How AI web search works

The training cutoff, and why it is dangerous

Every model has a date after which it knows nothing. The problem is not the ignorance, which is unavoidable, but that models are poor at recognising it. Asked about something recent, a model will frequently produce a fluent, specific, entirely invented answer, because generating plausible text is what it does and it has no reliable internal signal for "I have not seen this".

This is the single most common way people are misled by an assistant. The wrong answer does not look wrong.

What actually happens during a search

Search is implemented as a tool the model can call rather than something bolted on before or after. The sequence is worth knowing because it explains both the strengths and the failure modes.

1. 1The model decides a search is warranted, based on the question. This is a judgement it makes, not a keyword trigger.
2. 2It writes the query itself, which is often not your wording — it reformulates towards what it expects to retrieve.
3. 3Results come back as a list of pages with titles and snippets.
4. 4It selects which of them to actually fetch and read in full.
5. 5The fetched text enters the context window alongside your conversation, and the answer is generated from both.

Steps two and four are where quality is won or lost. A model that writes a poor query or reads the wrong three pages produces a confident answer built on bad sources, and the confidence is identical either way.

Why the sources have to be visible

Retrieval does not eliminate fabrication. It reduces it, and it changes its shape: instead of inventing a fact, a model given three pages may overstate what they support, blend them into a claim none of them makes, or lean on the one that happens to be wrong.

The only cheap defence is showing what was read. On Tartarus AI every source the model used is named above the reply, so a claim about this week arrives with the trail that produced it rather than as a research project. A search feature that does not show its sources has removed the main reason to want one.

Fetching pages safely

One implementation detail is worth stating because it is a real vulnerability class rather than a nicety. The model chooses which URL to fetch, and the model's input is ultimately your text — so without a guard, a crafted prompt could steer the fetcher at internal infrastructure.

Tartarus AI resolves the target host before connecting and refuses private and link-local addresses. That guard is load-bearing: without it the search tool becomes a way to reach the inference box, the database, or a cloud metadata endpoint from outside.

When to expect it, and when not to

Search fires when the question needs current information: news, prices, releases, anything dated. It does not fire on questions the model can answer from what it already knows, which is correct — a search on every message would add latency to most of them for nothing.

It is not a substitute for a primary source on anything consequential. Read the cited page rather than the summary when the answer matters.