Training data

It is the only explanation that makes a model's blind spots predictable rather than mysterious. Confidently wrong about last month, thin on your local market, fluent on subjects the internet writes about endlessly, and quietly carrying errors it inherited: all of that follows from the material, and none of it is random. It is also where the field's legal arguments live, which reaches you as a buyer through the indemnities in a contract. And it is the honest answer to whether a model knows your business, which is that it almost certainly does not.

The bulk is public web text, joined by books, code repositories and licensed collections, and in some cases material generated for the purpose. Scale is the whole point: these are volumes no person could read in a lifetime, and that breadth is where a general model's ability to be passable at almost anything comes from.

Cleaning and filtering matter as much as gathering. Duplicates are removed, low-quality pages are dropped, and whole categories of material are filtered out before training begins. Those decisions shape the result considerably, they are made by the provider, and they are rarely published in enough detail for an outsider to evaluate.

The material has an end date, usually called the training cut-off. Anything after it is unknown to the model unless supplied at the time you ask. The specific failure worth watching for is not silence but confidence: models frequently answer questions about the period after their cut-off in the same assured tone they use for anything else.

Coverage is deeply uneven and the unevenness is inherited exactly as it was found. Subjects the internet documents endlessly in English are covered in depth. Smaller languages, specialist trades, local regulation and anything recent are thin or absent. A model is not equally good everywhere, and it has no sense of where its own coverage runs out, which is why it does not warn you.

Whether your own conversations become training material for a future model is a question about the provider's terms rather than about the technology. It varies by product, by plan and by whether you are on a business agreement, and it is one of the few things in this whole subject you can settle definitively before buying by reading the contract.

• Ask a general assistant about something that happened recently and watch whether it says it does not know or answers anyway, which is the cut-off showing itself.

  ChatGPT

• Put the same recent question to a tool that searches before answering, where the material arrives at the moment you ask rather than from what was learnt months earlier.

  Perplexity

• Load your own documents into a notebook tool and ask about them, which makes the difference between what a model learnt and what it has been given concrete in one step.

  NotebookLM

Does the model train on what I type into it?

It depends entirely on the product and the plan, and the answer is in the terms rather than in the technology. Consumer tiers and business agreements frequently differ on this point within the same company. It is worth checking before anything confidential goes in, because it is one of the few questions here with a definite documented answer.

What is a training cut-off and why does it matter?

It is the date after which nothing reached the model. It matters because the model rarely announces it: questions about later events get answered in the same confident tone as everything else. Where currency matters, the fix is a tool that searches and cites rather than a newer model, which will simply have a later cut-off.

Does it know anything about my company?

Only whatever was public and well covered, which for most organisations is very little and often out of date. Prices, processes, policies and anything internal were never in the material. If a tool appears to know them, it has been given them, and it is worth confirming which documents.

Was copyrighted material used to train these models?

That is disputed and being litigated in several jurisdictions, and providers describe their sourcing and licensing differently. It is not a question this page can settle. The version of it you can act on is contractual: ask what indemnity your agreement gives you if a claim ever touches output you have published.

• Everything a model appears to know came from a fixed pile of material assembled before release.
• Filtering and mix shape the result as much as volume does, and are rarely published in detail.
• The cut-off is a hard edge, and the failure to watch for is confidence rather than silence.
• Coverage is uneven and inherited, and the model cannot tell you where its own coverage thins.
• Whether your conversations train future models is a contractual question you can answer before buying.

• ChatGPT

  The cut-off is easiest to feel here, by asking about something recent.

• Perplexity

  Searches before answering, so the material is current rather than learnt.

• NotebookLM

  Answers from documents you supplied, which is the contrast case in one step.