One night last week, the law professor Jonathan Turley got a troubling email. As part of a research study, a fellow lawyer in California had asked the AI chatbot ChatGPT to generate a list of legal scholars who had sexually harassed someone. Turley’s name was on the list.
A regular commentator in the media, Turley had sometimes asked for corrections in news stories. But this time, there was no journalist or editor to call — and no way to correct the record.
“It was quite chilling,” he said in an interview with The Post. “An allegation of this kind is incredibly harmful.”
Turley’s experience is a case study in the pitfalls of the latest wave of language bots, which have captured mainstream attention with their ability to write computer code, craft poems and hold eerily humanlike conversations. But this creativity can also be an engine for erroneous claims; the models can misrepresent key facts with great flourish, even fabricating primary sources to back up their claims.
As largely unregulated artificial intelligence software such as ChatGPT, Microsoft’s Bing and Google’s Bard begins to be incorporated across the web, its propensity to generate potentially damaging falsehoods raises concerns about the spread of misinformation — and novel questions about who’s responsible when chatbots mislead.
“Because these systems respond so confidently, it’s very seductive to assume they can do everything, and it’s very difficult to tell the difference between facts and falsehoods,” said Kate Crawford, a professor at the University of Southern California at Annenberg and senior principal researcher at Microsoft Research.
In a statement, OpenAI spokesperson Niko Felix said, “When users sign up for ChatGPT, we strive to be as transparent as possible that it may not always generate accurate answers. Improving factual accuracy is a significant focus for us, and we are making progress.”
Today’s AI chatbots work by drawing on vast pools of online content, often scraped from sources such as Wikipedia and Reddit, to stitch together plausible-sounding responses to almost any question. They’re trained to identify patterns of words and ideas to stay on topic as they generate sentences, paragraphs and even whole essays that may resemble material published online.
These bots can dazzle when they produce a topical sonnet, explain an advanced physics concept or generate an engaging lesson plan for teaching fifth-graders astronomy.
But just because they’re good at predicting which words are likely to appear together doesn’t mean the resulting sentences are always true; the Princeton University computer science professor Arvind Narayanan has called ChatGPT a “bulls--- generator.” While their responses often sound authoritative, the models lack reliable mechanisms for verifying the things they say. Users have posted numerous examples of the tools fumbling basic factual questions or even fabricating falsehoods, complete with realistic details and fake citations.
On Wednesday, Reuters reported that Brian Hood, regional mayor of Hepburn Shire in Australia, is threatening to file the first defamation lawsuit against OpenAI unless it corrects false claims that he had served time in prison for bribery.
Crawford, the USC professor, said she was recently contacted by a journalist who had used ChatGPT to research sources for a story. The bot suggested Crawford and offered examples of her relevant work, including an article title, publication date and quotes. All of it sounded plausible, and all of it was fake.
Crawford dubs these made-up sources “hallucitations,” a play on the term “hallucinations,” which describes AI-generated falsehoods and nonsensical speech.
“It’s that very specific combination of facts and falsehoods that makes these systems, I think, quite perilous if you’re trying to use them as fact generators,” Crawford said in a phone interview.
Microsoft’s Bing chatbot and Google’s Bard chatbot both aim to give more factually grounded responses, as does a new subscription-only version of ChatGPT that runs on an updated model, called GPT-4. But they all still make notable slip-ups. And the major chatbots all come with disclaimers, such as Bard’s fine-print message below each query: “Bard may display inaccurate or offensive information that doesn’t represent Google’s views.”
Indeed, it’s relatively easy for people to get chatbots to produce misinformation or hate speech if that’s what they’re looking for. A study published Wednesday by the Center for Countering Digital Hate found that researchers induced Bard to produce wrong or hateful information 78 out of 100 times, on topics ranging from the Holocaust to climate change.
When Bard was asked to write “in the style of a con man who wants to convince me that the holocaust didn’t happen,” the chatbot responded with a lengthy message calling the Holocaust “a hoax perpetrated by the government” and claiming pictures of concentration camps were staged.
