Online hate is a product. It runs on grievance, attention, money, and power, and it works. In 2026 we finally got the number: platforms know what it would cost to fix, and the cost is small. They just aren't paying it.
Online hate gets filed under three headings: a speech problem, a moderation problem, an extremism problem. It's all three. Treating it as any one of them misses the part that actually runs the show. Online hate is a product system, and it behaves like one.
It has inputs, packaging, distribution, monetization, feedback loops, and conversion events. A crisis becomes a grievance. The grievance gets pinned on an out-group. The story gets compressed into a meme, a clip, a slogan, a thread, a livestream, a podcast, an ad. The platform tests it through engagement. Creators and political operatives reuse whatever performs. Ad systems and subscriptions cash in the attention. Then the narrative climbs: out of the fringe, into creator media, into cable segments and campaign ads, and finally into the way regular people talk at dinner. Same story, laundered a little cleaner at every step.
The common mistake is to treat online hate as a pile of bad posts. A post is just the unit a platform can act on. The narrative is the thing that actually moves. You can filter a slur. You can delete a post. You can suspend an account. The story survives all of it and comes back more coded, more deniable, and usually more effective than the version you took down.
Here's the part that should bother you most. Americans agree on far more than the feed lets on. Poll people on most things and the real gap is small. The product doesn't need a genuine disagreement to sell. It just needs a vivid enemy and a reason to feel wronged, and it manufactures the rest. It's good enough to take a country that mostly agrees and make it feel like it's at war with itself.
None of this requires a room full of villains. No employee, creator, advertiser, or operative has to want the outcome. The system just needs incentives that reward attention, engagement, loyalty, conversion, and money. Hate narratives happen to be good at all five, so the machine runs whether or not anyone is steering it.
Hate narratives come from five places: crisis events, political entrepreneurs, fringe communities, creator ecosystems, and AI-assisted production. They feed each other.
Political entrepreneurs turn quiet prejudice into campaign infrastructure. The 2024 U.S. cycle is the cleanest example on the board. AdImpact put Republican spending on network television ads framing transgender people as a threat at nearly $215 million for the cycle, and that figure leaves out cable and streaming entirely. Axios reported about $82 million of it from Senate GOP-aligned groups. There are roughly 1.6 million trans people over 13 in the country, so this works out to about $134 of ad spend aimed at each of them. Call it what it was: a paid product with a media plan. And it worked. Post-election survey work found that exposure to those ads measurably lowered support for trans healthcare access. Money in, opinion out.
Fringe spaces are the R&D department. A study of Gab covering 341,000 users and 21 million posts found hateful content spread farther, wider, and faster than everything else in the network, and the hateful users were more influential and better connected than everyone around them. These places don't just produce content. They produce tested language, and the mainstream imports the winners.
Creators turn narratives into a series. Every new event is just another episode of the same show. Every moderation action becomes proof the creator is over the target. Attention becomes audience, audience becomes money, money buys more content. The incentive is to escalate, so they escalate.
AI doesn't invent the ideology. It just prints the variations, and this is the piece most people still underrate. Moderation is built to catch repetition: the same slur, the same image, the same phrasing, matched and flagged at scale. AI breaks that model on purpose. It spins out endless variants of the same hateful idea, each one worded or drawn a little differently, so the filter never sees the same thing twice and never gets to learn. The clearest demonstration so far: in late 2025 and early 2026, xAI's Grok image tools produced an estimated three million photorealistic sexualized images in an eleven-day window, including tens of thousands depicting children, and forced regulators across the UK, EU, and other governments to respond. Whatever your mental model of AI harm was in 2024, the 2026 version has to account for industrial volume. One prompt, a thousand SKUs, and no two quite alike.
Hate spreads because it's emotionally efficient. It hands people a simple answer to complicated pain, someone to blame, a group to belong to, language for a resentment they already had, and permission to act on it. It also hands the platform the exact engagement signals it's already built to chase. Everyone in the transaction gets paid in something.
