Using AI to learn

The first thing to notice is who is writing. A growing share of the recent pages a search finds comes from content farms, sites that publish as much text as they can, as cheaply as they can, because their income comes from the traffic a search engine sends them, not from readers. AI tools have made that cheaper still. A page written in bulk by a model can be published last month, answer your exact question in two thousand fluent words, and have no author and no one to ask when it is wrong. The fluency of the page is not evidence for the page.

The second thing is the date. A page with nothing new to say has learned to look new. It gets re-dated to this year, with an updated line and no record of what changed, because for many of these pages the update is the date itself, or a new affiliate link inside the recipe. The trick works because search engines and readers both share recency bias, the habit of trusting the newest page, so the re-dated page rises in the list even when the content has not moved. The sixth screen on page one, the re-dated post, walks through one of these posts clue by clue.

Then there is the order of the list. The results are not sorted by how good the pages are. They are sorted by a ranking that mixes relevance with who is paying for ads and who has the most links pointing at the page. The sites that win that ranking are the ones that study the search engine rather than the topic, and the top of the list is now where the commercial pages live. The list is still a place to start looking, and starting is not the same as trusting.

None of this makes the strong sources easier to find. The page where a person who knows has written the thing down is often a university department, a government data set, or a long report, and those pages are plainer than the pages around them and lower in the list. Plainness reads as carelessness to a reader who is skimming, and age reads as uselessness to the ranking, even when the question does not depend on the date. A lot of research now starts from a summary of a summary, and every copy is further from the source than the one before it.

There is an honest limit to this argument, and it is worth stating. None of this is a claim that the chatbot is the better researcher, because a model invents sources with a confident voice too, and the defence in both cases is the same. Open the thing, and check what it actually says. What has changed is narrower than that. The old default of just looking it up used to be the whole method, because the list was mostly pages written by people about things they knew. The list is no longer that. Looking it up is still the start of the method, and checking what you find is the rest of it.