Start here
You do not need to be technical. You do need to not be guessing.
This page assumes you have never used any of these tools and do not know what the words mean. The jobs most exposed are at the bottom: logging, transcribing, the first pass, the entry-level task somebody used to be paid to learn on. Those are the jobs people without contacts get.
Last checked September 2026.
What they are
A good guess, made quickly.
A large language model has been trained on an enormous amount of text and learned which words tend to follow which; asked something, it produces the most plausible continuation. An image or video model does the same with pixels. They are not looking anything up and not reasoning about the world. Plausible is very often correct, which is what makes the wrong answers dangerous.
What they cannot do
The list.
They do not know what is true. A model will state a rate, a law or a credit with total confidence and be wrong.
They do not know what happened recently. Unless a tool is explicitly searching, its knowledge stops at some point in the past.
They cannot be relied on to repeat themselves. Ask twice and you may get two answers — a problem for anything that has to match across shots or days.
They do not understand consequence. A model has no idea that the schedule it rearranged puts a crew inside their turnaround, or that the face it generated belongs to a living person.
They are not private by default. What you paste in may be stored, and on some services used to improve the product.
Your settings
The cheapest thing you can learn today is what your defaults are already doing.
Almost every real-world breach we can find happens through a default rather than a decision. None of this needs technical knowledge and none of it needs permission from anybody.
Midjourney publishes your images
On the entry plans the default is public. An art department generating unreleased designs on a basic subscription is publishing them. Privacy, which Midjourney calls Stealth Mode, starts at the Pro tier, $60 a month on Midjourney’s published prices in September 2026. Under a non-disclosure agreement the cheap plan is arguably already the breach.
A free assistant tier may train on what you paste
OpenAI’s pricing page shows content used to train the models with an opt-out available. Anthropic’s states training is off by default on paid plans. Google’s consumer page does not state business data handling at all. Three companies, three answers, one setting you can check in a minute.
The commissioner’s rule is stricter than the tool’s
Netflix’s published guidance, August 2025, requires that tools used on a production “cannot store, reuse, or train on production data inputs or outputs”. A higher bar than most consumer plans meet, and it applies to whatever you personally happen to be signed into.
The vocabulary
Ten words that unlock most of the conversation.
| Word | What it means |
|---|---|
| Model | The trained system itself. The app you type into is a wrapper around one. |
| Prompt | What you ask for. The biggest lever you have over the quality of the output. |
| Context window | How much the model can hold in mind at once. Go past it and the earliest part is forgotten silently. |
| Token | The chunks text is broken into, roughly a short word each. Pricing and limits are counted in these. |
| Training data | Everything the model learned from. Who owned it, and whether they agreed, is the central legal fight. |
| Hallucination | A confident, fluent, wrong answer. A property of how the thing works, not a bug you can switch off. |
| Diffusion | The method behind most image and video generators: start from noise and refine towards the description. |
| Fine-tuning | Further training on a specific set of material, to make a general model behave in a particular way. |
| Open weights | A model whose files you can run yourself, so your material never goes to somebody else’s server. |
| Provenance | Metadata recording how a file was made and whether a model touched it. C2PA is the standard people mean. |
