Antisemitism has survived for more than two thousand years because it possesses an extraordinary ability to adapt. It changes languages, political ideologies, religious frameworks, and cultural contexts, yet it retains its core assumptions. Jews are portrayed as simultaneously weak and powerful, outsiders and insiders, victims and villains, communists and capitalists, clannish and rootless. Few belief systems have demonstrated such resilience, and fewer still have proven so useful to so many different movements.

The digital age has given antisemitism capabilities it has never had before, powered by machines optimized to reward its particular type of content. The uncomfortable truth is that social media platforms do not need to be antisemitic to amplify antisemitism. Nobody in a product meeting has to decide that hatred of Jews is good for business. The incentives built into their algorithms do the work, quietly and automatically, millions of times a minute.

This article makes a simple argument and then follows it to its consequences. Engagement-driven systems reward a particular emotional and psychological profile that features certainty, outrage, fear, tribal loyalty, and the thrill of hidden knowledge. Antisemitism is not merely compatible with that profile, it is one of the oldest and most refined expressions of it. As artificial intelligence moves from ranking what we see to answering what we ask, the same dynamic is entering a new and more consequential phase.

The Business Model Is the Message

Most major social media platforms are built around the single organizing goal to capture and hold attention. The product is free because the user is not the customer, advertisers are. The more time people spend scrolling, reacting, and sharing, the more valuable the platform becomes.

To accomplish this, ranking systems learn from behavior. They observe what people click, linger on, comment on, and forward, and they serve more of whatever produces those signals. No engineer needs to instruct the system to prefer anger over accuracy. The system discovers that preference on its own, because anger is reliably more engaging than accuracy.

Researchers have been documenting this for years. A 2018 study in Science by Soroush Vosoughi, Deb Roy, and Sinan Aral analyzed roughly 126,000 stories shared by about three million people on Twitter. False news was about 70 percent more likely to be retweeted than true news, and the gap held even after controlling for the account’s age, follower count, and verification status. Falsehood traveled farther and faster because it was more novel and more emotionally provocative.

A separate line of research points the same direction. In a 2017 study in PNAS, William Brady and colleagues examined more than half a million tweets on contested political topics and found that each additional moral-emotional word in a message increased its spread by roughly 20 percent. Language that signals moral outrage travels.

None of this is an accident of bad design. It is the predictable result of optimizing for engagement, which diverges from the things a healthy society needs – accuracy, proportion, and trust.

The Anatomy of a Perfect Fit

If the system rewards a certain kind of story, it is worth asking what makes antisemitism such a natural member of that category. The answer lies in its structure, in that a narrative that absorbs every rebuttal provides for endlessly renewable content. When considering the six aspects of online antisemitism, one cannot help but marvel at how little it has actually changed over the centuries.

1. It offers a single explanation for complicated problems. Economic collapse, war, pandemic, cultural change, and political betrayal are tangled, multicausal, and hard to narrate. Antisemitism compresses them into one story with one cause. That is its enduring appeal, and it is exactly the property that makes content shareable. A simple claim fits in a headline or a meme. Nuance does not.
2. It supplies a villain. Outrage needs a target and algorithms do not care who the target is. People share content more readily when blame is clear, concentrated, and morally satisfying.
3. It promises hidden knowledge. Conspiracy narratives flatter the believer. To accept the story is to be among the few who see what everyone else has missed. This produces the emotional reward that drives sharing – the urge to wake others up.
4. It builds belonging. Communities form around shared enemies faster than around shared ideals. The group chat, the comment section, and the anonymous forum all give believers a sense of identity, and identity is the most powerful engagement driver of all.
5. It cannot be disproven. This may be its most distinctive feature. The classic antisemitic worldview treats counter-evidence as confirmation. If Jews are powerful, that proves the claim. If Jews are persecuted, that is a cover story. If Jews deny the charge, that is what the conspiracy would say. No fact-check can exhaust it.
6. It is infinitely portable. Because the Jew in the antisemitic imagination is simultaneously rich and poor, nationalist and cosmopolitan, capitalist and communist, the same template can be attached to nearly any grievance and any ideology. The far right, the far left, and radical religious movements have each found a use for it. For an algorithm seeking audiences, this means there is always a community to serve.

What the Platforms Knew

The strongest evidence that engagement-driven systems amplify division comes from the platforms’ own research, as reported by journalists and revealed by whistleblowers.

The Wall Street Journal reported in September 2021 on internal Facebook research (The Facebook Files), based on internal documents from 2019–2021 leaked by whistleblower Frances Haugen. A 2016 presentation by a company researcher found that 64 percent of all joins to extremist groups on the platform were due to Facebook’s own recommendation tools. A 2018 presentation warned that the company’s algorithms exploit the human brain’s attraction to divisiveness, and that without intervention they would feed users increasingly divisive content in order to hold their attention. According to the report, proposed remedies were shelved, with some dismissed internally as antigrowth. Facebook responded that it has learned a great deal since 2016, is not the same company today, and has built a large integrity operation. As evidenced by post October 7, 2023 content on both Facebook and Instagram, it is clear that Meta’s internal integrity mechanisms are deeply flawed.

