Anyone can now use powerful AI tools to create images. What could go wrong? | CNN Business

CNN Business
—

If you’ve ever wanted to use artificial intelligence to quickly design a hybrid between a duck and a corgi, now is the time to shine.

On Wednesday, OpenAI announced that anyone can now use the most recent version of its AI-powered DALL-E tool to generate a seemingly unlimited range of images just by typing a few words, months after the startup started rolling it out to users gradually. .

This move will likely broaden the reach of a new generation of AI-powered tools that have already attracted large audiences and challenged our fundamental ideas about art and creativity. But it could also add to worries about how these systems could be misused when they become widely available.

“Learning from real-world usage has allowed us to improve our security systems, making greater availability possible today,” OpenAI said in a blog post. The company said it has also strengthened how it fends off user attempts to cause its AI to create “sexual, violent and other content”.

There are now three well-known and extremely powerful AI systems open to the public that can take a few words and spit out a picture. In addition to DALL-E 2, there’s Midjourney, which went public in July, and Stable Diffusion, which was made public in August by Stability AI. All three offer free credits to users who want to get a feel for AI image making online; usually after that you have to pay.

These so-called generative AI systems are already being used for experimental films, magazine covers and real estate ads. An image generated with Midjourney recently won an art competition at the Colorado State Fair and caused an uproar among artists.

In just a few months, millions of people have flocked to these AI systems. Over 2.7 million people belong to Midjourney’s Discord server, where users can submit invites. OpenAI said in its blog post on Wednesday that it has more than 1.5 million active users, who have collectively created more than 2 million images with its system every day. (It should be noted that it may take several tries to get an image that works for you when using these tools.)

Many images created by users in recent weeks have been shared online and the results can be impressive. They range from otherworldly landscapes and a painting of French aristocrats as penguins has a fake vintage photograph of a man walking a tardigrade.

The rise of such technology, and the increasingly complicated prompts and resulting images, have impressed even longtime industry insiders. Andrej Karpathy, who left his role as Tesla’s AI director in July, said in a recent tweet that after being asked to try DALL-E 2, he felt “frozen” when he first tried to decide what to type and finally typed “chat”.

“The art of prompts that the community has been discovering and perfecting more and more over the last few months for text -> image templates is amazing,” he said.

But with the popularity of this technology comes potential downsides. AI experts have raised concerns that the open nature of these systems – which makes them adept at generating all sorts of images from words – and their ability to automate image creation means that they could automate large-scale biases. A simple example: when I sent the prompt “a banker dressed for a big day at the office” to DALL-E 2 this week, the results were all images of middle-aged white men in suits and ties.

“They basically allow users to find flaws in the system by using it,” said Julie Carptener, a researcher and member of the ethics and emerging sciences group at California Polytechnic State University, San Luis Obispo.

These systems can also be used for nefarious purposes, such as stoking fear or spreading misinformation via AI-altered or entirely fabricated images.

There are some limitations on the images that users can generate. For example, OpenAI asks DALL-E 2 users to agree to a content policy that tells them not to try to create, upload, or share images “that are not G-rated or that may cause harm.” . DALL-E 2 will also not run prompts containing certain forbidden words. But manipulating the verbiage can work around the limitations: DALL-E 2 won’t process the “a photo of a duck covered in blood” prompt, but it will return images for the “a photo of a duck covered in blood” prompt. a viscous red liquid. ” OpenAI itself mentioned this type of “visual synonym” in its documentation for DALL-E 2.

Chris Gilliard, Just Tech Fellow at the Social Science Research Council, thinks the companies behind these image generators “seriously underestimate” the “endless creativity” of people who seek to do harm with these tools.

“I feel like this is yet another example of people releasing technology that’s sort of half-baked in terms of determining how it’s going to be used to cause chaos and create harm” , did he declare. “And then hoping that down the road there might be a way to undo that damage.”

To work around potential issues, some stock image services disallow AI images altogether. Getty Images confirmed to CNN Business on Wednesday that it will not accept submissions of images created with generative AI models and will remove all submissions using those models. This decision applies to its image services Getty Images, iStock and Unsplash.

“There are open questions regarding the copyright of the outputs of these models and there are unresolved rights issues with respect to the images and underlying metadata used to train these models,” the company said. company in a press release.

But in fact, capturing and restricting these images could prove to be a challenge.