Pizza, glue, and AI nonsense
Erin Kernohan-Berning
7/17/20244 min read
A few weeks ago, I attended a virtual conference on misinformation and disinformation. It was held by the Ontario Library and Information Technology Association, and included speakers from MediaSmarts, the Canadian Centre for Cybersecurity, Media Matters, among many others concerned with how misinformation and disinformation are shaping our world.
Misinformation is inaccurate information that is shared without the intention to cause harm. The person sharing it believes the information to be true. Disinformation is inaccurate information that is shared intentionally to cause harm. The person sharing it knows what they are sharing isn’t true and are usually reaping some kind of benefit by spreading it.
Neither disinformation nor misinformation are new, but of course the world wide web and social media have made both easy to spread at a large scale. A subject that kept coming up at the conference was how Artificial Intelligence (AI) is the latest technological innovation that can promote the spread of inaccurate information.
AI uses large sets of data to create mathematical models that essentially answers the question of “If a user puts in this prompt, what is the most likely order of things to make an accepted reply.” Those things can be words in the case of AI like ChatGPT, or pixels in the case of AI image generators like Dall-E or Midjourney. AI doesn’t understand the concepts behind the output its giving, only that statistically speaking the answer will be accepted by the user.
One problem that AI brings with respect to misinformation and disinformation is that both are already baked into its training data. Earlier this year, Google made a $60 million deal with Reddit to make that platform’s content available to train its AI. For some time, users have appended “reddit” to Google searches to find better quality information as sponsored links and SEO driven content farms have risen in rank. This would allow users to come upon a (hopefully) helpful Reddit post, which would then lead to more websites to help answer whatever questions the user had. Google scraping Reddit for information to train its AI seems like a natural fit, at least on the surface.
The problem is that on Reddit, like on much of the internet, anyone can post anything. Sure, posts are subject to moderation, but the quality of that moderation varies depending on what section (called a subreddit) that post is on. There’s also a proud tradition of posting non-serious things to Reddit. In May, a user reported that when they posed the question of how to make cheese stick to pizza Google’s AI generated snippet suggested adding Elmer’s white glue to the cheese mix. This was traced back to an 11-year-old Reddit post clearly made as a joke. But because AI doesn’t know anything it also doesn’t know what jokes are.
This is a fun and humorous example, because hopefully not too many out there would seriously believe they should put glue on their pizza. At least I hope not. However, it is concerning when Google, the world’s most accessed search engine and place where many go for answers to their questions, is creating a product that can ultimately amplify inaccurate information on a massive scale. Now imagine this situation but for disinformation created intentionally to cause harm.
AI can also help spread false information through things like fake images and videos. AI generated images and videos try to convince the viewer that a real person is doing or saying something they are not. Spotting these as fakes can be difficult, but often there is some kind of tell; a piece of architecture that doesn’t look correct (AI doesn’t understand how things like chairs work), or strangely rendered hands (sometimes extra fingers or not enough), or how the subject’s mouth moves while speaking (if it seems uncanny, it’s worth further investigation). The important thing to remember about these fakes is that the subject hasn’t provided consent for their likeness being used, so not only is the fake being made to fool you it’s also victimizing the person whose face (or body[1]) is being used for nefarious purposes.
Probably the scariest consequence when it comes to AI and the propagation of disinformation is that if we’re so busy trying to determine if something is fake, it’s easy for us to start distrusting accurate and reliable information out of hand. AI can undermine our shared perception of what is trustworthy and what is not. If we can’t agree fundamentally on what makes information reliable, our foundation for understanding our world and one another begins to crumble. Developing discernment so that we can separate true from false and reliable from unreliable is already a crucial social skill and will continue to be in the future.
Learn more
Google Is Paying Reddit $60 Million for Fucksmith to Tell Its Users to Eat Glue. 2024. Jason Koebler. (404 Media) Last accessed 2024/07/25.
Tracking AI-enabled Misinformation: 1,000 ‘Unreliable AI-Generated News’ Websites (and Counting), Plus the Top False Narratives Generated by Artificial Intelligence Tools. 2024. NewsGuard. Last accessed 2024/07/25.
How to clock AI. 2024. Elyse Michele Cizek. (TikTok) Last accessed 2024/07/25.
Correction log
[1] Reading this article about deepfakes and their impact on sex workers helped me understand that focusing on the face of the person being impersonated causes us to miss the fact that people's bodies may also be used without consent by people creating deepfakes.
Laws About Deepfakes Can’t Leave Sex Workers Behind. 2024. Samatha Cole. (404 Media) Last accessed 2024/07/25.