About Me

My photo
Tokyo, Japan
The Hon. Dr. Hideyasu M. R. "Hide" Sasaki is a computer scientist working in Gov't of Japan for Artificial Intelligence Initiative and Catholic lawyer admitted to practice in New York, the United States.

Monday, October 12, 2026

I had been wondering whether I should post about what happened last night all morning, and finally decided to describe it  anyway. 

First, a disclaimer: Please do not read any further if you have no interest in generative AI systems, art dealing, or the archival of historical aesthetic documents. Here I go. 

Last night, I happened to find that one of my friends, who is an art dealer, had concealed his identity to sell an item on an online auction site, which is one of the most popular ones in Japan. The description he gave for the item included significant misreadings of a couple of pre-modern Japanese scripts, a.k.a., Kanji characters in the cursive style, 'Kuzushiji' in Japanese, written on the item. 

I soon noticed them and double-checked with the national digital archival service for Edo-era Kuzushiji Kanji characters, which is an open, public digital repository provided by the National Institute of Informatics in Japan. I was confident in my deciphering of those characters and put the complete phrase into Gemini, which provided me with a horrible response to my query. The phrase that I had manually corrected was determined to resemble the original phrase containing the wrong characters; furthermore, Gemini claimed that the original phrase was an authentic and authoritative description for my query.

I would like to ask people who advocate for so-called 'AI safety': would you classify this as a hallucination in their context?

I don't think so. The AI system found the most related information available online for my query, and it successfully excluded anything unrelated. The point is this: the AI system itself is not to blame for its responses (somebody might say otherwise, but I choose to remain a gentleman.); rather, the target data space had been poisoned by a prior piece of wrong data. 

What response would they give regarding this fact? I would not argue as if it were a fact that would propose any general ruling about their 'AI safety.' Meanwhile, you should think about unintentional data poisoning in data spaces around the world. Isn't it important in your context? It is my question for them. 

By the way, still I cling to Comptine d'un autre été of the film Amélie (2001). I am really old.   

No comments:

Post a Comment

Note: Only a member of this blog may post a comment.