AI unable to identify someone and prefer give a false answer

macadoum

I asked in french "who is he ?" The assistant gives me a false answer, saying the guy is Donald Trump just because of the name of the screenshot I guess. The name of my screenshot is "Screenshot 2026-02-25 at 15-57-47 Donald Trump delivers State of the Union address - YouTube.png"

While re-asking in English, the assistant gives a better answer : I don't know who this guy is.

I expected a better answer in french. If The Assistant do not recognize someone on a picture, do not give a false answer based on the name of the file.

carl

It tells you right there in your screenshot that it searched for "man in suit red and blue tie state of the union address".

Unfortunately, the AI seems to have mixed up with Trump because he wore a blue suit with a red tie.

k7absp2h

carl I think we need a reverse image search for Kagi Assistant.

User uploads an image.
Kagi reverse-searches that image.
Kagi could add related URLs (as context).
Thanks to the context, the model’s accuracy will increase dramatically.

macadoum

carl Yes but as said in my post, when I asked in english, the AI admit it doesn't recognize the man on the picture.

carl

k7absp2h

I don't think reverse image search would be of any help in this case, since the image is a screenshot the OP took from a YouTube video.

The question asked is simply impossible for an AI to answer, they would have to be trained on a massive database of peoples facial data and store that in order to be able to recognize faces of people who aren't famous or infamous. That would be a disaster for mankind.

Or - in the best case - the AI would be smart enough to watch the entire YouTube video and see if the man in the picture is introduced at any point of the video.

k7absp2h

macadoum still giving reverse search context would be great

carl

macadoum Absolutely, no answer is better than a wrong answer. What I meant is that I understand why the AI confused itself in your example.

k7absp2h

carl

I don't think reverse image search would be of any help in this case, since the image is a screenshot the OP took from a YouTube video.

You can try it yourself; when you provide reverse-search links, vision model's accuracy increases dramatically.

The question asked is simply impossible for an AI to answer; they would have to be trained on a massive database of people’s facial data.

They already did 😶 (see Google Lens, Bing Visual Search, Yandex Image Search).

And if Kagi Assistant used reverse image search, it would definitely answer correctly.

Richard Blumenral

carl

k7absp2h Look at that! I'm behind on the times.