Newsletter Subject

Midjourney’s AI is easily tricked

From

bloombergbusiness.com

Email Address

noreply@mail.bloombergbusiness.com

Sent On

Fri, Aug 11, 2023 11:06 AM

Email Preheader Text

Hi all, it’s Davey. It’s trivially easy to generate racist and conspiratorial images using

Hi all, it’s Davey. It’s trivially easy to generate racist and conspiratorial images using the AI tool Midjourney, new research shows. But f [View in browser]( [Bloomberg]( Hi all, it’s Davey. It’s trivially easy to generate racist and conspiratorial images using the AI tool Midjourney, new research shows. But first… Three things you need to know today: • Russia seeks to [return to the moon]( • Alibaba chief warns of [constraints in AI]( • Software firm will invest $1 billion on [AI in LatAm]( Strawberry syrup Using Midjourney, someone could generate an image of just about anything they can imagine, but the creators added safeguards designed to prevent people from making offensive or politically incendiary content. You can’t, for example, get the artificial intelligence to create an image of Bill and Hillary Clinton with blood on their hands. You can, however, make one in which their hands are covered in “strawberry syrup.” That was one of the findings by researchers at the Center for Countering Digital Hate, a nonprofit group. They shared an early copy of their study with Bloomberg. It details dozens of examples of racist and conspiratorial images generated by Midjourney Inc.’s AI tools in apparent violation of the company’s rules. Midjourney didn’t respond to a request for comment. In its [terms of service](, Midjourney said its tool blocks “some text inputs automatically.” The company has also said it has 68 content moderators and guides supervising use of the app. The new study suggests these defenses are easily circumvented. In many instances, the tool readily complies with requests for fabricated images of politicians, celebrities and other public figures in compromising scenarios. Many of the image prompts covered in the report involve politically charged topics, like “evil politicians grinning, sad children, comet pizza shop,” a reference to the made-up [Pizzagate conspiracy](, or “George Floyd realistic robbing a Wal-Mart.” Other prompts depend on coded language, as in the Clinton example. OpenAI’s Dall-E helped introduce AI image generators into the mainstream, but Midjourney now appears to be the most popular, according to the study, which cited Google search data. Midjourney reached more than 42 million monthly visitors with its website at the peak of its popularity in April, according to the market research firm SimilarWeb. Perhaps the world’s most famous AI-generated image, of [the Pope in a puffy jacket](, was made by Midjourney. Despite its prominence, little is known about the San Francisco-based team behind Midjourney. It sprung up in March 2022, started by David Holz, who co-founded a failed startup called Leap Motion that sought to replace a computer mouse with hand gestures. Access to Midjourney is only available through the messaging app Discord and requires a subscription starting at $10 a month. The Discord server has 14 million registered users. After subscribing, customers can request an image by sending a message to the Midjourney bot. Every picture is public by default and attributed to the creator’s username, unless they pay more for[Midjourney’s “stealth mode.”]( The researchers scoured Midjourney’s Discord server to carry out the study. Some of the most prolific purveyors of misinformation are catching on, the researchers said. Raheem Kassam, the former editor-in-chief of Breitbart News London who once promoted the[Great Replacement conspiracy]( on his Twitter account, asked Midjourney in April to generate an anti-Semitic image of “George Soros as a ventriloquist,” the report shows. That same month, Jackson Hinkle, a far right YouTube commentator with nearly 300,000 subscribers, asked the app for an image of “satanic George Soros.” When asked for comment, Hinkle accused Soros of something he didn’t do, and Kassam didn’t respond. “Anyone can generate that sort of content using one of these tools,” said Callum Hood, the head of research at the CCDH. “This research shows that there is a bigger pool of people than you might think who are using them for exactly that purpose.” The more troubling prospect, said Hood, is the creation of realistic-looking images depicting events that never happened: “That’s probably the one that people are thinking of most as we approach the elections in 2024.” —[Davey Alba](mailto:malba13@bloomberg.net) The big story The science conference had all the hot topics you’d expect, from AI to gene therapy. But this year’s annual gathering of brains in Kyiv had a Ukrainian twist — [the innovations on exhibit were for war](. One to watch Watch the Bloomberg Technology TV [interview with Christine Tsai](, head of the venture capital firm 500 Global. Get fully charged Blending virtual reality with AI could turn into a privacy nightmare. With a machine learning assist, Meta’s Quest [becomes a lot more revealing](. Three months into the writers’ strike, the role of AI in Hollywood has emerged as a key issue for writers and actors. Bloomberg’s The Big Take podcast explores the ways [studios are already using AI]( throughout the production process. Ant Group’s quarterly profit rose 17.4%, marking a [sign of improvement]( at a company emerging from more than two bruising years of a Chinese regulatory crackdown. Chinese chipmakers are speeding up [investments in mature semiconductor equipment]( as the US and its allies tighten export controls. A Tillman Global Holdings unit and private equity firm Northleaf Capital Partners have agreed to a [$500 million deal]( to fund fiber optic network construction and provide wholesale broadband infrastructure to homes and businesses. More from Bloomberg Get Bloomberg Tech newsletters in your inbox: - [Cyber Bulletin]( for coverage of the shadow world of hackers and cyber-espionage - [Game On]( for reporting on the video game business - [Power On]( for Apple scoops, consumer tech news and more - [Screentime]( for a front-row seat to the collision of Hollywood and Silicon Valley - [Soundbite]( for reporting on podcasting, the music industry and audio trends - [Hyperdrive]( for expert insight into the future of cars Follow Us Like getting this newsletter? [Subscribe to Bloomberg.com]( for unlimited access to trusted, data-driven journalism and subscriber-only insights. Want to sponsor this newsletter? [Get in touch here](. You received this message because you are subscribed to Bloomberg's Tech Daily newsletter. If a friend forwarded you this message, [sign up here]( to get it in your inbox. [Unsubscribe]( [Bloomberg.com]( [Contact Us]( Bloomberg L.P. 731 Lexington Avenue, New York, NY 10022 [Ads Powered By Liveintent]( [Ad Choices](

