Technology

93.2% macro-F1 for human vs. AI detection and 68.4% macro-F1 for six-way authorship attribution

This isn't what I'd call "reliable".

I'm also not impressed with their methodology, which is heavily based on Gemini.

To generate mirrored AI stories, we reverse-engineer writing prompts from each human story by prompting Gemini 2.5 Flash (Gemini Team, 2025a) to infer the underlying premise

That's too many steps removed from anything you'd encounter in the wild. They're not even testing against human-prompted output.

And then they use Gemini again to analyze all the stories. Relying on proprietary cloud models for the core of your analysis is like building on sand.

I wonder how does this stand to writers who have learnt English as a second language

... which will be fed into the next generation of LLMs in order to stop sounding quite so much like LLMs.

Here’s the thing: Read enough AI output and everyone can spot it. And that’s the seam that runs through it. The answer is partly right and it’s important to be honest about it. That’s the load bearing statement. It’s this. But it’s also that. And that’s doing a lot of the work here.

It’s. So. Staccato. And it sounds like tech bro in a Patagonia gilet.

Find me an option that works like GDocs keeping a running history of writing and editing, and I'll consider it.

I have receipts for the integrity of my writing. I dont need some shitty tool to validate it.

Millions of college students silently scream.

"We could identify it because it's shit." - researchers

"On the one hand, this is good news. On the other hand..."