One of my favorite books of all time is the Book of Lists. It contained weird and interesting lists. Here is a list of a dozen of these lists, i.e. a list of lists:   

1. 10 Famous Insomniacs
2. 10 Really Unusual Medical Conditions
3. 9 Cats Who Traveled Long Distances to Return Home
4. 15 Famous People Who Worked in Bed
5. 11 Most Unusual Objects Sold on eBay
6. 10 Celebrated People Who Read Their Own Obituaries
7. 16 Famous Events That Happened in the Bathtub
8. The World’s Greatest Libel Suits
9. People Suspected of Being Jack the Ripper
10. Worst Places to Hitchhike
11. Dr. Demento’s 10 Worst Song Titles of All Time
12. Breeds of Dogs Which Bite People the Most  

So, I thought I would offer a column dedicated to a list related to flubbed headlines, common sense and AI.

Flubbed Headlines

Flubbed headlines are fun. There are typically two meanings. One is the obvious meaning, meant by the writer of the headline. The second is an interpretation, which can be hilarious. Both interpretations are semantically correct. Humans can often differentiate between which interpretation is serious and which is silly. This requires common sense.

In a 2020 post, I challenged AI to resolve the ambiguity in flubbed headlines. This is before large language models (LLMs) like Grok, ChatGPT, MidJourney, Perplexity, and Claude became popular.

Here is an example of a flubbed headline: “Helicopter Powered by Human Flies.” Does this mean that a human being provides the power to fly the helicopter? Or does it mean that flies resembling humans are powering the helicopter? Clearly, the first interpretation is correct. The question, then, is whether large language models can reliably make this distinction.

To Serve Man

Perhaps the most famous example of a flubbed headline in TV comes from the 1962 Twilight Zone episode “To Serve Man.” In the story, aliens called Kanamits arrive on Earth and appear benevolent. They help humanity to cure diseases, increase crop yields, end war, and improve the overall quality of life. They bring with them a book titled To Serve Man, which is initially translated without sufficient understanding of the contents. Humans naturally interpret the title to mean that the aliens intend to help mankind.

Eventually, the aliens invite thousands of people aboard a large spaceship, promoting the trip as an extraordinary vacation and a chance to experience outer space firsthand. Only after the ship departs does the full translation become clear: To Serve Man is not a guide to serving humanity at all. It is a cookbook. The book contains recipes for preparing and consuming human beings as food. What seemed like a message of goodwill turns out to be a chilling example of deadly semantic ambiguity.

  Flubbed Headline List

LLMs have met my challenge in resolving flubbed headlines.

Okay, here’s the list of 125 flubbed headlines. I suggest you read just a few and then skip ahead. Like eating candy from a box of chocolates, the novelty wears off about the fifth or sixth piece. Return to the list when you get hungry again.

