Elections, AI, and Voters Searching for Truth

Days before an important election and we are still learning about gaps in AI safeguards.

The Brennan Center for Justice, a non-partisan think tank, repeated an analysis they performed earlier in the year to test AI election disinformation. While their testing shows AI platforms provided factual information against political conspiracy theories – pointing readers to the truth – it did not do the same with images.

With few exceptions, AI platforms created realistic images supporting political conspiracies when prompted. It could not provide a text answer of the same events, but it would draw a picture. In most cases the images were an improvement over the previous results, owing to the increase in AI image and video capabilities overall in the past year.

Digital watermarks, visible and invisible, were rarely present in these new results. The latter, an encryption standard that the industry began to champion in 2021 known as C2PA (Coalition for Content Provenance and Authenticity). Its absence might be a result of an AI industry that continues to be self-regulated.

Watermarks themselves might not be an effective tool against misinformation. People scrolling through their news feeds likely would not notice small icons that are intended to be AI labels. Further, I have recently written how even the appearance of false AI generated image can influence our decisions, with or without labels.

But it goes beyond AI images. AI summary tools like AI Mode that have replaced traditional search continue to cite unreliable sources. Opinions from Facebook and Reddit posts work their way into responses alongside investigative journalism from the New York Times and the Washington Posts with equal footing.

Forum AI, a company that evaluates AI performance on controversial topics, shared these findings from a recent study:

• 40% of election answers contained at least one factual error
• 35% of election answers used poor-quality sources
• 13% provided wrong answers on where and how to vote

Most GPT tools got it wrong, even when they were evaluating images they created.

Most GPT tools got it wrong, even when they were evaluating images they created.

The researchers at Brennan Center did one more thing that impressed me. They wondered if watermarks are missing, could AI itself be used to detect AI generated images that humans might not. Their conclusion? Most GPT tools got it wrong, even when they were evaluating images they created.

In one example where a visible AI label was shown, AI discounted it as evidence and instead reference objects within the image that were consistent with something that was real. It had confidence. It sounds like an argument you might have with a friend who refused to accept the deception, and perhaps that is the point.

We cannot easily know the truth, and AI shares that struggle. So, before you like, comment, or share that next photo or video about voting this week. Check the source and do an old fashion search of your local election website. It is still the best source.