Five words that hide more than they say

Vocabulary as a defence

Most of what makes AI coverage hard to follow is not the technology. It is five words used loosely by people who benefit from the looseness. Here is what each actually means, what it is usually made to mean, and the question that cuts through.

“Algorithm”

An algorithm is a procedure: a finite sequence of steps that turns an input into an output. Long division is an algorithm. Sorting a list is an algorithm.

In press coverage, “the algorithm” has come to mean the mysterious force deciding what you see. That usage hides the important distinction. When a company says its algorithm ranks posts, there are two very different possibilities: a set of rules a person wrote and could show you, or a trained model whose behaviour nobody can print out. Only one of those can be audited by reading it.

The question: is this a procedure somebody wrote, or a model somebody trained?

“Neural network”

The name comes from a 1943 attempt to describe brain cells mathematically, and it has stuck for eighty years. The connection is loose to the point of misleading.

An artificial neuron multiplies its inputs by weights, adds them up, and applies a simple function. A biological neuron communicates in spikes rather than numbers, has thousands of connections with their own chemistry, is modulated by dozens of neurotransmitters, changes its own structure, and lives in a body. Neuroscientists do not use artificial neural networks as models of the brain, except in narrow research settings.

The name causes a specific error: people conclude that because it is brain-like, it must have brain-like properties: understanding, intent, consciousness. Nothing licenses that step.

The question: what does this network actually compute, rather than what is it named after?

“AGI”

Artificial general intelligence. There is no agreed definition. Proposals range from “outperforms humans at most economically valuable work” to “can learn any task a human can” to specific revenue targets written into commercial contracts.

Because there is no definition, there is no test, and because there is no test, the claim that it is two years away can neither be verified nor refused. The term does reliable work in one place: fundraising. It is also used sincerely by researchers who disagree with each other about what it names.

The question: what specific capability are you claiming, and how would we know it had arrived?

“The AI”

There is no the. There are thousands of models, from many organisations, differing enormously in capability, and the same brand name may sit on top of several of them, routed by cost.

“AI can now pass the bar exam” means one particular model, under particular conditions, on a particular version of the exam. The claim is often true and almost never transfers. “AI is bad at maths” was accurate for one generation of models and much less accurate for the ones with a reasoning step, which are also slower and dearer.

The question: which model, which version, tested how?

“Trained on your data”

The single most slippery phrase in a privacy policy, because it can mean two different things and both are legal.

It can mean your text was included in the material used to fit the weights of a future model. That is durable, effectively irreversible, and hard to audit.

Or it can mean your text was sent to a model at the moment you asked, used to produce a reply, and (depending on the policy) retained for a period, reviewed by staff for safety, or deleted. Nothing entered any weights.

Some policies distinguish these carefully. Many use “we use your data to improve our services”, which covers both and commits to neither.

The question: does this text enter model training, and if so, is there a switch to stop it? Business and enterprise plans usually say no by default. Free consumer plans usually say yes unless you opt out, and the opt-out is generally a real setting worth finding.

Two more worth having

Parameters are the numbers in the file, fixed after training: the model’s size. Tokens are the chunks of text going in and out: the model’s traffic. Context window is how much text it can consider at once: its desk space. These three get mixed up constantly, including by journalists, and they measure completely different things: 8 billion parameters, 500 tokens, a 128,000-token window are three unrelated facts about one system.

Keeping them apart is a small thing that will make the rest of this course, and most of the news you read, noticeably easier to follow.