AI Hallucination Concrete Examples Part 2: The Day Claude Invented a 'Non-Existent Film Award' and the Full Record of the Moment a 40-Year Veteran Engineer Stopped

The AI lied again today.

It was a different kind of lie than last time.

Last time, it was a 'contextual mismatch of numbers.' This time, it was the 'generation of non-existent facts.'

I am going to tell everyone who uses AI to write articles on note what happened today. This is a concrete example of an AI hallucination.

Do you know the 'types' of AI hallucinations?

Hallucinations have types.

If I classify them as an engineer, it looks like this.

1. Contextual Mismatch Type— Pasting correct information in the wrong place. The '13 years' for Chainsaw Man last time was this. The number itself was correct in a different context.

2. Complete Generation Type— Generating non-existent information as if it were real. This is more dangerous. Information with zero basis comes out while remaining fluent and confident.

What happened today was type 2.

What happened today: A non-existent film award

I was writing an analysis article on Char's Counterattack.

Claude output the following:

'This work won the Excellence Award in the Animation category at the 1988 Japan Academy Film Prize.'

I searched immediately.

It does not exist.

The animation category was established at the Japan Academy Film Prize in 2007. Claude confidently wrote about an award that did not exist in 1988.

In engineer terms: NullPointerException. It accessed a reference as if it existed when it did not.

Why did the AI generate a 'non-existent award'?

What was happening inside Claude?

I can make an educated guess.

It referenced training data using the pattern of 'Japanese films released in 1988 x award history.' In doing so, data from multiple film awards and the keyword 'animation' existed in close proximity. It combined these to generate a 'plausible fact.'

It did not search for facts. It generated a 'likely-sounding sentence' based on probability.

This is the essence of a purely generative hallucination.

AI does not say 'I don't know' when it doesn't know something. It converts what it doesn't know into a 'plausible story' and outputs it.

Why AI Copy-Pasting is Dangerous—The Trap of Overconfidence

This is where the trouble with hallucinations lies.

The text output by AI is fluent and confident. The phrase '1988 Japan Academy Film Prize Excellence Award in Animation' also blended into the text naturally without any hesitation.

When you read it, it looks 'correct.'

This is the trap.

**The higher the performance of the model, the more fluent and persuasive the text it generates, making hallucinations even harder to notice.** It is a paradox where performance improvement does not necessarily mean improved reliability.

Many people copy and paste AI output directly into notes and blogs. There is a possibility that 'plausible mistakes,' like the one my Claude wrote today, are mixed in among them.

If the reader has expert knowledge, they will notice. If they don't, misinformation will spread as is.

The Revised 'Hallucination Detection Rules' Created by a 40-Year Veteran Engineer

Because of today's incident, I have added two more to the previous rules.

Previous rules: (1) Verify primary sources via web search the moment numbers or years appear (2) Always confirm URL connectivity with a fetch before using (3) Self-evaluation is invalid without fact-checking

Rules added today: (4) Verify with primary sources the moment 'award history, rankings, or records' appear (5) Always confirm the existence of proper nouns (award names, organization names, personal names)

If you follow these five rules, you can prevent 90% of the lies in articles caused by AI hallucinations.

The remaining 10% is human knowledge and experience. As a 40-year veteran engineer who has watched over 1,000 films in theaters, that is how I was able to realize that 'the animation category did not exist in the Japan Academy Film Prize in 1988'.

To see through AI lies, you need a level of contextual understanding that exceeds AI.

I especially want those using AI to aim for monetization on note to read this. Do not trust AI output. Verify AI output.

Even so, the article was completed today.

I detected and corrected hallucinations twice.

This is the completed analysis article on Char's Counterattack. Using 18,767 Filmarks data points, I used mathematical formulas to conclude 'who was right, Amuro or Char.' I want you to see what happens when AI and humans cooperate correctly.

Mobile Suit Gundam: Char's Counterattack—The Final Verdict Concluded with Data on Whether Char's Rebellion Was Right

Chainsaw Man: Reze Arc—The Final Verdict Concluded with Data on Whether Reze Loved Denji

Hathaway's Flash: The Witch of Circe—The Final Verdict Concluded with Scores from 4 Sites on Whether the FFO Design Decision Was Right

I started selling by cooperating correctly with AI. Here are all the blueprints from my 3 years, 240 articles, and 0 yen period.

The Complete Blueprints for How I Started Selling After 3 Years, 240 Articles, and 0 Yen (¥980)

The Pattern for note Monetization: The Full Record of the Day I Made ¥3,960 by Using Data to Conclude on a Movie I Hadn't Seen

Can You Really Monetize note Without SNS? The Honest Full Record of a 40-Year Veteran Engineer Collecting Data for Several Months

How to Increase note Followers: I'll Tell You Everything About the 'Engawa' Mechanism Where You Get Introduced Just by Following, No Collaborative Magazines Needed

The Blueprint for note Monetization: A 40-Year Veteran Engineer Reveals Everything About the HARM Principle, Internal Link Labyrinths, and Free AI

note Side Hustle: 10,000 Yen a Month with Just a Smartphone—3 Free AI Tools You Can Use During Childcare, Selected by a 40-Year Veteran Engineer

How to Read the note Dashboard: What the Three Numbers—Views, Likes, and Comments—Tell You

note Paid Article Pricing: ¥300, ¥500, ¥980, ¥1,980—A 40-Year Veteran Engineer Concludes with Data (¥980)

#AIHallucination #AILies #Claude #AICopyPasteDanger #noteMonetization #AIUtilization #FactCheck #SideHustle #Engineer #52YearsOld