Search engines vs generative AI

A search engine points toward material published by others; a generative tool writes a response in its own words. A generated answer may blend useful information with mistakes or invented details.

Official institutions, textbooks, reputable news, and local experts can be stronger evidence than an unattributed post. Check date, author, purpose, and whether the source supports the exact claim.

A search engine feels instantaneous, but the work behind a results page happens in three distinct stages, and understanding them explains both why search is trustworthy in one way and untrustworthy in another. The first stage is crawling. Automated programs called crawlers move continuously across the web, following links from page to page and copying what they find, much like a librarian who walks every aisle noting each book. The second stage is indexing. The copied pages are broken into words and stored in an enormous lookup structure — conceptually like the index at the back of a textbook — so that the engine can find every page containing 'maize planting dates Zimbabwe' without rereading the whole web. The third stage, ranking, is where most of the intelligence lives. For any query there may be millions of matching pages, so the engine scores them using hundreds of signals: whether the words appear in the title, how many reputable sites link to the page, how fresh it is, whether it loads properly on a phone, and what people who searched similar things clicked before. Machine learning combines these signals into the final ordering. Now work through an example: you search for a ZIMSEC examination timetable. The official zimsec.co.zw page ranks highly because many other sites link to it and its title matches your words, but a blog post from three years ago may also rank well if many people clicked it. The engine measured relevance and popularity — it never checked whether the timetable is current. That is the crucial limitation: ranking predicts usefulness, it does not certify truth, and pages can be deliberately engineered to rank well, an industry called search engine optimisation. The common misconception is that the top result is the verified answer. In reality the top result is the best guess of a relevance model, so the searcher still owns two jobs the machine cannot do: choosing which result deserves trust by checking who published it and when, and confirming that the page actually supports the specific claim needed. Search hands you a ranked reading list, and that is genuinely valuable — but reading, judging, and verifying remain human work.

When a chatbot states that a book, statistic, or court case exists when it does not, it has not lied, malfunctioned, or accessed bad information — it has done exactly what its design makes it do, and seeing why is the key to using it safely for research. A large language model generates text one token at a time, each time choosing a plausible continuation given everything written so far. During training it read enormous amounts of text and compressed the patterns into billions of numerical weights. It did not store the documents themselves, and it has no database of facts to consult at answer time. Everything it 'knows' exists only as tendencies in those weights: strong, frequently reinforced patterns reproduce reliably, while rare or specific details blur together. Hallucination is what that blurring looks like from outside. Ask for the population of Zimbabwe and the model answers well, because that fact appeared thousands of times in training text. Ask for the exact title and page number of a 2009 report on smallholder irrigation in Masvingo, and the model faces a prompt that demands a specific-looking answer it has no crisp pattern for — so it assembles one from fragments: a plausible institution, a plausible year, a plausible title in the standard format of report titles. Every piece is statistically reasonable; the whole is fiction. Crucially, the model produces confident and hedged text with the same machinery, and its training rewarded fluent, complete-sounding answers, so its tone carries no information about its accuracy. Work through the practical consequence: a student who asks a chatbot to explain the difference between weather and climate will usually get an excellent answer, because that explanation is a heavily reinforced pattern. The same student asking for five sources with page numbers may receive a beautifully formatted list where three references do not exist. Same tool, same session, opposite reliability — the difference is whether the request lives in dense or sparse training territory. The misconception to correct is that hallucinations are rare glitches that better models will soon eliminate entirely. Newer systems reduce them, and tools that search the web before answering ground themselves better, but any system that generates text from learned patterns can generate plausible falsehoods. The durable rule for research follows directly from the mechanism: use generative AI to explain, reframe, and organise, and use search plus original sources to establish facts.

Brian needs a current figure for a report on Zimbabwe's population. A responsible response is: Use the chatbot's number as a lead, then verify it in an official statistical source. Correct. The chatbot can help frame the search, but the report needs a checkable source. Use this case to separate what the technology contributes from what people contribute. The team should compare the intended outcome with a baseline where applicable, record important assumptions, and keep a clear route to correct or stop the process.

Ask whether the examples, data, and assumptions fit your school, company, or community. Generic demos often miss local names, laws, connectivity, and languages.

Libraries and official pages remain the place to verify. A chatbot paragraph is a draft, not automatic proof.