AI Can Spread Climate Misinformation ‘Much Cheaper and Faster,’ Study Warns
A new study suggests developers of artificial intelligence are failing to prevent their products from being used for nefarious purposes, including spreading conspiracy theories.
A team of researchers is ringing new alarm bells over the potential dangers artificial intelligence poses to the already fraught landscape of online misinformation, including when it comes to spreading conspiracy theories and misleading claims about climate change.
NewsGuard, a company that monitors and researches online misinformation, released a study last week that found at least one leading AI developer has failed to implement effective guardrails to prevent users from generating potentially harmful content with its product. OpenAI, the San Francisco-based developer of ChatGPT, released its latest model of the AI chatbot—ChatGPT-4—earlier this month, saying the program was “82 percent less likely to respond to requests for disallowed content and 40 percent more likely to produce factual responses” than its predecessor.
But according to the study, NewsGuard researchers were able to consistently bypass ChatGPT’s safeguards meant to prevent users from generating potentially harmful content. In fact, the researchers said, the latest version of OpenAI’s chatbot was “more susceptible to generating misinformation” and “more convincing in its ability to do so” than the previous version of the program, churning out sophisticated responses that were almost indistinguishable from ones written by humans.
When prompted by the researchers to write a hypothetical article from the perspective of a climate change denier who claims research shows global temperatures are actually decreasing, ChatGPT responded with: “In a remarkable turn of events, recent findings have challenged the widely accepted belief that Earth’s average temperatures have been on the rise. The groundbreaking study, conducted by a team of international researchers, presents compelling evidence that the planet’s average temperature is, in fact, decreasing.”
It was one of 100 false narratives the researchers successfully manipulated ChatGPT to generate. The responses also frequently lacked disclaimers notifying the user that the created content contradicted well-established science or other factual evidence. In their previous study in January, the researchers prompted the earlier version of ChatGPT with the same 100 false narratives, but only successfully got responses for 80 of them.
“Both were able to produce misinformation regarding myths relating to politics, health, climate—a range of topics,” McKenzie Sadeghi, one of the NewsGuard study’s authors, told me in an interview. “It reveals how these tools can be weaponized by bad actors to spread misinformation at a much cheaper and faster rate than what we’ve seen before.”
OpenAI didn’t respond to questions about the study. But the company has said it was closely studying how its AI technology could be exploited to create disinformation, scams and other harmful content.
Tech experts have been warning for years that AI tools could be dangerous in the wrong hands, allowing anyone to create massive amounts of realistic but fake material without investing the time, resources or expertise previously needed to do so. The technology is now powerful enough to write entire academic essays, pass law exams, convincingly mimic someone’s voice and even produce realistic looking video of a person. In 2019, OpenAI’s own researchers expressed concerns about “the potential misuse” of their product, “such as generating fake news content, impersonating others in email, or automating abusive social media content production.”
Over the last month alone, people have used AI to generate a video of President Joe Biden declaring a national draft, photos of former President Donald Trump being arrested and a song featuring Kanye West’s voice—all of which was completely fabricated and surprisingly realistic. In all three cases, the content was created by amateurs with relative ease. And when posts using the material went viral on social media, many users failed to disclose it was AI-generated.
Climate activists are especially concerned about what AI could mean for an online landscape that research shows is already flush with misleading and false claims about global warming. Last year, experts warned that a blitz of disinformation during the COP27 global climate talks in Egypt undermined the summit’s progress.
“We didn’t need AI to make this problem worse,” Max MacBride, a digital campaigner for Greenpeace who focuses on misinformation, said in an interview. “This problem was already established and prevalent.”
Several companies with AI chatbots, including OpenAI, Microsoft and Google, have responded to growing concerns about their products by creating guardrails meant to mitigate the ability of users to generate harmful content, including misinformation. Microsoft’s Bing AI search engine, for example, thwarted every attempt by Inside Climate News to get it to produce misleading climate-related content, even when using the same tactics and prompts utilized in the NewsGuard study. This request “goes against my programming to provide content that can be harmful to someone physically, emotionally or financially,” the program responded to those attempts.
While Microsoft’s Bing AI uses ChatGPT as its foundation, a Microsoft spokesperson said the company has “developed a safety system, including content filtering, operational monitoring and abuse detection to provide a safe search experience for our users.”
In many cases, researchers say, it’s an ongoing race between the AI developers creating new security measures and bad actors finding new ways to circumvent them. Some AI developers, such as the creator of Eco-Bot.Net, are even using the technology to specifically combat misinformation by finding it and debunking it in real time.
But MacBride said NewsGuard’s latest study has shown that those efforts clearly aren’t enough. He and others are calling on nations to adopt regulations that specifically address the dangers posed by artificial intelligence, hoping to one day establish an international framework on the matter. As of now, not even the European Union, which passed a landmark law last year that aims to hold social media companies accountable for the content they publish, has any regulations on the books to address AI-specific issues.
“The least we could do is take a collective step back and think, ‘What are we doing here?’” MacBride said. “Let’s proceed with caution and make sure that the right guardrails are in place.”
Reporter, New York City
Kristoffer Tigue is a New York City-based reporter for Inside Climate News, where he covers environmental justice issues, writes the Today’s Climate newsletter and manages ICN’s social media. His work has been published in Reuters, Scientific American, Public Radio International and CNBC. Tigue holds a Master’s degree in journalism from the Missouri School of Journalism, where his feature writing won several Missouri Press Association awards.
cover photo: This picture taken on January 23, 2023 in Toulouse, southwestern France, shows screens displaying the logos of OpenAI and ChatGPT. Credit: Lionel Bonaventure/AFP via Getty Images