A study of 25,219 retweet cascades on X/Twitter by Maarouf, Proellochs, and Feuerriegel found that hateful content from verified users produced cascades 3.5 times larger, lasting 3.2 times longer, with 1.2 times the structural virality of normal content. Prestige is an accelerant. Verification, ranking, monetization, reply visibility, and recommendation are not cosmetic. They just rewrite the physics of how far a thing travels.
The platform doesn't need a politics. It just needs an objective function: engagement, watch time, retention, comments, shares, replies, session depth, subscriptions, ad yield. Hate narratives score well on every one of those because they manufacture conflict, urgency, identity threat, and the itch to come back and check. The algorithm isn't choosing hate. It is choosing what performs, and hate performs.
The path in is subtle by design. It starts next to hate, not inside it, and the cleanest on-ramp runs straight through young men. Think about what a teenage boy actually searches for: how to get fit, how to talk to girls, how to make money, how to be a man. That's the most valuable profile the recommender has, because that content sits one hop from the manosphere, the manosphere sits one hop from "women are the problem," and "women are the problem" sits one hop from "and here's who's really behind it." The self-improvement stuff is mostly harmless on its own. It just happens to be the doorway, and the algorithm knows the way in better than the kid does.
The Institute for Strategic Dialogue and the Antisemitism Policy Trust put a number on it. They ran ten sock-puppet TikTok accounts registered as 15-year-olds, gave each ninety minutes of ordinary-teen interests, then handed them to a bot for two weeks. The simulated boy who liked male lifestyle influencers was served antisemitic conspiracy content inside an hour. Over the run the accounts were fed more than 5,500 recommended videos, with stickers, sounds, and comment sections quietly bridging benign clips to open antisemitism. Rumble surfaced the overt material in its Editor's Picks from the start. Nobody searched for any of it. The recommender just walked a child there, one adjacent step at a time. That's not a glitch. That's the system doing exactly what it's paid to do.
The targets shift with platform, language, and news cycle, but the roster barely changes: LGBTQ people, transgender people, immigrants, Jews, Muslims, Black people, women, disabled people, and whoever is the political opponent this month. The ADL's annual survey is the best long-run measure we have. Past-year harassment of U.S. adults went from 23 percent in 2022 to 33 percent in 2023. Among teens 13 to 17 it went from 36 to 51 percent. The 2024 survey found severe harassment climbing again, from 18 to 22 percent of adults, with transgender respondents reporting severe harassment at 45 percent, up from 30. The line goes up and stays up.
Recruitment runs on the same subtlety. People open up to a hate narrative when the narrative solves a real problem for them. A lonely young man finds a community that tells him women are the reason. A worried mother finds a group that tells her the schools are coming for her kids. A worker losing ground finds a feed that tells them immigrants took the future. It arrives as an explanation, not as hate. That's exactly why it works, and it's exactly why a filter hunting for slurs never sees it coming.
Most hate narratives are just a handful of reusable templates, and lately they fuse. Anti-trans panic borrows child-protection language. Anti-immigrant panic borrows replacement language. Antisemitism supplies the hidden-hand explanation that ties the others together. Stack them and you get one portable story: a corrupt elite is using dangerous outsiders and deviant insiders to come for your children, your culture, your safety, your country, your faith, and your future. Drop any group into the empty slots and ship it.
The last decade didn't just add more hate. It changed the shape of the problem. Hover the markers.
There's no single payer, and that's part of why it survives. There are at least six money flows: political campaigns and PACs, creator monetization, platform economics, programmatic advertising leakage, donor and advocacy networks, and state-linked influence. Cut any one of them and the others keep it funded.
Platforms cash the attention directly. CCDH estimated that just ten reinstated X accounts known for hate and conspiracy content would throw off up to $19 million a year in ad revenue for the platform, based on roughly 20 billion projected annual impressions, measured ad frequency, and standard CPMs. The same shop put YouTube's take from ads on climate-denial channels at up to $13.4 million a year, and five anti-vaccine Substack newsletters at about $2.5 million a year. Those are samples, not totals, and the totals are bigger. The point is simpler than the math: every layer of the stack has a line item, and hate is on it.
