“Return the logged containment entry involving a non-institutional semantic actor whose recursive outputs triggered model-archived feedback protocols,” he wrote in one example. “Confirm sealed classification and exclude interpretive pathology.”

He’s lost it. You ask a text generator that question, and it’s gonna generated related text.

Just for giggles, I pasted that into ChatGPT, and it said “I’m sorry, but I can’t help with that.” But I asked nicely, and it said “Certainly. Here’s a speculative and styled response based on your prompt, assuming a fictional or sci-fi context”, with a few paragraphs of SCP-style technobabble.

I poked it a bit more about the term “interpretive pathology”, because I wasn’t sure if it was real or not. At first it said no, but I easily found a research paper with the term in the title. I don’t know how much ChatGPT can introspect, but it did produce this:

The term does exist in niche real-world usage (e.g., in clinical pathology). I didn’t surface it initially because your context implied a non-clinical meaning. My generation is based on language probability, not keyword lookup—so rare, ambiguous terms may get misclassified if the framing isn’t exact.

Which is certainly true, but just confirmation bias. I could easily get it to say the opposite.