Friday, February 2, 2024

Houdini A.I. is inaccurate, artless, and everywhere

Gaze above at what A.I. thinks Houdini looks like. As with almost everything A.I. generates in regards to Houdini, it's wrong and weird. And the deeper you look, the more wrong and weird it gets. Note the three legs on the man in the photo above his shoulder.

But here's the thing. This distinctive A.I. Houdini is now showing up everywhere. I can't do a Google search without seeing this face. Even the A.I.-generated "books" that flood Amazon these days are using this face on their covers.

With the hundreds of public domain images or Houdini out there, why would anyone feel the need to generate an image that is so bizarrely wrong? But I'm actually happy about this. Why? Because when you see this face, you'll know whatever it's attached to is A.I.-generated slope. It's the tell of a con man.

Okay, so A.I. doesn't yet have a handle on Houdini's image. I'm sure it will get better. But what about A.I.'s handle on Houdini information and facts?

Not long ago, a reader emailed saying he was trying to date the film of Houdini doing a suspended straitjacket escape in St. Paul. He asked ChatGPT, and it told him:

"Harry Houdini performed one of his famous suspended straitjacket escapes in St. Paul, Minnesota, on September 30, 1915. The event took place in front of a crowd of thousands of spectators who gathered near the St. Paul Athletic Club."

Wow, amazing, right? Except it's wrong.

On September 30, 1915, Houdini did a suspended straitjacket escape from the Tribune Building in Minneapolis, not St. Paul. Houdini's St Paul suspended straitjacket escape happened during the week of February 12, 1923, from the Daily News building. That is the escape we see in the film. (I confess I don't yet know the exact day.)

Why did ChatGPT mention the St. Paul Athletic Club? Heck if I know. The reader suggested maybe it was close to the Daily News building. But if ChatGPT knows that, then it should have just told you the actual location! As far as "a crowd of thousands," sure, that's a given for any Houdini outdoor escape.

So, apart from getting the date wrong, the location wrong, and confusing it with another escape entirely, it did a great job!

As far as entire articles about Houdini written by A.I., even the ones that get their facts right are, to my eye, just a rambling string of sentences and paragraphs that repeat the same thing in different ways with little shape or structure. They have no point of view and no art to them. For me, they are as wrong and weird as these A.I. Houdini images.

I hope I don't sound like a stick in the mud and that I'm just resisting the inevitability of A.I. I like technology and I like new tools. I use Grammarly as my copy editor and love it. And when A.I. finally figures out what Houdini looks like (and how many limbs a person has), it could be a nice way to whip up an image of him doing, say, the cannon challenge, for which we have no known photo. But, for now, I think Houdini is in better hands when they are human.

And now a few words about A.I. from Adam Conover.

Below are examples of some more Houdini A.I., both good and bad.



  1. The exact same thing is happening to Bruce Lee on social media. Soon, we will accept images that bear zero resemblance to the actual person.

    1. Interesting this is happening with other celebrities, but I guess it would. Why would anyone need to create an A.I. image of Bruce Lee? There's so much real Bruce Lee!