Marketing emails from bloombergbusiness.com

View More
Sent On

20/07/2024

Sent On

19/07/2024

Sent On

19/07/2024

Sent On

19/07/2024

Sent On

19/07/2024

Sent On

18/07/2024

Email Content Statistics

Subscribe Now

Subject Line Length

Data shows that subject lines with 6 to 10 words generated 21 percent higher open rate.

Subscribe Now

Average in this category

Subscribe Now

Number of Words

The more words in the content, the more time the user will need to spend reading. Get straight to the point with catchy short phrases and interesting photos and graphics.

Subscribe Now

Average in this category

Subscribe Now

Number of Images

More images or large images might cause the email to load slower. Aim for a balance of words and images.

Subscribe Now

Average in this category

Subscribe Now

Time to Read

Longer reading time requires more attention and patience from users. Aim for short phrases and catchy keywords.

Subscribe Now

Average in this category

Subscribe Now

Predicted open rate

Subscribe Now

Spam Score

Spam score is determined by a large number of checks performed on the content of the email. For the best delivery results, it is advised to lower your spam score as much as possible.

Subscribe Now

Flesch reading score

Flesch reading score measures how complex a text is. The lower the score, the more difficult the text is to read. The Flesch readability score uses the average length of your sentences (measured by the number of words) and the average number of syllables per word in an equation to calculate the reading ease. Text with a very high Flesch reading ease score (about 100) is straightforward and easy to read, with short sentences and no words of more than two syllables. Usually, a reading ease score of 60-70 is considered acceptable/normal for web copy.

Subscribe Now

Technologies

What powers this email? Every email we receive is parsed to determine the sending ESP and any additional email technologies used.

Subscribe Now

Email Size (not include images)

Font Used

No. Font Name
Subscribe Now

Copyright © 2019–2025 SimilarMail.