieces, tokenization, retrieval, and recursion, and you’ve got something that can look a lot like thinking. When you ask AI to do some task, or answer some question, it uses a very large statistical model, called a Large Language Model (LLM), to predict the next most likely token based on the training corpus, a massive body of human-written literature and/or code, used in its development.
Turns out humans have written an enormous amount of content over the past several thousand years, and there are patterns in language and code that make it predictable enough that a tool like AI can make inferred predictions about what comes next in increasing increments: words, sentences, paragraphs, or whole documents. That prediction is probability without authority or verification, not understanding which is where hallucinations come from. With sampling and weighting, developers introduce a degree of randomness into what the model predicts so that outputs are varied instead of identical every time. This makes responses feel more natural and less mechanical.
That means the training data has a massive impact on what the model produces, and this is where bias comes into play. If you feed the model writings from a particular period, culture, or worldview, it's going to reflect the biases present in that material. AI developers inject safeguards to skew responses away from those biases. But even that is simply testing outputs, sometimes with and sometimes without humans in the loop, against rules representing socially accepted norms of the present day. This is done through processes like reinforcement learning, where the model produces a response, it's reviewed, ranked, and sometimes rejected. If you've ever seen a model show you two responses and ask you to pick the better one, that's an example of reinforcement learning being outsourced.
How does AI make these predictions? An enormous mathematical representation is put together that represents relationships of tokens in training data. A token is a unit of data (input/output) which can be a word, part of a word, punctuation, etc. OpenAI provides an example of how text is tokenized https://platform.openai.com/tokenizer. It learns that given "X" token, probabilistically, "Y" is the next token. The model takes your tokenized input and produces a response. Sampling introduces controlled variability so responses are not perfectly deterministic. That is what gives it personality. Not consciousness.