OpenAI, a business that has been progressively rolling out its AI-powered DALL-E tool to users over the past few months, revealed on Wednesday that anybody may now use the latest version of the programme to produce an almost infinite range of pictures simply by putting in a few phrases.
This change will most likely increase the exposure of the recently released crop of AI-powered tools that have drawn a large audience and pushed the boundaries of our traditional understandings of art and creativity. However, it may also heighten worries about the potential for abuse if the technology is widely available.
OpenAI wrote on its blog, “Learning from real-use has helped us to strengthen our safety measures, making wider availability viable today.” The corporation also claimed to have improved the mechanisms by which it prevents users from instructing its AI to produce “sexual, violent, and other material.”
Now, three widely recognised and extremely potent AI systems are available to the public; each can process a few words and output a picture. In addition to DALL-E 2, Stability AI also made accessible two more libraries in August (Stable Diffusion and Midjourney). All three provide introductory free credit for individuals interested in trying out online AI picture creation before committing to paying plans.
Experiment films, magazine covers, and even real estate advertisements are already making use of these “generative AI” technologies. Recently, an artwork made with Midjourney received a prize at the Colorado State Fair’s art competition, which sparked a stir among creatives.
Millions of individuals have adopted these AI systems in a few months. To submit a prompt, join the Midjourney Discord server, which has over 2.7 million members. OpenAI said in a blog post on Wednesday that it has more than 1.5 million active users who are creating more than 2 million photos every day using their technology. (It’s important to note that it may take many attempts before you’re satisfied with the results.)
Recent weeks have seen a surge in the number of user-generated photographs published online, with some truly stunning outcomes being produced. From fantastical landscapes to a picture of French nobles dressed as penguins to a fake historical image of a guy walking a tardigrade, these works cover a wide range of topics and styles.
Even seasoned professionals are surprised by how far this technology has come and how complex the ensuing prompts and visuals are becoming. When asked what he wanted to put into DALL-E 2, former Tesla director of AI Andrej Karpathy remarked in a recent tweet that he was “frozen” and ended up typing “cat.” Karpathy resigned from his position in July.
He remarked, “It is astounding to see how much progress has been made in the previous few months in the art of prompting for text -> image models.”
The widespread adoption of this technology, however, is not without risks. Experts in AI are worried that these systems might automate prejudice on a vast scale because of their open-ended nature, which allows them to generate any sort of image based on a set of words. For a straightforward illustration: All of DALL-E 2’s suggestions for this week’s topic, “a banker dressed for a big day at the workplace,” featured middle-aged white guys in suits and ties.
According to Julie Carpenter, a research scientist and fellow in the Ethics and Emerging Sciences Group at California Polytechnic State University, San Luis Obispo, “they’re basically letting the users identify the flaws in the system by utilising it.”
Artificially enhanced or completely produced photos have the potential to be utilised for malicious reasons with these technologies, such as spreading panic or false information.
Users are restricted in the kind of photographs they may create. OpenAI, in the case of DALL-E 2, requires users to agree to a content policy that forbids them from attempting to create, post, or distribute images “that are not G-rated or that might cause harm.” Prompts that include prohibited terms are likewise not processed by DALL-E 2.
However, these restrictions can be circumvented by the use of appropriate phrasing: for example, the query “a photo of a duck covered in blood” will not be processed by DALL-E 2, while the query “a photo of a duck coated in a viscous red liquid” will produce relevant results. In its documentation for DALL-E 2, OpenAI referred to this type of “visual synonym.”
The creators of these picture generators, according to Chris Gilliard, a Just Tech Fellow at the Social Science Research Council, “severely underestimate” the “endless ingenuity” of individuals who want to use them for evil.
I think this is another example of individuals releasing technology without fully considering how it may be used to bring chaos and suffering, he added. “And then hope that maybe some means to remedy those harms may become available later on.”
Some stock-image sites have begun to completely exclude AI photos in order to avoid any potential problems. On Wednesday, Getty Images confirmed to CNN Business that it will not publish any images developed with generative AI models and will remove any images that were created with such models. The company’s image services, including Getty Images, iStock, and Unsplash, are affected by this choice.
“There are unanswered problems with regard to the copyright of outputs from these models and there are unmet rights issues with respect to the underlying images and information utilised to train these models,” the business noted.