Definition
What AI is — and is not
In everyday talk, “AI” usually means software that was trained on examples and then produces a new output — text, an image, a transcript, a suggested next click. The useful picture is not a digital brain. It is a very large pattern matcher: given an input, it predicts a plausible continuation.
What the term usually names now
Modern consumer AI is mostly machine learning. Instead of a person writing a rule for every case, the system adjusts internal numbers so that its outputs better match examples. After that training run, it can be asked new questions. That second step is inference, not a new education each time you type.
The current wave people meet is often a generative model: it produces new text or pixels that look like the data it saw. Looking fluent is not the same as knowing a fact. The system is optimizing for a likely next token or pixel, not for a verified record.
What AI is not
- Not a person. It has no legal personhood on this page, no inner life we can point to, and no duties of its own. When a chat sounds caring or sure, that is style learned from text, not evidence of a mind.
- Not automatically correct. Fluency hides gaps. The same system can draft a clean paragraph that invents a citation, a date, or a law. See risks without hype.
- Not one product with one owner. “AI” is a category. Many labs train models; many companies wrap those models in products. Mixing the category with a brand is how marketing wins arguments it has not earned. See models versus products.
- Not magic, and not nothing. These systems are useful at drafting, summarizing, translating, and spotting patterns in messy input. Usefulness does not require consciousness, and skepticism does not require pretending the tools do nothing.
A short history, without a myth of sudden birth
The field is old. Rule-based expert systems, statistical language models, and neural networks all predate the current chat products. What changed for the public is scale, interface, and distribution: larger models, a box you can type into, and companies that ship that box to phones.
This desk will not invent a founding year, a single inventor, or a percentage of “all jobs” that will vanish. Those figures circulate. They are not load-bearing here unless a named public source is attached — and even then they are estimates, not measurements you can treat like a bank balance.
How to use the rest of this site
If you need the split between the trained file and the app around it, read models versus products. If you need the everyday harms, read risks without hype. If you need a method for press releases, read how to read an AI claim.