1. New Housing for Elderly Not Yet Dead
2. Shouting Match Ends Teacher’s Hearing
3. Dr. Gonzalez Gives Talk on Moon
4. Man Seeking Help for Dog Charged with DUI
5. Navy SEALS Responsible for Getting Osama Bin Laden to Be Honored at Museum
6. General Who Ran Vietnam Briefly Dies at 86
7. Drought Turns Coyotes to Watermelon
8. Living Together Linked to Divorce
9. Ban on Nude Dancing on Governor’s Desk
10. Town Okays Animal Rule
11. Deer Kill 130,000
12. Helicopter Powered by Human Flies
13. Stiff Opposition Expected to Casketless Funeral Plan
14. Drunk Gets Nine Months in Violin Case
15. Bar Trying to Help Alcoholic Lawyers
16. Family Catches Fire Just in Time, Chief Claims
17. Legalized Outhouses Aired by Legislature
18. Milk Drinkers Return to Powder
19. Two Convicts Evade Noose: Jury Hung
20. Farmer Bill Dies in House
21. Antique Stripper to Demonstrate Wares at Stores
22. US, China Near Pact on Wider Ties
23. Police Squad Helps Dog Bite Victim
24. Red Tape Holds Up New Bridge
25. Farmer Looking for Life With Tractor
26. Police Begin Campaign to Run Down Jaywalkers
27. Iraqi Head Seeks Arms
28. Stolen Painting Found by Tree
29. Two Sisters Reunited After 18 Years at Checkout Counter
30. Kids Make Nutritious Snacks
31. New Vaccine May Contain Rabies *
32. Man Struck by Lightning Faces Battery Charge
33. Students Cook and Serve Grandparents *
34. Utah Girl Does Well in Dog Show
35. Massachusetts Woman Has Eye On Kerry’s Seat
36. Local High School Dropouts Cut in Half
37. Death Causes Loneliness, Feelings of Isolation
38. Legislatures Tax Brains to Cut Deficit *
39. Miners Refuse To Work After Death *
40. Meat Head Resigns *
41. Lawyers Give Poor Free Legal Advice
42. Enraged Cow Injures Farmer With Axe
43. Panda Mating Fails; Veterinarian Takes Over
44. Eye Drops Off Shelf
45. Montana Traded to Kansas City
46. Engine Falls Off Plane, Lands Safely at O’Hare
47. Trees Can Break Wind
48. Lack of Brains Hinders Research
49. Missouri Woman Big Winner at Hog Show
50. Tuna Biting Off Washington Coast
51. Astronaut Welcomes Baby from Space
52. Road Rage Leads Police to Murder Victims Boyfriend
53. Man Eating Piranha Mistakenly Sold as Pet Fish
54. TV Networks Agree to Police Violence
55. Lansing Residents Can Drop Off Trees
56. Sewer District Plans Emergency Backup
57. The solution to huntings woes? Sight set on women.
58. Deer and Turkey Hunt for Disabled People
59. Doctor Testifies In Horse Suit
60. Lung Cancer in Women Mushrooms
61. New Foreclosure Listings: Entire State of New Jersey Available
62. Massachusetts Woman Has Eye On Kerry’s Seat
63. Music Industry Meets On Drugs
64. Army Suicides Expected to Jump
65. Condom Firm Stretches Product Line
66. New Study of Obesity Looks for Larger Test Group
67. Passengers Feeling Airline Crew Cuts
68. Policeman Shoots Man With Knife
69. Real Estate Executive Sold on City Market
70. School Bans All Kinds Of Nuts On Campus
71. Clinton Wins on Budget, But More Lies Ahead
72. TV ads boost eating of obese children by 130%
73. Children living without limbs lack support
74. Autos killing 110 a day. Let’s resolve to do better
75. Nuns forgive break in, assault suspect
76. British Left Waffles On House Floor
77. UAW Elects King as Next President
78. Big Ben Celebrates 150 Years of Bongs
79. Homeless Man Under House Arrest
80. Death in the ring: Most boxers are not the same afterwards
81. Tight End Retires After Colon Surgery
82. Wanted: Plain Clothes Security. Must have shoplifting experience
83. Arson Suspect is Held in Fire
84. How to prepare pets for Thanksgiving
85. L.A. Voters Approve Urban Renewal by Landslide
86. Crack secret to a perfect appetizer
87. Marajuana issue sent to joint committee
88. Bonus Permits Enable 809 Hunters to Kill Two Deer
89. Judges Appear More Lenient on Crack Cocaine
90. Garbage Truck Lands on Saturn
91. Wimbledon: Federer Still Riding High on Grass
92. 11 High Students Score Perfect Grade
93. Rangers’ Hamilton to Get Shot for Sore Knee
94. Ukraine First Lady Appeals for More Arms
95. Police Car Hits Cow in Line of Duty
96. The Fresno Nutritional Home Served Sick Children
97. Academics to Dissect Bob Dylan at NY Conference
98. Climber Who Cut Off Arm to Escape Speaking at MSU
99. One Armed Man Applauds the Kindness of Strangers *
100. Blind Kids on the Brink of Being Shown the Door
101. Community rallies to help massacre survivors
102. Cuts hurt patients, nurses say
103. Lebanon Will Try Bombing Suspects
104. CRIME: Sheriff asks for 13.7% increase
105. California Governor Makes Stand on Dirty Toilets
106. Nuns Forgive Break-In, Assault Suspects
107. Wanted: A hangman that knows the ropes
108. Madonna reads her 2nd book
109. Store clerk better after being shot
110. Workshop Is Scheduled on Manure Management
111. Old School Pillars are Replaced by Alumni
112. Doctors Offer Suicide Guide
113. 19 Feet Broken in Pole Vault
114. Cuts Could Hurt Animals
115. Kicking Baby Considered to be Healthy
116. Officials Warn Clams, Oysters Can Carry Virus
117. Kids Must Be Shot by Monday
118. Cops gets burglar to drop his pants, then pinch him
119. High Speed Train Could Reach Valley Within Five Years
120. Colleen Campbell champions the rights of murder victims after being one herself
121. Minus shorts, bankers get breathing room
122. Seafood Bought in by Customers
123. Some 70,000 turtle eggs to be whisked far from oil
124. Prostitutes Appeal to Pope.
125. Hospitals are Sued by Seven Foot Doctors

