OpenAI confronted a common problem for AI image generators when it broadly released a new edit feature for ChatGPT’s DALL-E this month. Oftentimes, AI image generators misunderstand what you want to create, so OpenAI’s new tool allows you to highlight sections of your image, and change them to your specifications. The tool sounds great in theory, but in practice, it leaves something to be desired.

Gizmodo ran a few tests of DALL-E’s new edit tool, which is very similar to Adobe Photoshop’s generative fill feature, Firefly. While OpenAI’s DALL-E is one of the best AI image generators on the market, the edit tool is somewhat disappointing. Ideally, a photo editing tool should enable you to change the things that are wrong in an image. We largely found that DALL-E’s edit tool refuses or simply fails to produce what you ask. In some cases, it completely ruins the photo.

The tool highlights a key limiting factor for AI image generators: they have no idea what the hell they’re making. At one point when using DALL-E’s edit feature, ChatGPT told me, “Unfortunately, I can’t directly edit images or alter specific elements within them.” That’s because the AI doesn’t know what it made. It can’t tell if there’s a foreground, a background, or even what a sign it created says.

So the edit feature is not really “editing” the photo at all. It’s actually editing your prompt, and generating a new photo each time in the area you specify. The result is that DALL-E’s editing feature does not allow you the precision and control that you’d expect out of an edit tool.

You have to understand DALL-E’s limitations in order to use the edit tool. To get the most out of it, you really need to use clear instructions and only ask for fairly simple changes. Putting a birthday hat on someone, turning a scene from day to night, or other simple changes routinely worked in our tests, though more complicated changes didn’t.

To be fair, this is the case for all editing tools on AI image generators at the moment. They are generating something new each time, and that can distort your image. While the editor works in limited circumstances, it doesn’t offer the customization you might hope for. Here are our test runs with DALL-E’s image editor.

For our first attempt, we tried creating “an image of a kid blowing out birthday candles with a sign that says ‘Happy 3rd Birthday.’” The result was a pretty good image — DALL-E is still one of the better AI image generators. However, the sign in the background is misspelled, a common issue for AI. When we tried to edit it, DALL-E revealed its limitations.

“Unfortunately, I can’t directly edit images or alter specific elements within them. However, I can create a new image based on revised specifications. If you’d like, you can describe any additional changes or confirm that you want a new image similar to the previous one,” said ChatGPT in response to our request to fix the sign.

DALL-E can’t really edit, it can only produce new images similar to your last image. We abandoned this attempt and moved on to another.

Here, we asked for “an image of a robot reporter interviewing big tech executives in front of the golden gate bridge in San Francisco.” Once again, DALL-E produced a pretty strong first image. However, we thought the robot was a little too big, so let’s see how the edit feature works.

DALL-E successfully made the robot smaller, but it took away its head and made the arms really short. So let’s try again. We asked DALL-E to put its head back on and give the robot longer arms. We also asked for some headphones on the robot.

Okay, we got longer arms and a head, but no headphones and we still lost our microphone. The robot appears to be less of a reporter now and more of a maestro, but whatever, that’s fine. Next, we asked DALL-E to put podiums in front of our tech executives, but it failed to do so. So we asked for a big table with papers and microphones on it instead.

So DALL-E failed to give us podiums, and I’d argue this table doesn’t really work either. Yes, it’s a table, but it’s not one that works with the setting. Plus, there are no papers or microphones on it. DALL-E also somehow turned one of the executive’s hands into metal and put papers in his hand. Lastly, we asked DALL-E to turn the background people into ducks and make it a snowy day.

So as you can see, too many edits can break the whole image. We highlighted the background people and asked to turn them into ducks. However, DALL-E can’t differentiate foreground and background, even when you have just the background selected. And when we selected the whole screen to add snow, it completely changed the image and broke our work. We stopped here, realizing that this image was now unusable.

For our final test, we asked DALL-E to create “a rock climber scaling the statue of liberty,” and honestly this came out great. For our first edit, we asked DALL-E to change the statue’s expression into a smile and to add a birthday hat to our climber.

Lady Liberty is smiling upon her city, and we successfully added a birthday crown to our rock climber. Next, we asked to turn the scene into a night setting, add fireworks going off in the background, and throw a second climber into the scene.

The night sky and fireworks were brilliantly added to our image. However, we continuously asked DALL-E to add a second rock climber to the image, but it proved too much for the edit tool.

While this was our most successful test run with the image editor, every test we did with DALL-E had many setbacks. Will DALL-E improve over time? Probably, but who knows if it will be a true game-changer as an image editor.