Why it cannot look it up

Two very different things wearing the same interface

When you type a question into a chatbot, one of two things happens, and the box looks the same either way.

Mode one: the model answers from what it learned. Nothing is consulted. No page is opened. The answer is produced word by word from the patterns absorbed during training, which finished months ago. This is what a language model is. It cannot check, because there is nothing to check against: the training text is not stored as a library it can open, it is dissolved into billions of numerical weights.

Mode two: the tool searches first. Many products now bolt a web search onto the model. Your question is turned into a search, some pages are fetched, and the model is handed their text along with your question. It then writes an answer that summarises what it was given. This is a genuine improvement for anything recent or specific, and it changes what you should worry about.

How to tell which happened

Look for sources. A searched answer shows links, citations, or a line like “Searched the web”. An unsearched answer shows nothing. If you see confident specifics and no sources, you are in mode one, and the specifics are reconstructed from old patterns.

Most free tiers search automatically for questions that look like they need it, and not otherwise. You can force it: “Search the web before answering, and show me the pages you used.” You can also forbid it, which is useful when you want the model’s general explanation rather than a summary of whatever ranked first on the web this morning.

What search fixes

• Staleness. The training cutoff stops mattering for the question at hand. The new syllabus, this year’s dates, the latest results are reachable.
• Local specifics, sometimes. If your district’s population is on a page the search finds, you get it. If no page has it, you are back to guessing, and the model may not tell you that it found nothing useful.
• Checkability. You now have a page to open. This is the real gain. Open it.

What search does not fix

The model still has to decide which pages to trust, and it decides the way it decides everything: by pattern. A confident, well-formatted page ranks with an accurate one. A content farm that copied an old syllabus looks like the exam board. If three pages disagree, the model may blend them into a fourth answer none of them said.

It also summarises, and summaries lose things. The exceptions, the footnote that says “for the 2024 cohort only”, the conditions: these fall out first, because they are the least predictable part of a sentence.

And the search itself can go wrong. A vague question produces a vague search, which fetches irrelevant pages, which produce an answer built on the wrong material. When an answer seems oddly off, look at what it searched for. Often the search query was not your question.

The teacher’s rule

For anything that must be right (a date, a figure, a rule from the board) the model’s job is to find the page, and your job is to read it. Treat a searched answer as a research assistant’s note: “I think it’s on this page, and it seems to say this.” Then open the page.

For anything conceptual (an explanation, an analogy, a set of likely misconceptions) search usually makes the answer worse, because it replaces the model’s broad synthesis with a summary of two websites. Turn it off for those.

An example that comes up every year

A teacher asks for the marks allocation in this year’s Grade 12 physics paper. Without search, the model produces a plausible table from previous years. With search, it may find the board’s circular, or it may find a tutoring site’s guess from last year, and it will present either with the same tone. The only safe version of this question is: “Find the official circular from the board’s own website and give me the link.” Then you read the circular.

That is not the tool failing. That is the tool doing exactly what it is: a very good guesser, sometimes with a web page in hand.