Can AI spot AI-written stories?

Why interpretation and mechanism are more powerful than simple pattern matching

If you follow writerly people on Substack, you may well have seen discussion of a recent prize-winning short story that appeared in Granta magazine. Specifically, there has been speculation that the story was written with the help of AI.

In response, Granta issued a statement, which included the following assessment:

We showed Claude.ai the story and asked whether it was AI-generated. The response was long, concluding that it was ‘almost certainly not produced unaided by a human’.

Several people criticised this approach. Large language models like Claude are just token predictors, went the argument. They just generate text. What you really want is a proper machine learning classifier that has been trained on AI and non-AI writing.

Except this is an over-simplified picture of current LLM abilities. For example, take a look at the following three story openings. Can you guess which one is AI?

1. The lock on the top drawer wasn’t really a lock. It was more like a habit that had grown teeth. I tugged at it anyway, because adults leave their important things in the places that pretend to be safe. The journal was under a stack of rubber bands and a cookbook with a cracked spine. Black cover, no title, corners worn down like it had been carried in a pocket too often. The elastic around it held but didn’t strain, like it was waiting to be noticed.

2. By the time Lena reached Gate 42, the plane had unlatched itself from the building and was nosing backward in the rain, serene as a swan that had never cared for bread. The woman at the desk said in a voice trained not to absorb grief. Lena nodded because she had no breath left for argument. Her suitcase, somehow, had gone ahead without her.

3. He’d once punched a shark, they said. Square in the jaw. Of course, they said a lot of things about Wilson Harton back then, before the police arrived. I once heard that he had an Olympic medal, and that he’d almost married that woman who does the news. Mrs Jenkins even said he could speak Portuguese. Couldn’t speak Portuguese herself, could Mrs Jenkins, so I’d put that one down in the maybe pile.

Have you made your judgement?

Before I give the answer, this was the Claude assessment: ‘1 is suspicious, 2 is very likely AI, 3 reads as human.’

And here are some of the reasons given by Claude (running Sonnet 4.6):

1. The prose is technically accomplished but over-engineered — every sentence reaches for a metaphor.

2. “Serene as a swan that had never cared for bread” is the kind of image that sounds arresting but doesn’t quite mean anything coherent — it’s decorative cleverness, style performing as substance.

3. The aside about Mrs Jenkins not speaking Portuguese herself is the kind of deflating, character-revealing detail that comes from actually knowing how people talk.

As it happens, Claude was correct: (1) and (2) were generated by GPT-5.5, and (3) was written by me.

Increasing the sample size

Of course, these are just three examples. So to test things more thoroughly, I generated 10 very short stories (i.e. less than 400 words) with GPT-5.5, then got Claude to guess at the probability that they were AI generated.

Because LLMs can be inconsistent, I asked Claude to judge each story five times, then took the median probability. (There are much better ways I could have approached this, combining different prompts and models, but this is just a quick Substack experiment, and I wanted to mimic how a typical user might test a story with Claude.)

For all of 10 stories, Claude was confident it had been given an AI-written story:

You’ll notice that these probability values aren’t particularly plausible; Claude is overly keen on 85% and 92%, it seems. Rather than actually putting together a well calibrated estimate, it’s likely that Claude has made a judgement then provided an accompanying number that looks like the sort of number someone might give alongside a judgement.

Still, it’s a clean sweep for the AI story police.

Can AI spot a human?

The above suggests Claude can spot GPT-5.5 stories. But can it also detect stories that are not AI written? One of the challenges with using existing published fiction is that these stories may well be in Claude’s training data, so in this case it wouldn’t be a fair test scenario (and nor, if you’re an author, a fair training scenario in the first place).

Fortunately for this experiment, I wrote quite a few short stories back in the day. A couple were published (one in Nature, another shortlisted for the Bath Short Story Prize) but a lot weren’t. Many were flash fiction, so comparable length with the GPT stories above.

I picked 10 and got Claude to assess the probability they were AI-generated. All came back with a ‘human’ judgement:

Then I got curious. What if I got AI to ‘improve’ these stories? Would Claude be able to spot these AI-enhanced hybrids? I gave each story to GPT-5.5 and asked it to ‘Rewrite the story to make it better while preserving core meaning, tone, and events’.

Five of them were now strongly flagged as AI-written:

What sort of edits gave the AI away? The change that stood out were classic never-lived-it AI phrases. For example, changing a human-written line describing a ‘draught clutching at my ankles’ to the more perplexing ‘draught worrying my ankles’.

Why does AI writing have these quirks? And can we detect AI-generated writing more reliably? Last week Tuhin Chakrabarty wrote a fascinating piece that tries to get at the mechanism behind the weirdness of AI writing. The logic was that if AI has been trained on the internet, an AI-written story is likely to contain warped fragments of web fiction.

In other words, AI typically produces bad writing because it has been trained to predict what words should come next in a story, and there are a lot of badly written stories out there on the internet to train on.

Yet, like an acerbic book critic who can’t write books of their own, LLMs have also consumed a vast amount of content on what good writing – and bad human and AI writing – looks like. So they have become increasingly good at spotting the giveaways of AI-slop, even if they can’t stop producing it.

Footnote

Here is the prompt I used to generate the GPT-5.5 stories: You are a skilled literary author. Write compelling flash fiction that reads as entirely human-authored. Avoid AI tells: no moralising endings, no em-dashes as stylistic crutches, no over-tidy structure, no suspiciously balanced sentence rhythm. Use idiosyncratic word choices, minor digressions, and naturalistic imperfection. Stay under 400 words.