I minus the A ?
I know skepticism about AI and "2+2=5" style AI stories abound and I remain a "fan with tons of huge doubts". But I have to say this little experiment is making it harder. Here's what happened:
We have a thing where you can type in someone's name and get back any recent news about them that might influence your willingness to do business with them. For example type "Hank Green Vancouver crime stories since 2025" and get back a link to a CBC story about Hank having robbed a bank. Maybe you decide not to hire him after all.
Systems to do these things exist and we offer it via a third party supplier, who does a fairly bad job. So we were wondering if AI could do better.
I was pretty surprised that Gemini, Perplexity and Deepseek all consistently came up with totally fabricated stories. I was even more surprised at Gemini's explanation:
"Gemini is a "next-word predictor." If it doesn't find a news story in its training data, it doesn't say "I don't know." It simply predicts what a news story about that person might look like."
"When an AI generates a URL, it isn't "finding" a link; it's imagining what a URL for that story should look like (e.g., https://bbc.com/news/person-name-event). These links almost never exist in reality."
"You need to tell the model that its job is to be a librarian, not a storyteller. Use a System Instruction like this: You are a factual news aggregator. Use ONLY the provided search results. If no recent news exists for the person, state 'No recent news found.' DO NOT invent stories. If you cannot provide a verified URL, do not provide a link at all."
Yeah, I knew AI is fundamentally a next-word predictor but I had been operating under the assumption that it predicted the next word based on FACTS, not what a response "might look like".
Comments
Post a Comment