Post

AI RANT Following the cancellation of three major debut book deals over AI suspicions, I decided to put my own (needless to say, 100% human-written) prose to the test. My detector of choice was Originality AI, and the results proved to be jarring, to say the least. 🧵1/5

AI RANT (cont’d) The picture above isn’t a case of “false positives.” It’s a case of ALL sentences in a 300-word sample being classified as either ~100% AI- or ~100% human-written *depending on a single punctuation mark*. 🧵2/5

AI RANT (cont’d) Originality AI also offered me some “helpful” suggestions on how I could “improve” my style. These suggestions, along with my response to them as a two-time Nebula and one-time Compton Crook nominee, are displayed below. 🧵3/5

AI RANT (cont’d) The most troubling thing about this is, Originality AI is apparently used by agents & publishers alike. Dear agents and publishers, PLEASE DO NOT FORGET: AI “detectors” are slop machines same as the slop machines whose output they purport to analyze. 🧵4/5

I've had the em dash beaten into me through high school, lots of older books (like Lem) and then additionally through my university where Chicago Style was law. seeing people assume that using the em dash is a surefire, 100%, never-miss way to identify AI writing kills me inside every time.

Interesting that the publishers/editors can’t just read the copy and decided for themselves, but I guess that would take too much time. Slop machines are ok for them but not for anyone else I guess.

Per my other comment, Pangram was also the ONLY one mentioned in the two canceled book deals and the other major accusation. They're built in a completely different way, and I think it's really important to know of any examples of false positives there vs. the super explicitly scammy ones!

Hasn’t Pangram also gotten called out for false positives? Also even if it’s better, it doesn’t change the problem that if agents/publishers rely on “AI detection” products, given ALL products’ possibility of false positives they’re inevitably going to make false judgments that there’s genAI text

Nor does it solve the problem that writers’ integrity to their own writing is being actively warped by the knowledge that editors, agents, and publishers are using these deeply flawed tools to judge their writing. slate.com/technology/2...

I have yet to see an example of Pangram doing a false positive on a story or portion of a novel, and I very much want to in order to better understand how it might happen. Let me know if you have any!

Oh originality is EXPLICITLY a massive scam/fraud. I told them they have basically built an em dash detector a year ago after a similar result (fully human-written chapter flagging as 90%+ AI, switching four natural em dashes to commas over 1,000+ words changed it by over 40%).

Yikes, Yare! This is both disturbing and annoying.

Absolutely--but at the same time, it's great to hear from you, Beth!

I think something that needs said as well is that they're slop machines that are guided initially by human intent. A HUMAN undoubtedly set up the training to make it highly suspicious of things like em dashes, because right now it's trendy to point to them as a signature of AI.