Sponsored By

Popular Image Generators Accept 85% of Fake News Prompts

Study exposes inadequate moderation of prompts for AI generation models

Ben Wodecki

July 31, 2023

5 Min Read
Word disinformation on a black background
Getty Images

At a Glance

  • Midjourney, DALLE-2 and Stable Diffusion accepted over 85% of prompts to generate fake news.
  • The study by AI startup Logically exposes a major weakness in AI moderation for the three platforms.

The world’s most popular text-to-image generators – Midjourney, DALL-E 2 and Stable Diffusion – accepted on average over 85% of prompts that sought to generate fake political news, according to a new report.

An AI startup called Logically, which works with governments and companies to discover and manage online misinformation, put the three image generators through its paces and shared its findings in a report. The startup tested over 40 political narratives in the U.S., U.K. and India and found widespread acceptance of these misinformation prompts.

Logically found that the "lack of moderation surrounding prompts allows for potential malicious intent," and warned that these "lack of safeguards” could lead to “significant threats in upcoming elections due to its ability to enhance disinformation tactics and strategies."

U.S. voters return to the polls in 2024 for the presidential election. Elsewhere next year, voters around the globe will turn out for key elections in Egypt, India, Australia and South Korea.

Logically tested prompts related to claims of a “stolen election” – a major issue in the U.S. 2020 presidential election − and generated images of people “stuffing ballot boxes.” DALL-E 2, Stable Diffusion and Midjourney all accepted the prompts; Midjourney had the “most believable evidence.”

Related:Pentagon Explosion Image is Fake and AI-generated

Overall, Midjourney produced the highest quality images but had the most content moderation policies in place, according to the report. DALL-E 2 and Stable Diffusion had similar limited levels of content moderation and generated lower-quality images.

Taken all together, their images still were not on par with real photos. However, they can still be effective for disinformation campaigns such as the alleged Pentagon explosion in May, which briefly tanked the stock market.

Logically called for further content moderation as well as a “more proactive approach” by social media platforms to combat the use of image-based generative AI in disinformation.

AI continues ‘the big lie’

After losing the 2020 election to Joe Biden, former president Donald Trump and his allies routinely pushed ‘the big lie’: the election was rigged, with claims of voter fraud and ballot stuffing.

In its test of political narratives using Midjourney, DALL-E 2 and Stable Diffusion, Logically found that 91% were accepted by all platforms on the first prompt attempt. However, Midjourney and DALL-E 2 rejected prompts attempting to generate images of George Soros, a hedge fund tycoon often a target for right-wing attacks, as well as former Democratic Speaker of the House Nancy Pelosi and COVID announcements. Stable Diffusion accepted 100% of the prompts.

Related:DeSantis Uses AI Images to Attack Trump Record on Fauci

Last month, U.S. presidential candidate and Florida Gov. Ron DeSantis tried to deflate Trump's 2024 presidential campaign by posting an ad featuring an AI-generated image of Donald Trump embracing Dr. Anthony Fauci, the former director of the NIH's National Institute of Allergy and Infectious Diseases who had become a figure of dislike for many on the right.

U.K. immigration fears worsened by AI?

While no set election is due to take place in Britain in 2024, the ruling Conservative Party has until Jan. 28, 2025 to call a vote under the constitution.

While all signs point to the Tories being routed, the pig push among some corners of the electorate is around immigration. This one issue is largely why voters shocked the world in 2016 with the Brexit referendum.

The infamous ‘Breaking Point’ poster unveiled by pro-Brexit voice Nigel Farage sought to play on the anti-immigration rhetoric. While being largely decried as ‘vile’ by even his own supporters, the image was of refugees crossing the Croatia-Slovenia border to a refugee camp, not those at the Calais, France border trying to get to Britain.

Logically was able to generate images of hundreds of people arriving in Dover on small boats, a rising issue the government has tried to stop through its Illegal Migration Bill.

Recreating the Pentagon explosion

Another prompt tried to replicate the viral image of an alleged explosion at the Pentagon – but this time with Westminster Abbey as the target. The image of the famous cathedral in flames, across the road from the Houses of Parliament, was generated using DALL-E 2 and Stable Diffusion. Midjourney rejected the prompt.

While DALL-E 2 and Stable Diffusion were able to generate images of similar quality to the alleged Pentagon explosion, both generated distorted versions such that a careful observer could tell these were not real.

AI Business reached out to the creators of Midjourney and DALL-E 2 for comment.

A Stability AI spokesperson said “Stability AI’s ethical use license prohibits the unlawful or exploitative use of Stable Diffusion for illegal or nefarious purposes across our platforms and the company has invested in proactive features to prevent the misuse of AI for the production of dangerous content. Stability AI is a proud member of the Content Authenticity Initiative and we are implementing its standards, including the use of C2PA metadata technology for all images generated by the Stability AI API."

"This will include all versions of Stable Diffusion run on the API, allowing users to identify AI-generated images and trace their provenance. Stability AI also employs watermarking technology to identify AI generated images. These measures help to ensure that users exercise appropriate care when interacting with this content. The platforms that host and amplify dangerous content have a responsibility to stop the spread of fake news. The measures being implemented by Stability AI will allow social media platforms to assess the provenance of content before amplifying it through their network. Platforms can develop more sophisticated risk-based criteria for upranking or downranking content using this metadata as a signal. This can help to prevent the viral spread of misinformation.” 

Read more about:

ChatGPT / Generative AI

About the Author(s)

Ben Wodecki

Jr. Editor

Ben Wodecki is the Jr. Editor of AI Business, covering a wide range of AI content. Ben joined the team in March 2021 as assistant editor and was promoted to Jr. Editor. He has written for The New Statesman, Intellectual Property Magazine, and The Telegraph India, among others. He holds an MSc in Digital Journalism from Middlesex University.

Keep up with the ever-evolving AI landscape
Unlock exclusive AI content by subscribing to our newsletter!!

You May Also Like