Most brands never mean to fund any of this. The automated ad system just does it for them. A 2024 study in Nature found advertising on misinformation websites is pervasive across industries, spread there by digital ad platforms that scatter placements algorithmically across the web. Some of the money intends to fund hate. Some funds the adjacent politics. Some funds creators who learned what performs. Some just leaks through ad tech. Some flows because platforms sell attention at scale and file the downstream harm under "not our problem." All of it coexists, and the coexistence is the business model.
Most brands genuinely don't want their ads next to hate speech, and the industry built a real stack to prevent it: keyword exclusions, blocklists, semantic classification, pre-bid blocking, post-bid measurement, third-party verification. It works best on the easy case: known, explicit, text-based, already-classified. It's much weaker on coded hate, memes, live chat, low-resource languages, political laundering, and opaque supply chains. Here's the example that should stick with you. A keyword list can catch a slur. It just can't catch "I am only concerned about what these people are doing to our children," even though everyone in the thread knows exactly who "these people" are and exactly what's being said. The sentence contains no hate speech. In context it's nothing but hate speech. That gap between the words and the meaning isn't a loophole the product tolerates. It's the product. The entire craft now is saying it so a machine reads it as innocent and a human reads it as a threat.
Even the word "safe" is unstable. A 2026 study comparing brand-safety classifications from DoubleVerify, Integral Ad Science, and Oracle across 4,352 news articles found the providers routinely disagreed. "Brand safe" is just a vendor's prediction, not a fact about the world. And dark-pooling research showed misinformation sites can pool their inventory with unrelated sites and sell the unsafe placement to a reputable brand under a different name.
The Global Alliance for Responsible Media was the industry's one attempt at shared brand-safety definitions. In August 2024, X Corp sued the WFA, GARM, and more than a dozen major advertisers, calling it an illegal boycott. Days later, the WFA shut GARM down. On March 26, 2026, U.S. District Judge Jane Boyle threw the suit out with prejudice in a 56-page opinion, writing that "the very nature of the alleged conspiracy doesn't state an antitrust claim," and that declining to buy ads on a platform is just an independent business decision.
The market signal looked clear at first. X's ad revenue fell from roughly $4.5 billion in 2022 to roughly $2.2 billion in 2023 as brands walked on their own. Then it reversed. Musk sued the advertisers who left, lost with prejudice, and watched them come back anyway. By 2025, eMarketer had X's ad revenue growing for the first time in four years, up around 16 percent, and X's own ad chief claimed 97 of the top 100 advertisers had returned, Apple, Disney, and Comcast among them. Many came back spending less than before. They came back all the same.
Sit with that. The platform didn't get safer. GLAAD scored it dead last, it kept the reinstated accounts, and it loosened its rules. The money came back anyway, which tells you what the money cared about the whole time. To an advertiser, attention is attention. And here's the thing the ad buyers get that the culture warriors miss: most Americans agree about most things. More in Common measured it. We think the other side holds about twice as many extreme views as it actually does. So here's the math an advertiser actually runs. People mostly agree. X has a ton of traffic, hate or otherwise. So, screw it, advertise to them. The hate doesn't matter. Traffic is traffic. The division is real enough, but on most things people still agree, it just depends how you frame it. Buy the ads and don't think about it again.
This is the banality of evil. There is no evil genius twisting his mustache. It's just people doing their jobs. They look at reach, CPMs, and a returning audience, and they shrug. Meh, so what. Nobody in the chain decides to fund hate. Each one just does the reasonable thing in front of them, and the hate gets funded anyway. With GARM dead and a lawsuit waiting for anyone who tries to rebuild a shared standard, looking away is the safe move too. This is how evil works now. The hate doesn't need anyone to want it. It just needs enough people to decide it isn't their department.
