Every election presents an opportunity for disinformation to find its way into the public discourse. But as the 2024 US presidential race begins to take shape, the growth of artificial intelligence (AI) technology threatens to give propagandists powerful new tools to ply their trade.
Generative AI models that are able to create unique content from simple prompts are already being deployed for political purposes, taking disinformation campaigns into strange new places. Campaigns have circulated fake images and audio targeting other candidates, including an AI-generated campaign ad attacking Joe Biden and deepfake videos mimicking real-life news footage.
The Guardian spoke with Renée DiResta, technical research manager at the Stanford Internet Observatory, a university program that researches the abuses of information technology, about how the latest developments in AI influence campaigns and how society is catching up to a new, artificially created reality.
Concern around AI and its potential for disinformation has been around for a while. What has changed that makes this threat more urgent?
When people became aware of deepfakes – which usually refers to machine-generated video of an event that did not happen – a few years ago there was concern that adversarial actors would use these types of video to disrupt elections. Perhaps they would make video of a candidate, perhaps they would make video of some sort of disaster. But it didn’t really happen. The technology captured public attention, but it wasn’t very widely democratized. And so it didn’t primarily manifest in the political conversation, but instead in the realm of much more mundane but really individually harmful things, like revenge porn.
There’s been two major developments in the last six months. First is the rise of ChatGPT, which is generated text. It became available to a mass market and people began to realize how easy it was to use these types of text-based tools. At the same time, text-to-still image tools became globally available. Today, anybody can use Stable Diffusion or Midjourney to create photorealistic images of things that don’t really exist in the world. The combination of these two things, in addition to the concerns that a lot of people feel around the 2024 elections, has really captured public attention once again.
Why did the political use of deepfakes not materialize?
The challenge with using video in a political environment is that you really have to nail the substance of the content. There are a lot of tells in video, a lot of ways in which you can determine whether it’s generated. On top of that, when a video is truly sensational, a lot of people look at it and factcheck it and respond to it. You might call it a natural immune response.
Text and images, however, have the potential for higher actual impact in an election scenario because they can be more subtle and longer lasting. Elections require months of campaigning during which people formulate an opinion. It’s not something where you’re going to change the entire public mind with a video and have that be the most impactful communication of the election.
How do you think large language models can change political propaganda?
I want to caveat that describing what is tactically possible is not the same thing as me saying the sky is falling. I’m not a doomer about this technology. But I do think that we should understand generative AI in the context of what it makes possible. It increases the number of people who can create political propaganda or content. It decreases the cost to do it. That’s not to say necessarily that they will, and so I think we want to maintain that differentiation between this is the tactic that a new technology enables versus that this is going to swing an election.
As far as the question of what’s possible, in terms of behaviors, you’ll see things like automation. You might remember back in 2015 there were all these fears about bots. You had a lot of people using automation to try to make their point of view look more popular – making it look like a whole lot of people think this thing, when in reality it’s six guys and their 5,000 bots. For a while Twitter wasn’t doing anything to stop that, but it was fairly easy to detect. A lot of the accounts would be saying the exact same thing at the exact same time, because it was expensive and time consuming to generate a unique message for each of your fake accounts. But with generative AI it is now effortless to generate highly personalized content and to automate its dissemination.
And then finally, in terms of content, it’s really just that the messages are more credible and persuasive.
That seems tied to another aspect you’ve written about, that the sheer amount of content that can be generated, including misleading or inaccurate content, has a muddying effect on information and trust.
It’s the scale that makes it really different. People have always been able to create propaganda, and I think it’s very important to emphasize that. There is an entire industry of people whose job it is to create messages for campaigns and then figure out how to get them out into the world. We’ve just changed the speed and the scale and the cost to do that. It’s just an evolution in propaganda.
When we think about what’s new and what’s different here, the same thing goes for images. When Photoshop emerged, the public at first was very uncomfortable with Photoshopped images, and gradually became more comfortable with it. The public acclimated to the idea that Photoshop existed and that not everything that you see with your eyes is a thing that necessarily is as it seems – the idea that the woman that you see on the magazine cover probably does not actually look like that. Where we’ve gone with generative AI is the fabrication of a complete unreality, where nothing about the image is what it seems but it looks photorealistic.
after newsletter promotion
Now anybody can make it look like the pope is wearing Balenciaga.
Exactly.
In the US, it seems like meaningful federal regulation is pretty far away if it’s going to come at all. Absent of that, what are some of the sort of short-term ways to mitigate these risks?
First is the education piece. There was a very large education component when deepfakes became popular – media covered them and people began to get the sense that we were entering a world in which a video might not be what it seems.
But it’s unreasonable to expect every person engaging with somebody on a social media platform to figure out if the person they’re talking to is real. Platforms will have to take steps to more carefully identify if automation is in play.
On the image front, social media platforms, as well as generative AI companies, are starting to come together to try and determine what kind of watermarking might be useful so that platforms and others can determine computationally whether an image is generated.
Some companies, like OpenAI, have policies around generating misinformation or the use of ChatGPT for political ends. How effective do you see those policies being?
It’s a question of access. For any technology, you can try to put guardrails on your proprietary version of that technology and you can argue you’ve made a values-based decision to not allow your products to generate particular types of content. On the flip side, though, there are models that are open source and anyone can go and get access to them. Some of the things that are being done with some of the open source models and image generation are deeply harmful, but once the model is open sourced, the ability to control its use is much more limited.
And it’s a very big debate right now in the field. You don’t want to necessarily create regulations that lock in and protect particular corporate actors. At the same time, there is a recognition that open-source models are out there in the world already. The question becomes how the platforms that are going to serve as the dissemination pathways for this stuff think about their role and their policies in what they amplify and curate.
What’s the media or the public getting wrong about AI and disinformation?
One of the real challenges is that people are going to believe what they see if it conforms to what they want to believe. In a world of unreality in which you can create that content that fulfills that need, one of the real challenges is whether media literacy efforts actually solve any of the problems. Or will we move further into divergent realities – where people are going to continue to hold the belief in something that they’ve seen on the internet as long as it tells them what they want. Larger offline challenges around partisanship and trust are reflected in, and exacerbated by, new technologies that enable this kind of content to propagate online.