There is a website called FIle 770 which I have been following since they asked to quote from one of my posts. I hadn’t heard of them until then because I’m not a scence fiction “fan”.

I have been writing science fiction for 50 years, and reading it extensivley for 65, but I don’t discuss it with friends and I don’t join fan clubs. I even wrote a post about that once, so I never ran across File 770 until they ran across me.

File 770 covers everything in science fiction, fantasy, and horror, including books, television, and movies, with a big et cetera following that list. I have learned a lot from them. You might check them out.

The October 29th post quoted an excerpt from SF Concatenation, which quoted from an article in Nature titled Can scientists reliably tell if a novel is written using AI.

That got my attention.

Being a scholar (snoop, busy body, sceptic) I tracked the quotation backward and sideways. I need to tell you where I went so you can check things out for yourself, but that will be tedious, so I will place the URLs at the bottom of the post.

The Nature article was a bit of a mess. I can partly put this down to being an interview rather than a closely written scientific report, but the categories given in the description of the study are a hopeless jumble. Think apples and oranges, with a few peach pits thrown in. The original scientific paper was more reasonable, but the categories raised questions from the outset.

The main example of AI generated text was much more troubling. AI (typically the program Pangram) searching for AI generated material seems to boil down to looking for identical phrases which are “too long and specific to occur in exactly the same way by coincidence”. The example given was “They collapsed together, a tangle of sweat-slicked limbs and racing hearts.”

What!  You’ve got to be kidding me.

Writing about sex in romantic novels calls for indirection mixed with enthusiasm. You can’t name body parts and describe what they are doing, but you must be able to bring your reader to a state of sexual excitement. That isn’t easy, and the turgid text which often results sounds exactly like the example provided. The author of the study is surprised to find this identical sequence is two separate books. I would be surprised if you didn’t find that sequence a dozen times, given a large number of romance novels to study.

But that isn’t the main problem with equating long identical word sequences with AI. If such identical sequences are found, what does it really mean?

1.  It might mean that two different authors had independently created the sequences, no matter how unlikely.

2.  It might mean that it was out and out plagarism.

3.  I might mean that the sequence was unconsciously remembered and reused.

4.  It might be deliberate, unreferenced quotation, perhaps an homage, perhaps a quotation thought too insignificant to note.

5. Or, it might be AI.

Simply assuming that it is AI is indefensible.

If a dozen or more such sequences were found in a single book, it still would not indicate AI. It would more likely be a dishonest human writer.

I will suggest an experiment that would deal with the question of AI hunting down AI. Take a few thousand novels from the twentieth century (or the nineteenth) and subject them to Pangram. See how many pop up as AI generated. I’m taking bets that it would “find” hundreds, a hundred years before AI was invented.

For even more negatives on Pangram, check out the Atlantic article below.

* * * * *

The original File 770 can be found at (https://file770.com/pixel-scroll-10-28-26-i-find-your-lack-of-pixels-disturbing/). No, you aren’t reading that wrong, Mike Glyer likes goofy titles.

From File 770, I went to the Nature article, available at (https://www.nature.com/articles/d41586-026-03394-4)

The Nature article carried me back to a précis of the original scientific study at (https://arxiv.org/abs/2607.20349). Once there, I clicked View PDF and got the whole thing.

Along the way I also found an even more interesting article in the Atlantic at (https://www.theatlantic.com/technology/2026/05/pangram-ai-detection-accuracy/687381/)