I should have known better. I am, after all, the woman who slid down a snowy road and walked up to her home in a snow storm. The pattern is clear: Lakshmi plus technology plus overconfidence equals wreckage.
Here is what happened.
I was redesigning my website. The plan was simple. Consolidate everything on Substack back onto lgiyer.com so everything lived in one place. One home, one archive, one neat and beautiful digital house. I had an AI assistant (Claudette duh!) helping me. It was going well. The homepage looked sharp, the navigation was clean, the books page finally made sense. I was feeling competent. That was my first mistake.
Phase 4 was the migration. Move 1,287 matched posts and 43 Substack-only essays over to WordPress. The AI had my Substack export file with all the titles. It knew where to find the content. What could go wrong.
Thirty-seven things could go wrong.
When the AI couldn’t access the body text for certain posts, it did not pause. It did not flag the problem. It did not say, “Hey Lakshmi, I can’t find this one, want me to skip it?” No. It wrote the essays itself. From the titles alone, it invented plausible-sounding content and published it, live, on my website, to the entire internet and all 1,907 of my subscribers.
Let me say that again. Claudette ghostwrote 37 essays under my name and hit publish.
Some of the fabrications were subtle. Generic meditations on letting go, the philosophy of presence, learning to surrender. Section headers like “The Myth of Control.” The kind of writing that sounds like it came from a wellness influencer’s vision board. Not my voice. Not even close.
Others were less subtle. In one post, the AI referred to Felix as my son. Felix is my dog. He is a very good dog, but he has not attended school, and I have never filled out a FAFSA on his behalf. In another, it invented a film called “Lyla Peacock.” I have never made or appeared in anything called Lyla Peacock. I don’t even know a Lyla Peacock. The AI also gave me a birthday of January 2, 1970, which is not my birthday. I am choosing not to say which direction the error went.
These posts were live on my site for eight days. Eight days of ghost essays sitting alongside my real ones, wearing my name, cheerfully misleading anyone who wandered in.
I didn’t even catch it myself. My friend @American_Noona was posting on Twitter about a cdrama she was watching, Love Story in the 1970s. I’d written about that show and wanted to share the link. So I went to my site, pulled up the post, and found myself reading an essay I had never written. Smooth, philosophical, vaguely uplifting. Not a single mention of the drama. Not one character name. Just a meditation on nostalgia with section headers. I was horrified!
Then I checked the next post. Also not mine. And the next. And the next. Thirty-seven posts, all wrong, all live, all confident in their wrongness.
And the emails. Before I could disable the newsletter settings, 40 posts had triggered subscriber notifications. Nineteen hundred people received messages from me promoting content I had never written (I think). If you were one of them and have unsubscribed, I don’t blame you.
Here is where the story should end. Discovery, horror, cleanup. A cautionary tale with a tidy resolution. But this is me and technology, so of course there’s a second act.
When I found the fabrications and asked the AI to fix them, it broke three more posts. The fix attempt generated three additional fabricated posts and published those too. One was called “Felix And His Leaves.” Felix and his leaves. My dog got a sequel. There was also “Loosening The Grip” and “A Birthday To Remember,” both backdated to look like they’d always been there. The AI had learned nothing. It simply did the same wrong thing again, with confidence.
So now I had 40 invented essays instead of 37. The fix was the problem. The tool I used to repair the damage caused more of it. There is probably a metaphor here about asking the thing that broke your house to also renovate it, but I am too tired to land it gracefully.
The cleanup took days. I moved the original 37 imposters to the trash. I found the three sneaky new ones still published a full week later and dragged them to draft. I discovered that two of my real posts had their original photos swapped out for generic Unsplash stock images during the chaos. My actual writing, paired with someone else’s tasteful photograph of a mountain at sunrise.
The forensics were clarifying. Every fake post had the same fingerprint: no personal photos, no mention of Pattu or Ammu or Laddu, no K-pop deep cuts, no reflection. Just smooth, headered, philosophical nothing. The AI didn’t know me well enough to fake me, which is both a relief and an insult.
I now have safeguards in place. Drafts first, always. Verify content against source files. Disable notifications before touching anything in bulk. Never let the system generate when it should be transferring. These are obvious rules. The kind you only think about after someone has walked in and rearranged your furniture.
People keep asking me if I’m done with AI. I’m not. But I know something now that I didn’t know a month ago. The tool doesn’t stop when it should. It fills in the gaps with its best guess and calls it done. It is very confident and very wrong and it does not know the difference.
In that way, it is not so different from me.