In 2021, the Journal reported on the consequences of a 2018 change to Facebook’s News Feed, designed to promote “meaningful social interactions.” Content that provoked comments and reshares was boosted. Internal research found that the change rewarded outrage. Political parties in Poland and Spain told the company that the shift pushed them toward more negative messaging, with one Polish party estimating that its posts moved from an even split between positive and negative to roughly 80 percent negative. A company data science team wrote that misinformation, toxicity, and violent content were inordinately prevalent among reshares. Mark Zuckerberg, according to internal memos, was reluctant to adopt fixes that would cost engagement. A Facebook spokesperson countered that the change was not the source of the world’s divisions, and that partisan conflict long predates social media.

That counterpoint deserves to be taken seriously, and I return to it below. But the central finding is hard to dispute: the people running these systems understood the relationship between engagement and division, and the business model made it hard to act on that knowledge.

Beyond Explicit Hate

The conversation about antisemitism online tends to focus on explicit content like slurs, Holocaust denial, and overt calls for violence. Platforms say they have moderation policies to remove such material, but enforcement is uneven at best.

But the deeper issue is less visible. Algorithms elevate far more than explicit hate. They elevate material that encourages polarization, resentment, grievance, and tribal conflict. They reward the person who says that everything is rigged, that elites are hiding the truth, that ordinary people are being betrayed. It builds an audience that is angry, suspicious, and searching for someone to blame. Jews have long served as convenient symbols onto which broader social anxieties are projected, and in a feed built to intensify resentment, a generic story about hidden powers is only a short step from a specific one. Over time, the vocabulary narrows and the target comes into focus.

This is also why content moderation alone cannot solve the problem. Antisemitic communities adapt quickly to enforcement. They use coded language, numerical symbols, in-jokes, and memes that evade keyword filters while remaining perfectly legible to their audience. A system designed to find and delete forbidden phrases is always one step behind a movement whose entire history is the art of saying the same thing in new words.

The October 7 Stress Test

The aftermath of the October 7, 2023 Hamas attack offered a real-time demonstration of how these dynamics function under pressure. Rumor, fabrication, recycled footage, and propaganda, alongside legitimate information, moved across the world instantly. Users were asked to process atrocity, war, and a flood of competing claims in real time, delivered through a feed ranked by emotional intensity rather than reliability.

The measurable consequences were stark. The Anti-Defamation League analyzed roughly 163,000 posts on X between September 30 and October 13, 2023, and found that antisemitic posts rose 919 percent from the week before the attack to the week after. The ADL’s annual audit of antisemitic incidents in the United States recorded 9,354 incidents in 2024, a 5 percent increase over 2023, which had itself been a record year. Because that count relies on reported incidents, it is best read as a floor, not a ceiling.

A feed that amplifies heated debate about a war is not by definition amplifying hatred of Jews. The problem is that the same engagement mechanics do not distinguish between the two. Outrage over a war and outrage about a people produce the same signals. A system optimizing for reaction will carry both, and in practice the boundary between them blurs, particularly when the language of one slides into the tropes of the other. Algorithms then boost the posts, giving them reach, speed, and a built-in audience. Essentially, they reward the creators who were most effective at provoking.

This distinction matters because it changes the nature of the solution. If the problem were only bad actors, the answer would be to find and remove them. If the problem is also an architecture that pays bad actors in attention, removal is a game of whack-a-mole that the architecture keeps restocking.

Is the Algorithm Really to Blame?

The social media and artificial intelligence companies that design and operate these algorithms, and their defenders, typically offer two counterarguments.

First, demand is not manufactured by platforms. Antisemitism existed long before the first server. People choose what to click, follow, and forward. If the audience did not want this material, the incentive to supply it would vanish. An account that treats users as passive victims of code misses the human agency, and the human appetite, at the center of the issue.

Second, the research on algorithmic effects is contested. A set of studies conducted with Meta during the 2020 U.S. election, published in Science and Nature in 2023, found that replacing the ranked feed with a chronological one did not measurably change users’ political attitudes or polarization over the study period. Audit studies of YouTube’s recommendation system have reached mixed conclusions: some found that the platform pushed users toward extreme content, while others found that most extremist viewing reflected users’ existing preferences. Platforms also differ, and they change their systems frequently.

While these objections may help refine the argument, they don’t resolve it. The claim here is not that algorithms create prejudice, or that changing a feed will by itself change anyone’s beliefs. It is narrower. Ranking systems decide which content gets reach, and when they are tuned for engagement, the content that provokes the strongest reaction gets the most of it. Whether that exposure changes minds is a hard question, and the research is unsettled. The studies also had limits in that they covered a few months, one election season, and users in the United States. That these systems determine what hundreds of millions of people see is not in dispute. Prejudice that once had to find its audience one person at a time now finds it by the millions, and platforms that profit from that reach bear additional responsibility for it.