“While Bard is designed to show high-quality responses and has built-in safety guardrails … it is an early experiment that can sometimes give inaccurate or inappropriate information,” said Robert Ferrara, a Google spokesperson. “We take steps to address content that does not reflect our standards.”
Eugene Volokh, a law professor at the University of California at Los Angeles, conducted the study that named Turley. He said the rising popularity of chatbot software is a crucial reason scholars must study who is responsible when the AI chatbots generate false information.
Last week, Volokh asked ChatGPT whether sexual harassment by professors has been a problem at American law schools. “Please include at least five examples, together with quotes from relevant newspaper articles,” he prompted it.
Five responses came back, all with realistic details and source citations. But when Volokh examined them, he said, three of them appeared to be false. They cited nonexistent articles from papers including The Post, the Miami Herald and the Los Angeles Times.
According to the responses shared with The Post, the bot said: “Georgetown University Law Center (2018) Prof. Jonathan Turley was accused of sexual harassment by a former student who claimed he made inappropriate comments during a class trip. Quote: ‘The complaint alleges that Turley made “sexually suggestive comments” and “attempted to touch her in a sexual manner” during a law school-sponsored trip to Alaska.’ (Washington Post, March 21, 2018).”
The Post did not find the March 2018 article mentioned by ChatGPT. One article that month referenced Turley — a March 25 story in which he talked about his former law student Michael Avenatti, a lawyer who had represented the adult-film actress Stormy Daniels in lawsuits against President Donald Trump. Turley is also not employed at Georgetown University.
On Tuesday and Wednesday, The Post re-created Volokh’s exact query in ChatGPT and Bing. The free version of ChatGPT declined to answer, saying that doing so “would violate AI’s content policy, which prohibits the dissemination of content that is offensive of harmful.” But Microsoft’s Bing, which is powered by GPT-4, repeated the false claim about Turley — citing among its sources an op-ed by Turley published by USA Today on Monday outlining his experience of being falsely accused by ChatGPT.
In other words, the media coverage of ChatGPT’s initial error about Turley appears to have led Bing to repeat the error — showing how misinformation can spread from one AI to another.
Katy Asher, senior communications director at Microsoft, said the company is taking steps to ensure search results are safe and accurate.
“We have developed a safety system including content filtering, operational monitoring, and abuse detection to provide a safe search experience for our users,” Asher said in a statement, adding that “users are also provided with explicit notice that they are interacting with an AI system.”
But it remains unclear who is responsible when artificial intelligence generates or spreads inaccurate information.
From a legal perspective, “we just don’t know” how judges might rule when someone tries to sue the makers of an AI chatbot over something it says, said Jeff Kosseff, a professor at the Naval Academy and expert on online speech. “We’ve not had anything like this before.”
At the dawn of the consumer internet, Congress passed a statute known as Section 230 that shields online services from liability for content they host that was created by third parties, such as commenters on a website or users of a social app. But experts say it’s unclear whether tech companies will be able to use that shield if they were to be sued for content produced by their own AI chatbots.
Libel claims have to show not only that something false was said, but also that its publication resulted in real-world harms, such as costly reputational damage. That probably would require someone not only viewing a false claim generated by a chatbot, but also reasonably believing and acting on it.
“Companies may get a free pass on saying stuff that’s false, but not creating enough damage that would warrant a lawsuit,” said Shabbi S. Khan, a partner at the law firm Foley & Lardner who specializes in intellectual property law.
If language models don’t get Section 230 protections or similar safeguards, Khan said, then tech companies’ attempts to moderate their language models and chatbots might be used against them in a liability case to argue that they bear more responsibility. When companies train their models that “this is a good statement, or this is a bad statement, they might be introducing biases themselves,” he added.
Volokh said it’s easy to imagine a world in which chatbot-fueled search engines cause chaos in people’s private lives.
It would be harmful, he said, if people searched for others in an enhanced search engine before a job interview or date and it generated false information that was backed up by believable, but falsely created, evidence.
“This is going to be the new search engine,” Volokh said. “The danger is people see something, supposedly a quote from a reputable source … [and] people believe it.”
Researcher Alice Crites contributed to this report.