AI-driven disinformation spreads rapidly ahead of US midterms

The rise of cheap disinformation

Artificial intelligence has fundamentally changed the economics of spreading lies online. Former Facebook data engineer Frances Haugen noted that information operations can now be run with a level of quality that was impossible during the 2016 election, all while remaining very cheap to produce.

Because AI can generate content at scale, bad actors can simply keep tweaking a post slightly each time they create it. This tactic is designed to bypass the automated filters used by social media companies to catch prohibited content. Haugen warned that these small adjustments allow many deceptive posts to slip through the cracks.

This shift is particularly concerning because social media companies have scaled back their human trust and safety teams. Following a counter-backlash over fears of speech restrictions, many platforms have seen significant layoffs in the departments responsible for monitoring election integrity.

How misinformation mutates

A recent case involving a Texas Senate candidate highlights how disinformation evolves as it moves across different audiences. A false claim was initially spread about the candidate being hospitalized in Ohio, but as it was picked up by Spanish-language media, it mutated into a claim that he had attempted suicide.

The Digital Democracy Institute of the Americas (DDIA) tracked this propagation and found that these mutations make original corrections insufficient. When a lie changes its core narrative as it spreads, a single fact-check cannot address every version of the story that people are seeing.

Furthermore, there are fewer guardrails in place for non-English languages. Roberta Braga of the DDIA noted that platforms enforce terms of service more congruently in English, leaving Spanish-language content more vulnerable to unchecked disinformation.

Government and platform roles

The influence of disinformation is no longer confined to fringe websites or anonymous accounts. In one instance, an official statement from the Department of Homeland Security amplified a false claim that a Democrat candidate had employed an undocumented immigrant as a campaign worker.

The campaign involved held records proving the individual had no paid position, yet the claim continued to circulate in ads. Activists have since filed complaints accusing the government of violating the Hatch Act, which limits the political activities of federal employees.

Meanwhile, major platforms like Meta and YouTube have narrowed their policies. YouTube stopped removing content that falsely claimed the 2020 election was marred by widespread fraud, citing a desire to avoid a chilling effect on political speech.

The limits of current moderation

Current platform rules are often very specific and may not cover many types of direct lies. Meta requires advertisers to disclose the use of synthetic images, but the company does very little active moderation of general political commentary. They primarily focus on removing content that misrepresents voting dates or locations.

YouTube uses a likeness detection system and allows civic leaders to request the removal of videos that falsely depict their deeds. However, the company prefers to amplify authoritative news outlets through information panels rather than taking down content that falls outside of their narrow definitions of misinformation.

X, formerly known as Twitter, replaced internal safety reviews with a crowdsourced community notes program. While the platform has a civic integrity policy, experts argue that the rules amount to little without active enforcement against synthetic media designed to mislead.

The challenge of distribution

As AI makes it harder to identify specific pieces of content, the focus of safety experts is shifting toward the systems of distribution. Because AI can slightly tweak a post until it gets through, it is becoming harder for humans to monitor every individual instance of a lie.

The goal for the future of election integrity may rely on understanding the biases and feedback cycles that allow certain information to go viral. Experts suggest that the community needs better ways to provide feedback to damp down the spread of these automated campaigns.

Campaigns across the country are currently scrambling to respond to these lies in real time. They face a difficult quandary as they try to determine who to turn to for help when fabrications propagate faster than they can be debunked.

What to watch for

The situation remains fluid as the midterm elections approach. Observers will be watching to see if platforms implement more robust guardrails for non-English languages and if government agencies face consequences for amplifying unverified claims.

Try it

Check the 'Community Notes' or 'Information Panels' on your social media feeds today to see how platforms are currently attempting to provide context to high-stakes political content.

Watch

Whether government agencies will face formal investigations for amplifying campaign-related disinformation.

How platforms will handle the 'mutation' of lies as they cross into different languages and demographics.

The effectiveness of C2PA metadata in identifying AI-generated images in political advertisements.