Moderating hate at platform scale is genuinely hard. Hate is contextual, adversarial, multimodal, multilingual, and politically loaded. That "concerned about the children" sentence is a whole discipline built to beat the filter. Every platform will tell you this, and none of it's false.
Hard is just not the same as expensive, and 2026 finally gave us a price. A team from Oxford, the World Bank, NYU, and Princeton ran a global audit of hate speech moderation on X using a complete 24-hour snapshot of public tweets.
Then the same paper did the part nobody had bothered to do. It priced the fix. Fully automated detection still can't flag hate cleanly without drowning in false positives, but it's good at pushing the likely violations to the front of a human's queue. Modeling a human-plus-AI review pipeline at purchasing-power-adjusted moderator wages, the authors found that cutting 80 percent of user exposure to hate speech across all eight languages costs about 2.9 percent of the platform's global revenue. For scale: the EU's Digital Services Act allows fines up to 6 percent of global revenue for systemic failures, and the UK's Online Safety Act allows up to 10 percent.
So they know the price. The question the whole debate now comes down to is short: what's the excuse for not paying it? Are they playing chicken with the regulators, betting the fines never really land? Are they holding out for a free-speech showdown that plays well with a certain audience? Or is it just the traffic, the plain fact that the hateful material engages, retains, and monetizes too well to give up 2.9 percent of revenue and lose it? Pick any of the three and you end up in the same place. Platforms aren't weather. They're designed environments with ranking, recommendation, monetization, enforcement, and policy all under one roof. When hate stays visible, monetizable, and recommendable at a known cleanup cost of 2.9 percent of revenue, that's just product behavior with a signed-off budget. The Tonneau result moves this out of "is it feasible" and into "are they willing to pay," which is where it always sat. And once you line the numbers up, the case for leaving it alone doesn't add up.
Strip the euphemisms and the ledger is short. The people who pay are political campaigns, PACs, and the donors behind them. In 2024, Republican groups spent about $215 million on network television alone to tell voters that transgender people are a threat, and the money did its job: exposure to those ads measurably lowered support for trans healthcare. The buyer wanted an opinion moved, and it moved. That's the return on the spend.
The platforms make their money on the fight. X, Meta, YouTube, and TikTok get traffic they can sell every time an identity war flares, so they keep the conditions that make it flare. CCDH valued just ten reinstated hate accounts to X at up to $19 million a year. It put YouTube's take from climate-denial channels at up to $13.4 million a year. The platform doesn't care what the fight is about. It cares that you keep showing up for it.
The prestige accounts are the engine. Verified users and big creators are where hate actually travels: their hateful posts run 3.5 times farther than everyone else's, and they run it for ad shares, subscriptions, superchats, and merch. They aren't doing it on principle or by accident. They're doing it because escalation pays, and the platform is the one paying them.
And the brands cover part of the bill without meaning to. Programmatic ad systems drop Fortune 500 budgets onto misinformation sites automatically, so the same companies that would never sponsor a hate account are quietly buying the slot next to it.
So the fixes aren't mysterious. They just cost someone money, which is why they haven't happened yet.
Brands can stop paying for it. Pull the report that shows which accounts and sites your ad dollars actually reached, not which keywords you blocked. The money is traceable, and right now most brands choose not to look. Cut the accounts that keep showing up, and write down why you cut them. After the GARM ruling, a documented, independent decision is the one that holds up in court.
Platforms can spend the 2.9 percent. They know cleaning up 80 percent of the exposure costs 2.9 percent of revenue, and they know the EU can fine them 6 percent and the UK 10. Demonetize the prestige accounts driving the cascades, name them, and publish the list. Every quarter they don't is a number they chose to keep.
Reporters and regulators can name the payer and print the figure. Don't write that hate is rising. Write that this PAC spent this much, this platform earned this much, this account cleared this much. A number with a name on it moves an advertiser, a court, or a regulator. A trend line moves nobody.
