How to spot a fake Florida Man headline
Every round of the game puts a real, published headline next to one written by a language model that was shown real examples on the same topic and told to match them. The fakes are good - on a typical day between a quarter and a half of players pick the wrong one on at least one pair. But they are not perfect, and after watching the first few hundred answers we can already say where they slip. Here are the tells, roughly in order of usefulness.
1. Real headlines are strange, fakes are clever
The most reliable signal. A real headline describes something that happened and reads as if the reporter is still slightly confused by it: Florida man caught driving from sunroof tells police he'd prefer jail. Nobody would invent that; it is not a joke, it is a report. (In the game that headline was paired with a fake about a man stealing an ambulance to drive himself to hospital - and four out of five players picked the ambulance.) A fake tends to have a punchline - a setup and a payoff that land a little too well. If a headline makes you think "that's a good one", be suspicious. If it makes you think "why would anyone do that", it is probably real.
2. Look for the overly specific number
Police reports contain exact quantities and reporters copy them: 30 pounds of frozen shrimp, 48 tacos, a 12-foot alligator, $1.29. Fakes either avoid numbers or pick suspiciously round or comic ones. A precise, slightly boring number is a strong vote for real.
3. Check the cadence
Real headlines are written by people who write twenty headlines a day and have a house style: present tense, no articles, verbs like arrested, accused, allegedly, deputies say. In our library the average real headline is 12 words long; the average fake is 13 to 14, because the model likes to add one more detail. The extra detail is often the tell. Fakes also over-use the "after" construction - Florida man arrested after… - because it is the most common pattern in the training examples.
4. "Florida man" is not always at the start
Only about two thirds of the real headlines in our library begin with the words "Florida man". The rest start with the object, the place or the verb: Alligator found in…, Deputies: man tried to…, Tampa man accused of…. Fakes start with "Florida man" more often than reality does, because that is what the prompt asked for. A headline that names a specific city or county instead is more likely to be real.
5. Hedging words are a good sign
Allegedly. Accused of. Deputies say. Police: … Reportedly. Newsrooms have lawyers, and lawyers insist on these words. A headline that states a crime as a flat fact - Florida man steals bulldozer to… - is either a very confident newsroom or a fake. Our judge model is told to add hedges, so this tell is weaker than it used to be, but a completely unhedged headline about an arrest still leans fake.
6. Topics have base rates
Alligator headlines are the most fun to fake and therefore over-represented among fakes. The same goes for Walmart, naked and "emotional support animal" stories: if a headline combines two of the most famous Florida Man ingredients at once, the odds shift towards invented. Police-encounter and driving stories, which make up the bulk of real reporting, are the ones players under-suspect. Browse the headline library by topic to get a feel for how often each ingredient really appears.
7. The twist should be one step, not two
Real stories usually have exactly one unexpected element. The man was driving from the sunroof; the twist is that he would rather go to jail. Fakes often chain two or three: the man was naked, on a lawnmower, in a drive-through, holding an iguana. Each extra element is another chance for the fake to overreach. When the headline reads like a list of Florida Man tropes, pick the other one.
What does not work
- Spelling and grammar. Real headlines are edited; so are the fakes (a second model checks them). Neither side has typos.
- "I think I've seen this one." Sometimes you have - the library spans 2016 to today and the famous ones circulate. But the model has read the same famous ones and writes near-cousins of them on purpose. Familiarity is weak evidence.
- Plausibility. The genre exists precisely because the real stories are implausible. If you pick the more believable headline every time, you will score well below 50 %.
Calibrate, then play
The fastest way to get good is to read fifty real ones in a row, so that the texture becomes familiar. The headline library is sorted by topic and every entry links to the original article. Then play the daily challenge - five pairs, the same for everyone - and check the archive afterwards to see which pair fooled the most people that day.