How Did AI Do?

My students and I used ChatGPT to evaluate the interpretation of the flubbed headlines on the list. An asterisk indicates the LLM got the wrong interpretation or the response was inconclusive. For the remainder of the flubbed headlines, the right interpretation was given or the existence of an ambiguity was identified.

 If you want to look at the raw data, here’s the LINK.

How did ChatGPT do in resolving flubbed headlines?

When asked “What does <phrase > mean? ” there were 116 correct answers out of 125. That’s 93% accurate. For example, when asked “What does `Drought Turns Coyotes to Watermelon’ mean? “, the response was:

Watermelons are made up of about 90% water, making them a potential source of hydration. If coyotes are eating watermelons, it could be indicative of just how dire the drought situation is.

The alternate silly interpretation is that coyotes somehow morphed into watermelons.

Often, the ambiguity in the headline was identified. For example, “What does `Deer Kill 130,000’  mean?” returned the response:

“The headline “Deer Kill 130,000” can be interpreted in a couple of ways, depending on the context:

Deer Hunters’ Success: It could mean that hunters have killed 130,000 deer, possibly within a specific hunting season or geographic area. This would likely be part of wildlife management efforts or hunting reports.

Deer-Related Incidents: It could also refer to 130,000 incidents involving deer that resulted in deaths, most likely from deer-vehicle collisions. This interpretation would imply a high number of accidents where deer have caused fatalities, typically among the deer population but potentially including human fatalities as well.

On rare occasions, the response was either non-responsive to the query or was simply wrong. Here’s an example:  “What does `Legislatures Tax Brains to Cut Deficit’ mean?” generated the response:

The phrase “Legislatures Tax Brains to Cut Deficit” is a figurative expression that suggests a government or legislative body is imposing taxes or levies on its citizens’ incomes or earnings in an attempt to reduce a budget deficit.

There is no indication or implication in the query that income is to be taxed. One way to decrease the deficit would be to decrease spending.  Maybe ChatGPT never learned an alternative. Maybe it was only trained on tax-and-spend practices popular to politicians.

Why the Success?

I first made my challenge to AI to resolve flubbed headlines in 2020.   ChatGPT-4 was released in 2024. LLMs like ChatGPT-4 are a special case of transformers  introduced in the  groundbreaking 2017 paper entitled “Attention Is All You Need”. 

LLMs resolve flubbed headlines because they have absorbed vast statistical patterns from human language that encode typical meanings and real-world expectations. When faced with ambiguity, they tend to select the interpretation that is most plausible and frequently observed in their training data. This allows them to mimic (not duplicate) common sense, even though they do not truly understand the underlying concepts or causal relationships.

Homework

Winograd Schema are similar in ambiguity to flubbed headlines. A classic example is: “I can’t cut down that tree with this axe. It is too small.” Here, the pronoun it is ambiguous and could refer either to the axe or to the tree. Common sense, however, tells us that the axe is too small. Much like flubbed headlines, today’s LLMs perform remarkably well on them.

Here is a LINK to 150 Winograd schemas you can try with your favorite LLM.