When it searches the web

A different failure, not no failure

Most assistants now retrieve web pages and cite them. This is a genuine improvement: the model is working from text it just fetched rather than compressed memory, so pure invention drops sharply.

What replaces it is subtler and easier to trust by mistake.

The link is real, the claim is not in it. The most common failure. A citation attached to a sentence the page does not support, sometimes because the page is about the topic generally and the specific claim was generated around it. Independent evaluations of search-grounded assistants keep finding this in a substantial minority of citations.

The source is weak. Retrieval optimises for what ranks, not for what is reliable. A content-farm article, an SEO page, a five-year-old forum post and a Cochrane review are all just retrieved text.

The summary distorts. Even from a good page, conditions and hedges drop out. May reduce risk in some populations becomes reduces risk.

Only part was read. Retrieval usually fetches a portion. The model can answer confidently about a document it saw a fragment of.

Circularity. An increasing share of the web is itself model-generated. A claim can be confirmed by three sources that all originate from the same unverified generated text. Multiple agreeing pages are weaker evidence than they used to be.

The check, which has not changed

Open the link. Find the sentence. That is the whole method, and it is why this course keeps repeating it: every generation of this technology changes what fails and leaves the check identical.

When you open it, ask three things:

1. Does this page say what the answer said?
2. Is this page a source or a repeater? A news article about a study is not the study. Follow it back.
3. Would I have cited this if I had found it myself? A page you would not have chosen does not become citable because a machine surfaced it.

Deep research modes

Several tools now offer a longer, multi-step research mode that runs many searches and returns a report with dozens of citations. These are more thorough, and they concentrate the risk rather than removing it: a twenty-page report with sixty citations invites acceptance precisely because checking it looks expensive.

Use them as a map, which is what they are good for: what the areas are, which names recur, which terms to search. Then do your own reading of the sources that matter. Never submit a report of this kind, or its reference list, as your own work, and note that a bibliography assembled by a tool you did not verify is a bibliography of sources you have not read, which is the misconduct finding from two lessons ago.

Wikipedia, used correctly

Wikipedia is an excellent starting point and not a citable source, and both halves are true for the same reason: it summarises, and it tells you what it summarised from.

The correct use is to read the article and then use the reference list at the bottom. Those are the real sources; go to them, read them, cite them. The article’s talk page is also unusually informative: it is where disputes about the content are argued out, so you can see immediately whether a claim is contested.

And for offline work, Kiwix packages Wikipedia for local reading, which is free and useful with intermittent data.

The one place it is unambiguously better

Retrieval is a real improvement for anything time-bound: this year’s entry requirements, the current version of a library, a deadline, an exam board specification, who currently holds an office. A model without search has no route to any of these and will answer from whenever its training stopped, often without saying so. If your question has a date in it, or implicitly has one, switch search on and open the page.

Building the habit

A short discipline for any answer with links:

• Open every link you intend to rely on. Not the ones you are ignoring.
• Search within the page for the specific term from the claim.
• If the page is reporting somebody else’s work, follow it back to the original.
• If you cannot find the claim in the page, the claim is unsupported. Delete it or find a real source. Do not keep it because it sounded right.

That last instruction is the one people break. A claim that survived because it was plausible, attached to a source that does not contain it, is the single most common way a careful student ends up in an integrity meeting.