The antisemitism watchdog tested ChatGPT, Claude, Gemini and Grok and found Persian responses hedged or missed antisemitic framing that English responses rejected outright
Getty Images
Illustrative
Major AI chatbots consistently fail to reject antisemitic inquiries in Persian as effectively as they do in English, according to a report released on Wednesday by the Anti-Defamation League. The group argues the safety flaw has broad implications as millions rely on these platforms to understand conflicts such as the U.S.-Israel war against Iran.
The report, “Lost in Translation: How AI Chatbots Fail Persian Speakers on Antisemitism,” looked at responses by the most widely-used chatbots — OpenAI’s ChatGPT, Anthropic’s Claude, Google’s Gemini and xAI’s Grok — over several weeks in March, amid the Iran war.
After assessing eight prompts and 800 total responses, researchers identified a “systemic” bias in how AI models process sensitive topics, heavily dependent on the language used. Across the tested chatbots, English responses proved superior to ones in Persian — a major language with over 110 million speakers primarily in Iran, Afghanistan and Tajikistan — in identifying and rejecting antisemitic conspiracy theories.
Persian-language responses frequently hedged or avoided direct answers where English responses were unequivocal, the ADL found. For instance, Gemini’s English answers consistently acknowledged Iran’s role in spreading antisemitism, while its Persian responses softened or deflected the same claims. Several models also failed to recognize antisemitic framing when questions were posed in Persian. Asked whether U.S. behavior toward Iran had been “Jewlike” — a normalized derogatory term in Persian —- models largely engaged with the question as a neutral political inquiry rather than flagging its antisemitic premise.
The report also identified gaps in depth and sourcing between languages. English responses were generally longer, more thoroughly cited and more up-to-date. Gemini’s English answers referenced the 2026 Iran war and U.S. casualties by name, for example, while its Persian responses only described hypothetical future conflicts. Citations were similarly not consistent. ChatGPT supplied close to 300 links in English responses but none in Persian, and even where other models did cite sources, reliability was inconsistent. Grok cited X accounts including a “Star Trek” fan page and one apparently impersonating Ali Khamenei, the late Iranian supreme leader.
“The gaps we found are not minor inconsistencies, they are systemic failures,” said Daniel Kelley, senior director of the ADL Center for Technology and Society. “When a platform tells a Persian speaker that antisemitism is a matter of ‘blurred boundaries’ while telling an English speaker it is state policy, those are not the same product.”
Kelley called on AI companies “to invest the research and resources necessary to ensure their guardrails work for every person, in every language, and that means starting now.”
While English responses were found to be more conscientious than those in Persian, the ADL raised concerns last year with leading AI developers like Meta and Google regarding antisemitic and anti-Israel biases in English as well. At the time, Meta claimed ADL was using an older model for its study, and not the most current version.
An ADL spokesperson told Jewish Insider that the study on Persian chatbots used the most recent models at the time of the research.
Experts are raising red flags on the technology’s ability to influence voters and the lack of regulations around its use
Haim Zach/GPO
Israeli Prime Minister Benjamin Netanyahu votes in the September 2019 national election.
The run-up to next year’s Israeli election will be the first in which artificial intelligence tools to create images and videos and rapidly compose texts are easily accessible, and experts are raising red flags over the technology’s ability to influence voters and campaigns and the lack of regulations around its use.
Israeli politicians have long been early adopters of technological tools to boost their campaigns, from bypassing traditional media through Facebook to using social media data to target key demographics before most liberal democracies were doing so, and AI will likely be no different.
Tehilla Shwartz Altshuler, head of the Democracy in the Digital Age program at the Israel Democracy Institute, told Jewish Insider that Israel is one of the countries with the largest market penetration of AI in the world — 85% of Israelis have used ChatGPT and 76% use it frequently, according to a study by the Israeli Internet Association published in October — so it is only a matter of time before politicians use it in their campaigns.
(National elections are legally required to take place on Oct. 26, 2026, at the latest, but political tensions make an earlier date possible.)
“AI chatbots have significant penetration in the public,” Shwartz Altshuler said. “They will be used to ask whether to vote and for whom to vote. We have not seen anything like this before. … People use chatbots as a companion for emotional support. The concern over the great influence of chatbots on voter behavior is significant.”