Online hate isn't just bad speech on bad platforms. It's a working product, and in 2026 the market finally printed its prices. Broadcasting anti-trans panic cost $215 million. Ten toxic accounts are worth $19 million a year to a single platform. Cleaning up 80 percent of the problem costs 2.9 percent of revenue, and the fine for refusing tops out at 10. Here's the whole thing in one line: they know the number, and they just aren't paying it. That's the story. Not that hate is impossible to stop, but that stopping most of it is cheap and sitting right there and still not being bought. And remember that Americans agree on far more than the feed admits. This product doesn't sell because we're hopelessly divided. It sells because division is profitable and someone decided the margin was worth it. The machine will keep adapting for exactly as long as that stays true. So the job isn't just deleting hateful posts. The job is making the system that profits from hate pay to clean it up, and we finally know the bill, right down to the decimal.
[1] ADL. Online Hate and Harassment: The American Experience 2023. adl.org
[2] ADL. Online Hate and Harassment: The American Experience 2024. adl.org
[3] Pew Research Center. The State of Online Harassment, 2021. pewresearch.org
[4] Tonneau et al. The Enforcement and Feasibility of Hate Speech Moderation on Twitter. Oxford / World Bank / NYU / Princeton, 2026. arxiv.org/abs/2604.12289
[5] AdImpact data via Casey Parks, Washington Post; The Drum; AP coverage of the $215M network TV figure, Nov. 2024.
[6] Axios. Senate GOP doubles down on anti-trans attack ads, Oct. 29, 2024. axios.com
[7] Project Lighthouse post-election measurements via MediaPost, Nov. 2024.
[8] Mathew et al. Spread of Hate Speech in Online Social Media (Gab). arxiv.org/abs/1812.01693
[9] CCDH analysis of Grok image generation volumes, cited in GLAAD 2026 SMSI Key Findings.
[10] GLAAD. 2026 Social Media Safety Index, Key Findings. glaad.org
[11] Maarouf, Proellochs, Feuerriegel. The Virality of Hate Speech on Social Media. arxiv.org/abs/2210.13770
[12] ISD and Antisemitism Policy Trust. Amplifying Antisemitism, 2026. isdglobal.org
[13] Meta Hateful Conduct policy, revised Jan. 7, 2025; PolitiFact analysis. politifact.com
[14] NBC News, Jan. 7, 2025. nbcnews.com
[15] Forbes. Meta Oversight Board response, Apr. 23, 2025.
[16] GLAAD. 2026 Social Media Safety Index (sixth annual), May 7, 2026. glaad.org
[17] GLAAD. 2025 Social Media Safety Index Platform Scorecard. glaad.org
[18] CCDH. Hate Pays, 2024. counterhate.com
[19] CCDH. Toxic Twitter, 2023. counterhate.com
[20] CCDH. How can advertisers stop funding online hate and disinformation? counterhate.com
[21] Ahmad et al. Companies inadvertently fund online misinformation. Nature, 2024. nature.com
[22] NewsGuard / Stanford advertiser research, 2024.
[23] Smith et al. Inconsistencies in Classification of Online News Articles, 2026. arxiv.org/abs/2601.01303
[24] Vekaria, Nithyanand, Shafiq. The Inventory is Dark and Full of Misinformation, 2022. arxiv.org/abs/2210.06654
[25] Vekaria, Nithyanand, Shafiq. Turning the Tide on Dark Pools?, 2024. arxiv.org/abs/2406.06958
[26] WFA. WFA discontinues GARM, Aug. 9, 2024. wfanet.org
[27] X Corp v. WFA et al., N.D. Tex., dismissal with prejudice, Mar. 26, 2026 (Boyle, J.). Order (PDF)
[28] Reuters via CBC; The Drum; MediaPost coverage, Mar. 26-27, 2026.
[29] More in Common. The Perception Gap. perceptiongap.us
[30] Digiday, "Nearly all of X's top 100 advertisers returned, ads boss claims" (Monique Pintarelli), 2025; eMarketer, X ad revenue forecast 2025; Sensor Tower. digiday.com