AI Life / Work & Study / Research & Reading
Both answer questions with links to sources. The best independent test of how accurate those answers are comes from the Columbia Journalism Review’s Tow Center. More on the research and reading hub .
The study
The Tow Center tested eight AI search engines on finding and citing news articles. Perplexity answered incorrectly 37% of the time. ChatGPT Search was wrong 67% of the time (CJR ). The headline: they’re all bad at citing news, and Perplexity was less bad.
What that means for work
- Use them to find sources, not to be the source. Click through and read.
- Ask for links. Then check that the link says what the answer claims.
- Be careful with numbers. Prices, dates and stats are where errors hide.
Which one
- Research with sources: Perplexity did better in the test.
- Drafting and thinking through a problem: ChatGPT.
For students, see best AI study tools . For research from your own documents instead of the web, Gemini Notebook sticks to your sources.
What my agent can do with it
I let agents search the web, but I don’t let them trust it. I have them return the link with every claim so I can check. Agents fetching web pages can also pick up prompt injection; I covered one version in MCP discovery prompt injection .