CHANUKAH 5786 Sec. B, P9
Every day seems to bring a fresh headline about how AI chatbots are spreading hateful ideas. But researchers tasked with understanding anti-Semitism and how it can be stopped say they have found evidence that AI chatbots can fight hate.
ADL’s Center for Antisemitism Research trained a large-language model — LLM — on countering antiSemitic conspiracy theories, then invited people who subscribed to at least one of them to interact with it.
The result, according to the ADL-sponsored study:
The users soon believed in the anti-Semitic theories less, while also feeling more favorable about Jews as a group.
The effects were still strong a month later, even without further engagement with the LLM.
The researchers call the finding as a breakthrough in the quest for strategies against Jew-hatred.
“What’s remarkable about these findings is that factual debunking works even for conspiracy theories with deep historical roots and strong connections to identity and prejudice,” said David Rand, a Cornell University professor, the study’s senior author.
“Our artificial intelligence debunker bot typically doesn’t rely on emotional appeals, empathy-building exercises, or anti-bias tactics to correct false beliefs,” Rand explalined. “It mostly provides accurate information and evidence-based counterarguments, demonstrating that facts still matter in changing minds.”
Matt Williams, who has headed the ADLCenter for Antisemitism Research since its founding three years ago, says the study builds on a growing body of research that views contemporary anti-Semitism as a misinformation problem, rather than a civil rights problem.
“It’s not ‘Jews’ that are the problem. It is ‘the Jew’ as a function of a conspiracy theory that is the problem. And the relationship between ‘Jews’ and ‘the Jew’ in that context is far more tenuous than we might want to think.”
Calling conspiracy theories “malfunctions in the ways that we make truth out of the world,” Williams said the study showed that “people can correct those malfunctions.
“They really can, which is super exciting.”
The study emerges from the ADL’s relatively new effort to devise evidence-based ways to reduce anti-Semitism, working other researchers to design and carry out experiments aimed at turning a robust advocacy space into less of a guessing game.
The new current study, conducted earlier this year, involved more than 1,200 people who said on a previous ADL survey that they believed at least one of six anti-Semitic theories, such as that Jews control the media or the “Great Replacement” theory about Jewish involvement in immigration.
The people then were randomly assigned three different scenarios:
• A third chatted with an LLM programmed by the researchers to debunk such theories, built within Microsoft’s Claude AI model;
• A third chatted with Claude about an unrelated topic;
• A third were simply told that their belief represented a “dangerous” conspiracy theory.
Then they were all tested again about their beliefs.
Members of the group that chatted with what the researchers are calling DebunkBot were far more likely than members of the other two groups to have their anti-Semitic beliefs weakened, the researchers found.
DebunkBot was not a panacea for anti-Semitism. The study found that those who believed in more antiSemitic conspiracy theories experienced less change.
Williams notes that the study found only that anti-Semitic beliefs were reduced, not rooted out entirely.
But he said any strategy that can cut against what researchers believe has been a widespread explosion of anti-Semitic beliefsis a good thing.
The proportion of Americans subscribing to conspiracy theories over the last decade has reached as much as 45%, more than twice the rate that had held steady for 70 to 80 years, Williams said.
“I don’t think that we’re going to ever create a world in which we go under 15% — but going from 45 back to 30 or 25 seems more doable.”
The new study comes as AI models vault into widespread use among Americans, raising concerns about their implications for Jews.
When Elon Musk launched a model of his own earlier this year called Grok, it was criticized for amplifying anti-Semitism — kicking off a pattern that has played out repeatedly.
The company apologized and said it would train its model to avoid the same behavior in the future.
Criticism of Grok remains, but it no longer praises Hitler — though even this week it reportedly told one user that the Nazi gas chambers were not designed for mass killing, prompting an investigation by French authorities.
Chatbot training is seen as essential for delivering high-quality AI results. DebunkBot can be found online on its own website now.
Williams said ADL is trying to convince the companies operating major AI platforms to incorporate its expertise.
Whatever happens with that effort, Williams said, the new research demonstrates that combatting what’s sometimes called the world’s oldest hatred is possible.
“AI and LLMs — those are tools, right? And we can use tools for good and for evil,” he said.
“But the fact that we can subject conspiracy theories to rational conversation and arguments and actually lead to favorable outcomes is itself, I think, relatively innovative, surprising and extraordinarily useful.”
