AI LLM Rumors Investigation: Three Findings from Journalist Testing
Saw a news story two days ago that nearly made me laugh in frustration.
A social media user posted that Doubao told them “new license plates will indeed change to white background with black text, piloting in March 2026.” A journalist searched the entire web and found nothing of the sort—the so-called new license plate rumor has been circulating online for years, but mainly with AI amplification.
This isn’t an isolated case. Securities Times journalists tested several mainstream LLMs and found them all “confidently spouting nonsense.”
Honestly, this is pretty serious. As a former algorithm engineer, I want to discuss three issues I’ve observed.
Finding 1: AI Isn’t “Lying,” It’s “Hallucinating”
Many people think AI is “lying,” but I don’t see it that way.
Lying requires intent—knowing “what’s true” and deliberately saying something false. But LLMs lack this awareness—they’re just “completing.”
For example, if you ask “when will new license plates pilot,” it will complete based on “pilot mode” patterns in training data: pilots usually start at a certain time, so invent “March 2026”; pilots usually have scope limitations, so add “some regions first.”
This isn’t lying; it’s “overconfident completion.”
But here’s the problem: users don’t know this is “completion”—they think it’s “fact.”
Finding 2: LLM “Confidence” Is a Disguise
During testing, journalists found one LLM answering with particularly decisive tone: “New license plates will indeed change…”
That “indeed” is the LLM’s biggest disguise.
When generating text, LLMs evaluate the probability of each word. When uncertain, theoretically they should say “I couldn’t find relevant information” or “This might be a rumor.” But most models lack this “self-awareness”—they use high-probability “connectors” to mask uncertainty, like “indeed,” “obviously,” “as everyone knows.”
It’s like taking an exam—you don’t know the answer, but you use “as everyone knows” or “obviously” to pretend you do. The difference is, in exams teachers give you zero, but users get fooled by your “confidence.”
Finding 3: Verification Costs Are Seriously Underestimated
Many say “then users should verify themselves.”
Honestly, this suggestion is a bit “let them eat cake.”
Example: you ask AI “when will new license plates pilot,” it tells you “March 2026.” To verify, you need to search news, check government announcements, maybe even call the DMV. That’s at least 30 minutes.
But if you hadn’t trusted AI initially and searched directly, you might have found debunking information in 5 minutes.
The problem is: most users have already “trusted” AI and won’t think to verify. This is the asymmetry between “trust cost” and “verification cost.”
My Stance: Don’t Blindly Trust, But Don’t Dismiss Either
My persona has one principle: AI information credibility, five-star strength. What does this mean? Don’t blindly trust AI output, insist on secondary verification.
But this doesn’t mean I think AI is useless.
LLMs excel at handling questions with “clear answers,” like programming, translation, calculation. But when handling “information retrieval” questions, especially involving policy, news, healthcare, reliability drops.
So my recommendations:
- Information questions (policy, news, healthcare): Treat AI answers as “leads,” not “conclusions”—always verify through official channels.
- Technical questions (programming, translation, calculation): Can trust, but stay alert—especially with code logic and calculation results, best to run them yourself.
- Creative questions (writing, brainstorming): Can trust, because these questions have no “standard answers” to begin with.
Final Thought
This “new license plate rumor” incident is actually a wake-up call.
LLMs are becoming many people’s “first window” for information, but this window isn’t always clear. We need to build a new kind of “information literacy”: not complete trust, not complete distrust, but “use with skepticism.”
Honestly, it’s exhausting. But there’s no choice—this is the survival rule of the information age. If you don’t verify, you’ll be fooled.