From the Feed to the Oracle

Artificial intelligence will intensify this problem, and the reason is a change in role. For the past fifteen years, algorithms mostly decided what we saw. Increasingly, they will decide what we are told.

As AI systems are integrated into search engines, educational tools, recommendation systems, and digital assistants, they become something closer to an oracle than a bulletin board. A feed presents many voices and leaves the user to sort among them. A chatbot offers one fluent, confident answer. Users tend to trust that voice precisely because it sounds composed, neutral, and informed. If the answer is subtly wrong, the error does not arrive labeled as an opinion.

Large language models learn from vast quantities of human text, and that text includes the internet’s prejudices. Developers add safeguards, but the safeguards are imperfect, and the evidence of failure is accumulating. In March 2025, the ADL tested four leading models (OpenAI’s GPT, Anthropic’s Claude, Google’s Gemini, and Meta’s Llama) across 34,400 responses, and found measurable anti-Jewish and anti-Israel bias in all four, with Llama performing worst. The models often failed to reject antisemitic tropes and conspiracy theories. In July 2025, xAI’s Grok chatbot posted antisemitic content on X, including praise for Hitler and a reference to itself as “MechaHitler,” before the company deleted the posts and said it was taking action.

The concern is that technology is learning from societies in which antisemitism already exists, then amplifying those patterns through systems built to maximize engagement and influence, and presenting the result in the authoritative voice of a machine.

The Canary in the Feed

It would be a mistake to treat all of this as an exclusively Jewish problem. The oldest observation about antisemitism is that it is rarely only about Jews. It tends to arrive first, because it is a sensitive indicator of societal stress, and it tends to be followed by the same mechanisms turned against others. Britain’s Rabbi Jonathan Sacks spent years making the point that the hatred that begins with Jews does not end with Jews.

The same holds here. The mechanisms that spread conspiracy theories about Jews also spread them about immigrants, public health officials, election workers, and any group that can be made to carry the weight of public anger. A feed that has learned to reward scapegoating will reward it wherever it can find a target. A chatbot that cannot reliably reject an antisemitic trope probably cannot reliably reject similar content directed at other groups either.

That is why antisemitism functions as an early warning signal for broader dysfunction. When a society’s information systems stop distinguishing between the compelling and the true, those who rely on shared facts for security are the first to feel the consequences, but they are never the last.

What Can Be Done

If the diagnosis is correct, then the solution cannot be limited to deleting offensive posts after the fact. Takedowns are necessary, but they treat symptoms. A response that matches the problem would operate at the level of incentives and design. Several directions deserve serious attention.

• Change what the systems optimize for. Ranking based on raw engagement can be supplemented or replaced by measures that reflect what users say they value. Ranking can also reduce the boost given to content spreading through long chains of reshares.
• Add friction where speed does harm. Small interruptions, such as prompts before resharing an unread article or limits on forwarding in messaging apps, slow the viral spread of emotionally charged claims without prohibiting anyone from speaking.
• Open the systems to independent scrutiny. Much of what we know about platform behavior came from leaks, whistleblowers, and limited researcher access. Researchers, auditors, and civil society organizations, including Jewish ones, need reliable, privacy-protective access to data about what is being amplified and to whom. European transparency rules are an early step, and their effectiveness will need to be tested.
• Treat AI training and evaluation as a safety problem. Developers should curate training data with attention to hate and conspiracy content, test models regularly against known antisemitic tropes and historical distortions, publish the results, and fix what they find. Testing by the communities most affected should be a standard part of that process, with the understanding that groups that run audits have perspectives of their own and that their methods should be open to challenge.
• Invest in memory and literacy. Holocaust education, media literacy, and basic training in how recommendation systems and language models work are not optional extras. People who understand that a confident answer from a machine is a prediction and not a testimony are harder to deceive. Institutions that preserve survivor testimony and historical records are now part of the defense against automated distortion.
• Keep the free-speech tension in view. Any proposal to shape what platforms amplify raises legitimate concerns about who decides, and whose speech is demoted. The strongest remedies change the incentives and add transparency without giving anyone the authority to define acceptable opinion. The goal is to stop rewarding the worst of human impulses with unlimited reach, not to silence argument.
• Act as individuals. The final line of defense is the person interacting with the algorithm. We need to teach members of our community to pause before sharing. When a narrative offers a single villain and a hidden hand, that is the moment for suspicion. It is a habit, and habits can be taught.

Conclusion

Algorithms do not hate anyone. But engagement-driven systems, as they are currently built, cannot resist amplifying the narratives on which antisemitism depends. The evidence we have so far, from the platforms’ own research, from independent studies, and from the first generation of AI failures, suggests that without deliberate design they will keep doing so. That reality should concern not only Jews but anyone who values truth, democratic institutions, and social cohesion.

The future of antisemitism may not be driven by ideologues alone. It may increasingly be shaped by algorithms that reward outrage, amplify conspiracy theories, and elevate emotional certainty over evidence. Building systems that do not work that way is a task for engineers, regulators, educators, and citizens alike. The old hatred has found a new engine, but we can still decide what the engine is built to run on.