Smell test

I’ve begun to realize how unreliable Gemini’s deep research has become. Or perhaps it always was. This is especially noticeable for questions that don’t have simple answers that aren’t easy to source. The results seem impressive, but compared to Claude’s research reports, Gemini is basically pulling data out of a hat.

• It restates a lot of my own observations as if it found supporting evidence. Sometimes it confabulates what I said, adding details I never said to make the ideas more concrete.
• Citations don’t always support what was said. Some citations are given more weight than they deserve. It makes up equations not found in any of its citations. And it doesn’t distinguish between trustworthy sources and marketing.
• Reports will be padded with extra material that I never asked for, seemingly to hit a length target.
• It’s overconfident. Citations are overused. And it gives too much credibility to sources that are pure speculation.

This is why it helps to have humans in the loop. As a human user of AI, I have to keep performing the basic smell test.

And then I can pass the output off to another AI and ask, “Does this smell right to you?”