No. of Recommendations: 18
I'm going to jump in here with something that I raised on the Fool, but is pertinent here.
People (including myself, formerly) seem to vastly overestimate the ability of AI tools to speak intelligently on current events.
After having been frustrated with some responses on Gemini a few times, I decided to dig under the hood and see what was causing the problems. So here's what you need to know:
- For nearly everything, Gemini's core parametric training data is only current up through January of 2025. The core of Gemini knows absolutely nothing about anything that's happened in the last two years.
- If you ask Gemini a question that requires knowledge of anything in the last two years, it will just a search engine (unsurprisingly Google) to look at the web.
- But the use of the search engine is extremely constrained. It will only run one or two searches. It will only look at about ten of the results. And it won't even read those ten results in full - it will only look at snippets of those ten results.
- It will then return an answer.
For shirts and goggles, I entered Jeff's initial query into Gemini. But I also asked it to give me an audit of how in-depth the research was that it conducted on the response. Warning to sensitive readers - the results may shock you.
Audit of Search Activity
Web Searches Conducted: 2
China US autonomous driving technology comparison 2026
China vs US robotaxi autonomous vehicle progress leadership 2026
Total Hits Returned: 10 (5 snippets per search query)
Hits Examined: 10 individual result entries across tech publications, industry analysis reports, and automotive deep dives.
Text Evaluated: Snippets only. The underlying context window processed short extracted search summaries (~150–400 words per result) rather than opening or reading the full HTML/webpages.
So....ummmm....yeah, don't rely on Gemini about anything where the answer depends on knowing anything that's happened in the last two years. It's going to give you a response that is based on the human equivalent of maybe 10-15 minutes of web searching, and the answer is going to be overwhelmingly dependent on which ten articles Google happened to return to the web search.