Fantas-Tech?
5 tips for getting better answers from AI
By Geoffrey A. Fowler The Washington Post
Published Sept. 8, 2025
Contact The Editor
Articles By This Author
Want to do better online research? Use artificial intelligence like a librarian.
Three professional librarians recently helped me evaluate AI search tools, from ChatGPT 5 to Google's AI overviews, by asking the bots a bunch of tough questions. Our goal: finding which AI could save us time, while also not giving wrong answers.
These days, many public libraries and universities have "AI librarians," like the ones that assisted me, to help people use the latest tech for research.
Watching the librarians at work, it became clear they approach AI with the same healthy skepticism you might direct at a used-car salesman. It's trust, but most definitely verify. "Instant information gratification, which I define as the need to accept information without rigorously interrogating it, is dangerous," says Trevor Watkins, a librarian at George Mason University.
Here are five practical lessons from our testing and conversations.
1. Start with Google's AI Mode - not AI Overviews
The winner the librarians chose for our AI search test isn't most people's first stop for online research. But it should be.
AI Mode is a hybrid version of Google with a chatbot. You can access it by tapping on the button in the top left corner of search results.
AI Mode beat all the AI tools we tested (including ChatGPT 5) - and especially outperformed Google's AI Overviews that appear by default on top of most search results.
Why is AI Mode better? Google says it uses AI "reasoning" to identify subtopics to dig into and then conducts multiple - possibly dozens and dozens - of searches all at the same time. Unlike AI Overviews, it also has access to newer tech updates to Google's AI model, which helps it analyze and summarize what it finds.
There are still plenty of questions AI Mode struggles to answer. But AI Mode is particularly useful for needle-in-a-haystack type research, where its ability to quickly search broadly is most useful.
2. Be very, very specific
Most AI tools aren't smart enough yet to ask key follow-up questions, a process human librarians call a reference interview. It's "one of many things that separates librarians from all generative AI tools," says Watkins. But with AI, "it is up to you, as the user, to rephrase your question and, when needed, provide more context."
When librarians do this, it's about drilling down to what exactly you want to ask but maybe haven't fully formulated yet, says Chris Markman, a librarian at Palo Alto City Library. "Nine times out of 10, what they're looking for is two or three layers beyond the initial question," he says.
"When I'm coming up with a prompt, I'm trying to build that thought process into the initial question while still being as concise as possible," he says.
Adding examples to your question can help, as can limiting the time period for potential sources. "Sometimes we get really lazy with remembering to give the AI model a role, providing context or constraints," says Sharesly Rodriguez, a librarian at San José State University.
Rodriguez suggests you could also ask the AI to criticize itself or show multiple examples, with prompt phrases such as "argue the opposite viewpoint," "identify weaknesses in your thinking," "list what might be missing" or "what biases might be here."
3. Beware of AI blind spots
In our test, all of the AI tools demonstrated weaknesses that you should keep in the back of your mind. These included questions about:
• Recent events: The process of making an AI model is lengthy, so their built-in knowledge is frozen in time. For example, without a web search, ChatGPT 5 doesn't know anything after Sept. 30, 2024. If you need to ask about something more recent, be sure to explicitly instruct the AI to check up-to-date information.
• Visual elements: AI is still much better with text than with images, and can get very discombobulated when you ask it to answer questions that involve analyzing images.
• Niche sources: Because of publishing agreements, AI access to academic, news and other media content may be limited. "The majority of scholarly research is still not ‘open,' findable or accessible, even on search engines," says Rodriguez. That means for some questions, you really may be better off asking a human librarian who has access to paid databases and sources.
4. Check the answer citations (really)
In our tests, AI tools would sometimes give an answer, complete with links to sources. Then the answer would still turn out to be wrong. "It's so easy to fall into that trap when they look and sound so authoritative and confident," says Rodriguez.
So you're still going to want to click on the sources to see what they have to say. Or if they're even real.
And if the link confirms the answer from the AI, can you really trust where it came from? "Click through to sources, verify authorship, look at the organization behind it and note the date of publication," says Rodriguez. AI tools have a hard time judging the authority of sources, so they can surface blogs, low-quality news or SEO junk instead of legitimate peer-reviewed sources.
One red flag is when the AI sources its answer to Reddit, the community message board, or a social network like X. There's nothing wrong with people's opinions, but they're not the same as facts.
5. Ask the same question twice (or three times)
If you've got an extra moment, open a few extra browser tabs and ask the same question to different AI tools. You might be surprised at how the answers diverge - or pleased when they don't.
You're a lot better off, says Markman, if you have at least some knowledge about the topic if you're exploring when you're trying to figure out whether to trust an AI answer. "I like to call this the ‘two book minimum'" for AI, he says. "You don't need to know everything about a subject, but you should know enough to spot a mistake and move on rather than going on wild goose chases with misinformation."
And when in doubt, "get help from librarians" says Watkins. "Many public and some academic libraries offer a service called virtual reference, where users can ask questions to library professionals (not chatbots) online."
(COMMENT, BELOW)
Every weekday JewishWorldReview.com publishes what many in the media and Washington consider "must-reading". Sign up for the daily JWR update. It's free. Just click here.