AI FactScan- AI summaries are a useful first pass for common, well-covered, low-risk questions.
- For exact figures, breaking news, named authorities, and high-risk medical, legal, or financial questions, return to Google Search and the original source.
- If an answer includes links, use AI FactScan to assess source quality quickly, then open the key sources and verify the details.
Type "look this up for me" into a chatbot and a tidy answer appears within seconds. For many people, AI has started taking over the job they once handed to Google.
The catch is that AI search is not equally reliable for every question. WIRED's guide to using AI search points to three practical factors: how common the question is, how much evidence is available online, and how much harm a wrong answer could cause. Ask the right kind of question and AI can save time. Ask the wrong one and you may get prose that looks like an answer without being one.
Three situations where AI search works well as a first pass
These situations have two things in common: there is plenty of information to compare, and a small error is unlikely to cause immediate harm.
1Simple, widely known facts
Think well-known landmarks, common definitions, or basic unit conversions. Many sites report the same facts, so an AI system has a better chance of finding the right answer.
2Popular topics with plenty of coverage
AI can sketch out how bubble tea calories are estimated or summarize the major differences between iPhone generations. Treat that as an overview. Check current prices, specifications, and model details on an official page.
3Questions with several conditions
For example: "Where can a family of five have dinner in San Francisco for under $200?" A conventional search may return a list of restaurants. AI can organize candidates by location, party size, budget, and dietary needs.
The more obscure the question, the easier it is for AI to retrieve the wrong source
A fifth-billed actor in an old show, a neighborhood restaurant's former menu, or an obscure product number may have only one or two useful pages online. If the target page is hard to find, a retrieval system may substitute something with a similar name, a different date, or a related topic.
In 2026, Stanford HAI tested six commercial chatbots on same-day news questions and classified 1,497 errors. Retrieval failure accounted for 38.8% of them. Another 32.7% came from source divergence, where a model found a related source that reported different facts. When the systems retrieved the correct source, they usually extracted the correct answer. The hard part was connecting the question to the right evidence.
AI search is most dangerous when it retrieves something that looks right but is not the source your question requires.
Four situations where you should return to Google and the original source
- Exact quotes, figures, and specifications. Summaries and rewrites can drop qualifiers or change small details. Opening the original is usually faster than untangling a near miss.
- Breaking news. Old reporting, new updates, and unverified social posts may all appear at once. Start with the people involved, official statements, and news outlets that keep updating the story.
- A specific authoritative source. If you need a health agency notice, court ruling, or research paper, Google Search gives you more control over the source. An AI system may substitute a blog post that covers similar ground.
- Medical, legal, and financial decisions. A mistake can affect your health, rights, or money. AI can help organize questions, but confirm the answer with official documents or a qualified professional before acting.
Why do AI summaries go wrong?
AI search combines web retrieval with a language model. The system interprets the question, selects sources, extracts passages, and writes a finished answer. Each step creates another place for details to drift.
When the source is correct, AI is often good at summarizing it. Trouble starts when the target material is missing, the question is misunderstood, or retrieval returns a similar page as a substitute. The model can still turn those fragments into fluent prose, which makes the mistake hard to spot from writing style alone.
Ask yourself two questions before asking AI
- What happens if this answer is wrong? Does it mean a minor detour, or could it delay treatment, cost money, or affect a legal decision?
- Do I need to know exactly where this claim came from? If a report, assignment, or decision needs a citation, an AI paraphrase is not enough.
If the cost of an error is high, or you need traceable evidence, use AI to find leads and then open the original sources.
Where does AI FactScan fit?
When ChatGPT, Gemini, Claude, or another AI includes links, AI FactScan collects the cited URLs and shows the source mix with A-F grades. That makes it easier to see which sources deserve attention first.
- 1Ask the AI to include sources. A conclusion without sources is difficult to trace.
- 2Scan the cited links. Review the mix of official sources, general media, content farms, dead links, and self-citations.
- 3Open the sources that matter. Check the date, figures, and context to confirm that the original supports the AI's conclusion.
You do not need to read every link from top to bottom. Start with source quality, then spend your time on the claims that could change a decision.