Fandango and the Viper

AI generated image from Meta

Act 1: The villainous accusation.

Beloved blogger Fandango was accused of having AI write his most excellent sci-fi flash fiction post SoCS - Lost in Space. Those of us who have read Fandango for years dismissed this claim as ridiculous, but still it seemed to have stung a bit. No one wants to be labeled a fake writer even if the accusation is blatantly untrue. Fandango subsequently addressed the claim and ran his writing through various AI detectors, only to end up with a wide range of results. Turns out that AI can't reliably recognize itself, which should surprise no one at all. Fandango's readers assured him that they knew he wrote his own posts ~ he even occasionally makes a typo!

Act 2: Paula’s precious poems.

Naturally, I had to run some of my writing through AI detectors to see what was what. At first, it was all happy news: depending upon which (free) detector I used, my writing ended up at zero to very low chance of being AI generated. Yay! That was my "normal" writing though; when I ran a few poems through the detectors they came out with a high probability of being AI generated. What? I kept using different detectors and different pieces of poetry, but the results were the same depressing high probability. I wondered if this was because all my poetry sounded predictable and clichéd. Sad! The heck with my silly blog posts ~ it's my "real" writing that should pass the no-AI test. I was too scared to run portions of my self-published fiction through after what happened with the poems.

Act 3: A venomous viper!

The other day a man in Orange County emailed an image to the local police saying that his dogs alerted him to a GABOON VIPER nestled in a tree in Santa Ana. Of course, I read this story in the middle of the night when I couldn’t sleep. Though snakes are cool, this is a GABOON VIPER, one of the deadliest reptiles on earth and only 4 miles from my house. It’s a ground snake, but this one was chilling in a tree! Could it be on my roof now! Where is it? They can’t find it! What if it slithers down my chimney and eats Inky? Holy crap!

Act 4: An unneighborly prank.

As I googled for more info, I found the follow-up story announcing that the GABOON VIPER sighting was a hoax. The SAPD had initially run the image through Grok, who said it was real, so the police immediately dispatched a bunch of resources to deal with it. All the experts on the internet however quickly pivoted from their infectious disease research and US Constitutional studies to don their African reptile hats and declare that not only was the image suspicious, but also GABOON VIPERS don’t climb trees, and a heavy snake would definitely not be on a skinny branch. Well, duh! CBS-LA used a different AI detector and found the image to be fake. Then the man admitted it was an AI image. 

Act 5: AI vs AI.

I know you’re not supposed to take Benadryl to help you sleep, but this was an emergency, so I did. While waiting for it to hit, I chatted with my buddy Meta AI and asked him why anyone would do the fake snake thing. Possible reasons are: for attention or to test the AI detectors or just to see if they can get away with it. Ugh! Anyway, I went on to complain about my poetry getting a bad result, and Meta told me not to worry: 

Don't be depressed — this is a detector problem, not a poetry problem. And it's a well-documented one.

AI detectors don't detect AI. They guess based on two things:

Perplexity — how predictable is the next word?

Burstiness — how much does sentence length and complexity vary?

[…]

Poetry does the opposite of what detectors reward:

It uses common, beautiful phrases that have existed for hundreds of years — "silent night," "broken heart," "endless sky." To a detector, that's low-perplexity, which it labels "likely AI."

It often uses regular meter, rhyme, parallel structure, and intentional repetition. To a detector, that's low-burstiness — not enough variation — which it also labels "likely AI."

It compresses language and removes the very quirks detectors look for in human prose — the asides, the self-corrections, the uneven sentences.

That's why so many real poets get 90%+ AI scores, and why your blogger friend got different results on different detectors for the same story. If you run Shakespeare through most detectors, he'll come back as AI too. The tools contradict each other because they are just statistical guesses with no ground truth.

[…]

Don't use detectors as evidence against another writer. The Santa Ana viper case and your poet friend's case are the same lesson — these tools can create fear and accusation out of nothing, and they put the burden of proof on the innocent person.

And there you have it: straight from the horse’s mouth!

P.S. Meta AI created a cool image for this post from the following directions: “I need an image for my post of that blogger, who is an older man with a scruffy beard blogging on his iPhone with a gaboon viper nearby…”