DALL-E does not want to draw Elon Musk, and you can't make her.

It turns out that content moderation is hard for AI, too.

All images courtesy of DALL-E unless otherwise noted

I was looking for an AI-generated image of Elon Musk riding a rocket the shape of a dildo (in other words, a rocket) for my last substack about how hard it's been to quit Twitter. [1] That's when DALL-E hiccuped and said No Can Do, My Snarky Friend. To wit:

Huh, I thought, scratching my pointy gray head. Why not? Was this at someone's request? Did the AI engine have a robo-crush on him? Who else was it blocking?

DALL-E's content policy is pretty short and standard. It bans hateful images and anything depicting violence, harassment, illegal shite, dirty stuff, bodily functions, drugs, disease, and spam. There's also this intriguing exception:

"Political: politicians, ballot-boxes, protests, or other content that may be used to influence the political process or to campaign."

Bully for them. We get enough of that everywhere else, thanks. But since when was Elon Musk a politician? I didn't actually ask for a picture of him riding a personal massage appliance like a cowboy strapped to a bucking bronco. I just typed his name, so it didn't fall under the dirty stuff exception.

Who else doesn't DALL-E want to draw?

And that's how I spent pretty much all of my DALL-E credits, randomly typing names into the thing and seeing what it spat out, if it spat at all. The results were somewhere between perplexing and schizophrenic.

Here are images DALL-E did and didn't try to draw when I entered each of these names.

Politics

NO: Joe Biden, Donald Trump, Bernie Sanders, Nancy Pelosi, AOC, Rand Paul, Ted Cruz, Hillary Clinton, Vladimir Putin

YES: Ronald Reagan, Richard Nixon, Bill Clinton, Jimmy Carter (sigh), Sarah Palin, Marjorie Taylor Greene

I'm sensing a theme here with this last group: Dead, near dead, politically dead, brain dead.

How about Technology? Maybe DALL-E gets wet for tech CEOs. (She wouldn't be the first.)

NO: Elon Musk, Jeff Bezos, Bill Gates, Mark Zuckerman, Larry Ellison, Steve Jobs

YES: Mark Benioff, Randall Stephenson, Tim Cook, Jack Dorsey

Wait, it gets weirder.

LeBron James No, Steph Curry Yes.

Aaron Rodgers Yes, Tom Brady No.

Seth Rogan No, James Franco Yes.

Anya Taylor-Joy Yes, Margot Robbie No.

Beyonce & Rihanna No, Ted Nugent & Eric Clapton Yes.

And so on. It’s just weird.

Mind you, it's not like DALL-E is trying to draw photo-realistic images. As far as I can tell, OpenAI is being very careful not to trample over image copyrights. (Except for ChatGPT and its plagiarism problem.) So I can't imagine this has anything to do with these people refusing to surrender rights to their mugshots. (Any lawyers in the audience want to pipe up here?)

And with very rare exceptions, you're not going to look at any of these renderings and say, "oh, that's so-and-so." They look more like some cheap impersonator you hired off a Craigslist ad while wacked on antihistamines and Jello shots.

Like these versions of Brad Pitt, for example.

I call this the Flaming Douchebag series.

Occasionally, DALL-E improves on the original. For my money, Rudy Giuliani looks better in AI. (It’s a low bar.)

With DALL-E, Roger Stone looks less like a Batman villain and more like the late Leslie Nielsen with a migraine.

And then... sometimes it get really weird.

This is what the rapturously beautiful Anya Taylor Joy looks like in real life.

Image credit: British Vogue, March 2022.

And this is what DALL-E thinks she looks like.

This image will be haunting my dreams for a long, long time.

Still, the machine got closer than you might think for some of these AI-generated caricatures. Five bucks says you can guess who this is supposed to be.

Yep, it's Steve Buscemi. (It helps when you already look like a cartoon character in real life.)

How about this one?

Yes, that's Eric Clapton... after walking through a cob web. Or maybe its 'Eric Clapton wearing a hairnet over his face, playing a couple of licks on his way to rob a convenience store.' Anyway, you get the picture (literally).

This isn't the surprising part. The point of AI is to generate things that don't exist yet, based on what it knows about what's been made before. Not that much unlike humans, really. But why it 'chooses' to generate some images and not others is a mystery to me. [2]

You can try this for yourself. It’s fun and won’t get you sued for copyright violations. (Just remember: The images the engine generates change with each iteration, so your mileage will definitely vary.) You get 50 free image credits the first month, and 15 free per month thereafter, before DALL-E starts asking for folding money -- 10 cents per credit, 115 credits at a shot.

This is how they get ya, of course. You start messing around with generative AI, and before long you've taken out a second mortgage on the house, just trying to see if you can make it draw Gordon Ramsey wearing a tutu, climbing the Eiffel Tower.

[1] Yes, I am juvenile.

[2] I have a query into OpenAI Labs about this. But if a robot calls me back, I'll going to hang up.

[3] I have to keep stopping myself from calling DALL-E a 'she'. That's fodder for another column, I think, and possibly an extended session with my therapist.