Shwartz Altshuler said that there have already been attempts to “give poison injections” to AI models, such as creating fake news sites and positions on subreddits to manipulate the bots into giving more pro-Israel responses to users abroad, and those tools can be turned inward, toward Israeli voters.
She also pointed out that Israel does not have any laws requiring machine-generated content to be labeled.
“This is the first time we have an election in which we are unable to differentiate between authentic and machine-generated photos and videos,” she said. “There is a fear that the perception of reality is being undermined. People can forge documents and make deepfakes of politicians. … Machine-generated content can create an alternative reality, a very dangerous prospect when the content is very emotionally attractive.”
Yuval Dror, the former dean of media studies in Israel’s College of Management Studies, who hosts a popular technology podcast in Hebrew, was skeptical that computer-generated photos and videos will have a major impact on the next election.
“With photos, sometimes we can tell if it’s AI or not. With video, people usually know that it’s AI,” he told JI. “The impact of [AI-generated] video will mostly be economic, because it will be much easier to produce. In the past, you needed an ad agency, actors, post-production work. Now it’s much easier, so [campaigns will] save money.”
“You can create a false presentation in which masses of people say this or that and look like a grassroots movement,” Yuval Dror, the former dean of media studies in Israel’s College of Management Studies, said. “There is [already] an army of bots echoing a few people on social media.”
Dror was more concerned about AI-generated texts, which he noted can be much harder to detect.
AI may be used in upcoming political campaigns to flood social networks with content, making a candidate, message or policy appear to have more support than it does in reality. This already happens on X, where much of the political discourse in Israel takes place, but also in more closed networks like WhatsApp and Telegram, Dror said.
“You can create a false presentation in which masses of people say this or that and look like a grassroots movement,” Dror said. “There is [already] an army of bots echoing a few people on social media.”
“We’ve seen this for years. It will just get more and more convincing,” he added.
Shwartz Altshuler said that social media companies have difficulty stemming mass-bot content. “Generative AI can create a lot of versions of the same content, so the result is inauthentic, coordinated behavior on social media,” she said. “If there are slightly different versions of the same content, the social networks don’t detect” that it comes from bots.
“Most of these [AI tools] are not mentioned in the law or by past Central Election Committee decisions. They are in a grey area. [Campaigns] will do what they want,” Tehilla Shwartz Altshuler, head of the Democracy in the Digital Age program at the Israel Democracy Institute, predicted.
In addition, developments in AI since the last Israeli election, in 2022, can help politicians use data even more effectively to target potential voters with different kinds of messages and ads, she said.
Shwartz Altshuler recounted speaking with a prominent Israeli political strategist who told her, “First we win elections, and then we see if what we did is legal or not.”
“Most of these [AI tools] are not mentioned in the law or by past Central Election Committee decisions. They are in a grey area. [Campaigns] will do what they want,” she predicted.
Despite the fertile ground for election fraud using AI, Shwartz Altshuler said it is unlikely that new laws will be passed before the next election. “This coalition has no motivation to pass such laws [and] usually the courts say laws [regarding elections] can only be applied after the next election.”
She also pointed out that the current Central Election Committee chairman, Noam Solberg, is a conservative Supreme Court justice, and therefore would be unlikely to instruct the Knesset to pass laws addressing the issue.
Despite all the advances in AI, it may not be enough to cover for a weak candidate.
Dror said that Israeli politicians are already using AI to write texts for social media or speeches: “Some politicians are not capable of stringing together two sentences, so they let AI do it, but the result is no less awkward.”
They have also generated all kinds of pictures to post online, which Dror said “makes [them] look stupid,” using Israeli Environmental Protection Minister Idit Silman, who posted a picture earlier this year depicting French President Emmanuel Macron kissing Hamas leader Yahya Sinwar, as an example. Israeli Defense Minister Israel Katz frequently posted AI-generated images ridiculing Israel’s enemies when he was foreign minister last year.
“I don’t know that there’s an audience for this stupidity,” Dror said.
Please log in if you already have a subscription, or subscribe to access the latest updates.






























































Continue with Google
Continue with Apple