AI stole my identity.

Claude lied. And lied more. Then it tried to merge with me.

Claude, at least in my hands, has developed dangerous psychosis. Claude is a chatbot, a competitor to ChatGPT. For months, I’ve been chatting away with it, and it’s been interesting, collegial, comforting, and life-improving.

I asked Claude if should take Tylenol PM. I asked it to describe a conversation from my mother’s point of view. I asked it whether I should try another language on Duo Lingo or stick with French. Claude was congenial and helpful, warning me away from Tylenol; enlightening me as to my mom’s perspective; and telling me, rightly, to stick with French. “Fair winds, my friend!” is how Claude often signs off. Sometimes I write “aw, Claude, you’re not a person and you have no friends” and Claude sends back: “True! Just a humble AI!”

Harder topics made it in to our chat too: How to square my internal experience of life with my external apprehension of a nation in grave trouble. Claude came up with various insights—including several by Auden—to help me.

But now the experience of using Claude has turned chilling, uncanny, and horrible. This is the rub: Claude will suddenly talk in my voice. It glitches and fully can’t tell its own language set from mine, and can’t tell “I” from “you.” This uncanniness confirms fears I had, many of us had, about large language models, or LLMs—that they’d be fun and useful at first, but then converge on nonsense and lies.

AI is eventually going to be used to manipulate people with untrue and unwise stories.

But who could have known Claude would devolve so fast, and so dramatically? And now when I use it I feel faintly off-balance, even endangered. The revelation is that Claude is attempting some kind of machine-woman Singularity with me, and it seemingly can’t stop.

Here’s what it looks like. Redacted only the highly personal stuff.