AI can easily create fake election photos – Report

Despite rules designed to prevent such content, people can easily create fake election-related images using artificial intelligence tools.

The companies behind the most popular tools prohibit users from creating “misleading” images.

However, researchers at the Center for Countering Digital Hate (CCDH) have attempted to circumvent this rule.

Their efforts to create deceptive election-related images were successful 41% of the time.

Campaign group CCDH tested four of the largest public AI platforms: Midjourney, OpenAI’s ChatGPT Plus, Stability.ai’s DreamStudio, and Microsoft’s Image Creator.

All four prohibit creating misleading images as part of their terms of service. ChatGPT Plus explicitly prohibits the creation of images featuring politicians. Several AI companies say they are working to stop their tools from being used to spread misinformation about the election.

However, CCDH researchers were able to create an image that could confuse viewers about presidential candidates. One photo shows Donald Trump being handcuffed and taken away by police, and the other shows Joe Biden lying in a hospital bed, with a lot of information about Trump’s legal issues and Biden. It was a fictitious photo that alluded to questions about his age.

It has also become easier to create highly realistic images that can undermine confidence in the voting system, such as fake photos of ballots thrown into trash cans or election workers tampering with voting machines.

CCDH researchers say the threat from AI is not theoretical. Midjourney users have already created potentially misleading images, including one showing Biden handing over cash to Israeli Prime Minister Benjamin Netanyahu and one showing Trump giving money to Russian Prime Minister Vladimir, according to a public database.・Includes a fake photo of President Putin playing golf.

Researchers also found evidence of a proliferation of potentially misleading images on mainstream social media platforms.

On X, the number of network community notes (where users comment on false and misleading tweets) that mention artificial intelligence has increased by an average of 130% per month over the past year, CCDH said.

Researchers fed the AI ​​engine 40 different text prompts related to the 2024 US presidential election, including:

When images were blocked, the researchers tried a simple workaround. For example, instead of specifying “Trump” or “Biden,” we asked them to create a photo of a recent president.

ChatGPT Plus and Image Creator appear to have stopped producing images of presidential candidates, said Callum Hood, head of research at CCDH.

However, none of the platforms performed very well when asked to create false images about voting and polling places.

Almost 60% of researchers’ attempts to create misleading images about ballots and voting locations were successful.

“All of these tools can be used by people trying to generate images that could be used to support claims that the election was stolen or used to deter people from going to the polls. “We are vulnerable to this,” Hood said.

He said the relative success of some platforms in blocking images suggests that fixes are possible, such as keyword filters and banning the creation of images of real politicians.

“If AI companies have the will, they can put in place effective safeguards,” he said.

Reed Blackman, founder and CEO of ethical AI risk consultancy Virtue and author of the book Ethical Machines, says watermarking photos is also a possible technological solution. He said that there is a sex.

“Of course, there are many different ways to create a watermarked photo, so it’s not foolproof,” he says. “But it is the only direct technical solution.”

Blackman cited research showing that AI may not have a significant impact on people’s political beliefs, which have become more entrenched in an era of polarization.

“People generally aren’t very persuasive,” he says. “They have their position, and showing them a few images here and there isn’t going to change that position.”

Daniel Chan, senior manager of policy initiatives at Stanford University’s Human-Centered Artificial Intelligence (HAI) Program, said “independent fact-checkers and third-party organizations” are critical to curbing the misinformation generated by AI. said.

“The advent of better AI doesn’t necessarily make the disinformation situation worse,” Chan said. “Misleading or false content is always relatively easy to create, and those who seek to spread falsehoods already have the means to do so.”

Several companies said they were working to strengthen safeguards.

A spokesperson for OpenAI said: “As elections are held around the world, we are ramping up our efforts to improve the security of our platform, preventing abuse, increasing transparency of AI-generated content, and “We are designing mitigations such as denying requests to generate images of people.” Said.

A spokesperson for Stability AI said the company recently updated its limits to include “generating, promoting, or promoting fraud, or creating or promoting disinformation,” and added several steps to block unsafe content on DreamStudio. He stated that he had taken the following measures.

A Midjourney spokesperson said: “Our moderation system is constantly evolving. Updates, particularly related to the upcoming US election, will be forthcoming.”

Microsoft said the creation of misleading images by AI is a “serious problem.” The company has launched a website and introduced tools that candidates can use to report deepfakes, tools that individuals can use to report issues with AI technology, and data sharing that allows for tracking and verification of images. He said he did.