When Gemini 3 Flash doesn't know the answer, he just makes it up.

But, according to an assessment made by an independent test group, if you ask them something that he doesn't really know about, hard or out of his knowledge, he'll almost always try to respond by lying or inventing something.

In the tests of “the degree of hallucinations” (recognition rate) at the Benchmark entry AA @Omniscience, Gemini 3 Flash reached a 91 percent fush, which means that even when there was no correct answer, he answered anyway, and often it was completely invented.

This phenomenon of “is a known problem in text generation patterns: to know when to stop and say “does not know” is as important as knowing how to answer. According to this test, Gemini doesn't do this very well, reports Telegraph, broadcast Periscope.

This does not mean, however, that 91 percent of his answers are wrong. This figure just shows how often he invents something in situations when the real answer would be “doesn't know”.

Even though Gemini 3 Flash can be very powerful and perform well on general tests, he is very self - confident even when he should be careful that it may be a problem in serious use. /Periscope