Stanford study offers guidance on AI-assisted research
Stanford researchers tested ChatGPT, Gemini and Claude on recent-event questions. The source gives AI-search guidance but no sample size, accuracy rate, model versions or product-level results.

- Researchers tested 3 AI systems—ChatGPT, Gemini and Claude—on multiple-choice questions about recent events.
- The test assessed the systems’ effectiveness as a news source, according to Zetlin.
- Zetlin’s guidance covers 4 practical uses, ranging from sourdough questions to multi-parameter queries.
- One example searched for dinner for a family of 5 in San Francisco for less than $200.
- The source reports no test date, sample size, accuracy percentage or model version.
Stanford University researchers tested three AI systems—ChatGPT, Google Gemini and Claude—on multiple-choice questions about recent events, according to Minda Zetlin’s analysis of the study. The exercise focused on the three systems’ effectiveness as a news source, a role Zetlin says is growing as more people use AI to find news.
The source frames the choice around researchers who need to learn a new topic quickly or answer an obscure question. It identifies no test date, sample size, accuracy percentage, model version or question count, so it cannot support a ranking of the three systems.
What did Stanford actually test?
Zetlin, author of Career Self-Care: Find Your Happiness, Success, and Fulfillment at Work, said the strengths and weaknesses uncovered in the Stanford work could apply to any type of query. That broader application is her interpretation, not a separately reported conclusion from Stanford.
Tech journalist David Nield used the Stanford results, along with other recent research, to create a brief WIRED guide on when to search with AI and when not to. The source does not identify that other research or provide its dates, sample sizes or findings.
When does Zetlin recommend AI-assisted search?
Zetlin presents four practical use cases based on the Stanford study, Nield’s guide and her own long experience with online research. These are expert recommendations, not scored findings attributed to Stanford.
- General orientation and how-to guidance. Zetlin used sourdough as her example and wrote that “AI can answer a question like that really well.” She attributes easier digestion, a smaller glucose spike and better absorption of iron and other minerals to sourdough’s long fermentation process. She also says AI has given her useful gardening and cooking advice.
- Widely known facts. Nield uses the height of the Empire State Building as his example. Zetlin’s reason is straightforward: the height is accurately reported in many places, so the odds are high that an AI system will give the correct answer.
- Complex information in plain English. In Zetlin’s account, a friend diagnosed with prostate cancer entered test results and antigen levels into Google Gemini. The system gave a straightforward description of what the medical information meant. The friend had never used AI before.
- Queries with multiple parameters. A family of five sought dinner in San Francisco for less than $200. Zetlin says AI can also handle lead generation, such as finding associations that cover a speaker’s topic and schedule events six to 12 months ahead.
Her point is not that AI always wins. Zetlin writes: “Search engines can’t handle several elements at once in this way.” That is her explanation for using AI on these compound requests, not a measured result from the Stanford test.
What does the evidence not establish?
The source supplies no side-by-side accuracy scores for ChatGPT, Gemini and Claude. It does not provide the Stanford paper’s title, publication date, sample size, scoring method or model versions, so no product ranking or overall research-accuracy percentage can be derived from the item.
The sourdough, medical, dining and lead-generation examples are illustrations in Zetlin’s expert opinion. Nield’s reference to “other recent research” is a pointer, not enough information to assign additional findings to those unnamed studies.
The excerpt asks when not to use AI but does not provide a separate list of tasks to avoid. It supports affirmative examples for AI-assisted search; any broader rule about what researchers should never delegate to those systems would exceed the supplied evidence.
What should readers monitor next?
Monitor any fuller disclosure of Stanford’s methodology, including test dates, sample size, model versions and results by system. Also watch whether follow-up work tests the broader research tasks Zetlin infers from a recent-events quiz, rather than only multiple-choice questions about current affairs.
For readers choosing a tool now, the source supports AI for orientation, plain-language explanation and multi-part requests. Treat the missing Stanford metrics as the key data to wait for before declaring a system the better research engine.
via inc.com (Source)
More from James Calloway
Show full bio
Market editor covering money, funding and market shifts around Search rankings and algorithm updates, generative engine optimization at SERP Journal.
220 articles