AI & Future
AI chatbots sound confident even when they are wrong. Here is a quick, practical routine for checking their answers so you do not get caught out by a slip.
The fastest reliable way to fact-check an AI answer is lateral reading: leave the chatbot open, start a new tab, and confirm each specific claim against an independent source you already trust. Fluent prose is a writing style, not evidence, so treat every name, date, number, quote, and citation as unverified until something outside the chat window backs it up.
Large language models like ChatGPT, Claude, Google Gemini, and Microsoft Copilot don't look answers up in a database. They generate text by predicting the next most likely token (roughly, the next chunk of a word) based on patterns in their training data. Accuracy is a side effect: true statements are common in the text these models learned from, so the likely-sounding answer is often the correct one. But there is no internal fact-checker, and no signal that separates "I'm certain" from "I'm improvising."
That gap produces what the field calls hallucination (or, more precisely, confabulation): the model states something false with the exact same fluency it uses for facts. It will invent a statistic, misattribute a quote, or cite a study that was never written, all in clean, confident prose. In human writing, confidence usually tracks expertise, so we read polish as reliability. With an LLM that link is severed, which is precisely why smooth answers are so easy to over-trust.
Two structural limits make certain errors predictable. First, every model has a knowledge cutoff, a date after which it wasn't trained, so anything recent (prices, software versions, election results, who currently holds a job) can be stale unless the tool is actively searching the web. Second, retrieval features like Perplexity's inline citations or ChatGPT and Gemini web browsing use RAG (retrieval-augmented generation) to ground answers in real pages, which reduces hallucination but does not eliminate it. The model can still misread or misquote a source it genuinely retrieved.
Fact-checking everything is exhausting and unnecessary. Scale your scrutiny to two things: how specific the claim is, and what happens if it's wrong.
High priority: names, dates, numbers, statistics, direct quotes, legal or medical claims, and any step-by-step instructions where a wrong step causes damage (dosages, tax filings, electrical work, code that touches production data). These are exactly where models fumble and exactly where errors cost you.
Low priority: broad conceptual explanations, brainstorming, summaries of well-known ideas, and drafts you'll rewrite anyway. If the AI explains how compound interest works in general terms, you rarely need a source. If it gives you the specific rate, contribution limit, or filing deadline, you always do.
Professional fact-checkers use a method called SIFT (Stop, Investigate the source, Find better coverage, Trace claims to the original), popularized by digital-literacy researcher Mike Caulfield. A stripped-down version works well for everyday AI use:
This "sideways glance" usually takes less time than re-reading the answer, and it catches the large majority of slips.
Asking "What's your source?" is smart, but it's only half the job, because models fabricate citations as readily as facts. A hallucinated reference can arrive with a plausible title, a real-sounding author, a journal name, and even a formatted DOI, all for a paper that does not exist.
So follow every citation to its origin:
The stakes here are not hypothetical. In the 2023 U.S. case Mata v. Avianca, a lawyer submitted a brief containing several court decisions that ChatGPT had invented; the judge imposed sanctions once opposing counsel couldn't locate any of them. A fabricated citation is more dangerous than none at all, because it borrows the authority of a reference while being entirely made up.
Some topics warrant caution every single time, because the downside is real harm:
Not reliably. Models don't have accurate insight into their own certainty, so "Are you sure?" tends to yield a plausible-sounding second answer rather than a genuine confidence check. Sometimes it stubbornly repeats an error; sometimes it abandons a correct answer under pressure. Verify against an independent source instead.
No, but they make checking faster. Inline citations point you to real pages, which is a head start, but the model can still overstate or misread what a source actually says. Click at least the citations behind any claim you'll act on or repeat, and confirm the page supports that exact statement.
Check whether the tool is actively browsing the web for that answer; if it isn't, treat time-sensitive facts as potentially stale past its knowledge cutoff. For prices, software versions, laws, and current events, confirm with a dated, primary source. When in doubt, ask the tool directly whether it searched, then verify anyway.
Lateral reading. The moment an answer contains a specific fact you'll use, open a new tab and confirm it somewhere you already trust. It takes seconds, needs no special tools, and catches the errors that fluency is otherwise hiding.
Priya translates the fast-moving world of AI and the internet into things you can actually use and understand. She's curious but skeptical, quick to separate genuine progress from hype, and keen to help readers use new tools wisely rather than fearfully.
Keep reading
AI news moves fast and most of it is noise. Here is a calm, jargon-free system for staying informed about what matters without burning out on every headline.
Priya Nadar
Recommendation systems shape what you watch, buy, and read every day. Here is a clear, jargon-free look at how they work and how to stay in control.
Priya Nadar