That raises a fascinating point, which is whether search results that default to general topics ever are the basis for LLM training or information retrieval as a general phenomenon.
Yes, any human will most likely recognize the result as random noise, as they will know whom they are searching for, and see this not a video from or about Benn. But AI, taking all results as valid, will obviously struggle with this, condensing it to bullshit.
Thinking about, it's probably not even a real hallucination in the normal AI-meaning, but simply poor evaluation and handling of data. Gemini is likely evaluation the new data on the spot, trusting them blindly; and without any humans preselecting and writing the results, it's failing hard. Which is showing that there is no real thinking happening, only rearrangement of the given words.
Yes, any human will most likely recognize the result as random noise, as they will know whom they are searching for, and see this not a video from or about Benn. But AI, taking all results as valid, will obviously struggle with this, condensing it to bullshit.
Thinking about, it's probably not even a real hallucination in the normal AI-meaning, but simply poor evaluation and handling of data. Gemini is likely evaluation the new data on the spot, trusting them blindly; and without any humans preselecting and writing the results, it's failing hard. Which is showing that there is no real thinking happening, only rearrangement of the given words.