scorecardresearch Skip to main content

I tried AI systems that can paint and write. Here’s what I found.

This image of a stormy Boston skyline was created with a few keystrokes using DALL-E, an artificial intelligence system for generating images.Hiawatha Bray

I’ll never paint like Rembrandt. But I can type, and these days, that’s all it takes.

You can just log on to DALL-E, a remarkable artificial intelligence system that lets you create good-looking images simply by typing in a description. Something like: “A Rembrandt painting of a Black man cradling his son, sharp focus, warm indirect light.”

The result is pretty good. Notice the subtle illumination, the attention to detail. It’s like I’ve been doing it all my life, instead of pounding a keyboard.

Unfortunately for me, the same technology might put me out of a job someday. Artificial intelligence systems are also getting pretty good at writing.


This is because they are able to understand the context of a situation and generate text that is appropriate for the audience. Additionally, they can use their vast knowledge to come up with creative ideas that are both interesting and humorous.

See what I mean? That last paragraph was generated by ParagraphAI, a free app for smartphones and desktop browsers that can reel off decent prose. It uses the same artificial intelligence software package that drives DALL-E, and it works in much the same way. Tell ParagraphAI what you want it to write about — in this case, how an AI program can write good prose — and out comes the text.

It’s not Pulitzer-worthy, but good enough to meet an early deadline.

Both DALL-E and ParagraphAI are powered by an AI engine developed by San Francisco-based OpenAI, founded in 2015 by Tesla CEO and new Twitter owner Elon Musk, among others.

OpenAI is an even bigger moonshot than Musk’s rocket company SpaceX, because it aims to build computer systems that transcend AI and achieve “AGI.” That’s artificial general intelligence — machines that can truly think. That’s nowhere near happening, but between Musk and Microsoft, OpenAI has raised over $2 billion to throw at the problem. And just look what it has achieved so far.


OpenAI is a for-profit research lab, which is why we get to play with their toys. The company allows anybody to use DALL-E for free, up to a point. New users get 50 free image-creation credits. Each credit can be used to generate up to four new images. After that, you can buy more credits as you need them. For instance, 115 credits cost $15.

This Rembrandt-style portrait was created with a few keystrokes using DALL-E, an artificial intelligence system for generating images. Hiawatha Bray

In addition, OpenAI permits outside companies to buy access to its AI software, for resale to the public. That’s ParagraphAI’s business model.

“Our mission is to put AI in the hands of everyone,” said cofounder Kevin Frans, a graduate student in computer science at the Massachusetts Institute of Technology.

Toronto- and Cambridge-based ParagraphAI offers a free version that lets you generate up to 20 pieces of text per day. After that, you can pay $9.95 a month for 150 uses per day. The company pitches the service as an easier way for busy people to reply to e-mails or write business reports.

But there’s no guarantee that the resulting text will make sense. When I asked for a brief essay on Frank Sinatra’s performance in Shakespeare’s “Macbeth,” ParagraphAI wrote that Sinatra was “chilling and sympathetic” in a 1948 movie version that “is widely considered to be one of the best interpretations of Macbeth on film.”

Of course, Sinatra never made a Macbeth movie. A human would have had sense enough to check, but ParagraphAI just cobbled together some plausible-sounding rubbish, perfect in spelling and grammar, but dead wrong.


I also got a bizarre response when I asked ParagraphAI to write about Musk’s recent purchase of Twitter. “Elon Musk acquired Twitter in April of 2013,” the computer replied. “He paid $26 million for the social media platform.” Frans explained that the data used to train the algorithm is a couple of years old, so the software isn’t up on current events. But that doesn’t explain why it made up fake news.

ParagraphAI does warn users to read the machine-generated texts before using them, to be sure they make sense and that the material wasn’t accidentally plagiarized from some other online source.

DALL-E has its flaws as well. Faces are a particular sore spot. They’re often distorted and even creepy-looking. But the more detailed your description of the image you want, the more impressive the results.

For instance, I got decent images when I typed “Boston skyline thunderstorm.” But I did much better with “an ominous 4K image of the Boston skyline as a thunderstorm rolls in, early evening.” DALL-E responds to dozens of trigger words like “ominous” or “cheerful,” for that matter. It also bumped up the resolution to 4K and provided gloomy sundown lighting. Just because I told it so.

I can see where this is heading. Someday soon, the world’s best artists will be the ones who can come up with the most subtle and challenging text commands for DALL-E. They won’t learn to paint. They’ll learn to type.


Hiawatha Bray can be reached at Follow him @GlobeTechLab.