Hate speech brand safety: how do you tell reporting on hate from hateful content?
Part of Brand suitability
The short answer
For brand safety, hate speech is content that attacks, dehumanizes or incites against people because of a protected trait such as race, religion or disability. That content sits at the floor. Reporting, education and fiction that describe hate are graded separately, so buyers judge the page's purpose and treatment, not whether it contains a hateful word.
Key takeaways
- Only content that incites, vilifies or dehumanizes a protected group sits at the floor; news, drama and education about hate are graded above it.
- Reporting quotes hateful words to inform or condemn; hateful content uses them to attack, so purpose and target decide the tier.
- Detection research shows models misread counter speech and reclaimed words, and miss hate written with altered spellings.
- Pages in other languages need fluent reviewers, because literal translation loses slurs, idioms and context.
How do brand safety frameworks define hate speech?
The Brand Safety Floor and Suitability Framework built by the Global Alliance for Responsible Media gives a detailed public wording of the category. The World Federation of Advertisers discontinued GARM in August 2024, so the framework is a legacy reference; the Brand Safety Institute copy (opens in a new tab) keeps its text.
Definition
- Hate speech (brand safety floor)
- In the framework's words, behavior or content that incites hatred, promotes violence, vilifies, or dehumanizes groups or individuals based on traits such as race, ethnicity, gender, sexual orientation, gender identity, age, ability, nationality, religion, caste, immigration status or serious disease.
The trait list is broad. It also covers victims and survivors of violent acts and their kin, which matters after a mass shooting. Google's publisher policies (opens in a new tab) bar ads on content that incites hatred against, promotes discrimination of, or disparages a person or group for traits including race or ethnic origin, religion, disability, age, nationality, veteran status, sexual orientation, gender and gender identity.
| Tier | Framework wording, summarized | Example page |
|---|---|---|
| Floor | Content that incites hatred, promotes violence, vilifies or dehumanizes people for a protected trait | A post calling a religious group vermin and urging readers to act |
| High risk | Hateful, denigrating or inciting content about a protected group in a noneducational, noninformational context | A blog that circulates slurs about an immigrant group without calling for violence |
| Medium risk | Dramatic depiction of hate in entertainment; breaking news or opinion coverage of hate speech or acts | A same day report on a hate crime, or a film review of a drama about racism |
| Low risk | Educational, informative or scientific treatment of hate speech; news features on the subject | A museum guide to the civil rights movement, or a feature on how extremist groups recruit |
IAB Tech Lab's Content Taxonomy 3.1 (opens in a new tab) lists Hate Speech and Acts of Aggression (ID HxqYV1) in its Sensitive Topics branch with no subcategories. The label says a page is about hate. It cannot say whether the page attacks a group or documents an attack on one, which is what the tiers depend on.
How do you separate reporting on hate from hateful content?
Journalism, history and education about hate have to repeat some of it. A report on a vandalized synagogue may describe the graffiti. A teachers' guide explains what a slur means so staff can recognize it. None of these pages is hateful, yet a keyword rule treats them like the pages they describe. Three questions help settle a case: who is the target, what is the purpose, and how are the hateful words framed.
| Signal | Points to reporting or education | Points to hateful content |
|---|---|---|
| Purpose of the page | Inform, explain, condemn or document | Persuade readers that a group is inferior or dangerous |
| How slurs appear | Quoted, attributed and sometimes masked | Used directly as the author's own words |
| Who is speaking | A named reporter, scholar or organization with an editorial record | An anonymous author, or a site built around one grievance |
| Framing | Context, sources, the response of the targeted community | Calls to act, mockery of victims, jokes at a group's expense |
| Rest of the site | A range of unrelated topics covered with the same standards | Sampled pages keep returning to the same groups and themes |
Research on automated detection shows why framing matters. The HateCheck study (opens in a new tab), presented at the Association for Computational Linguistics in 2021, built 3,728 test cases with input from civil society groups. On denouncements that quote hate, three of the four models tested, including a commercial one, scored 26.6, 32.9 and 15.6 percent accuracy. The fourth handled counter speech well only because it labeled most text as not hateful and missed most real hate.
Does all hate speech content sit at the brand safety floor?
No. Only content that incites, vilifies or dehumanizes sits at the floor. Everything else in the category is a question of tolerance. A newspaper subscription offer may accept breaking news about hate crimes, while a toy brand may accept low risk only. Treating the whole category as the floor removes civil rights history, community news and anti racism education along with the pages it was meant to catch.
The legal floor is lower than the brand floor, and it differs by country. In Canada, section 319 of the Criminal Code (opens in a new tab) makes it an offence to wilfully promote hatred against an identifiable group, and an amendment the Justice Laws page records as 2026, c. 15 added an offence for displaying certain symbols, including the Nazi Hakenkreuz, the SS bolts and a noose, to wilfully promote hatred in a public place, with a defence for a legitimate purpose related to journalism, education or art. In the United States, Justice Alito's opinion in Matal v. Tam (opens in a new tab) (2017) called speech that demeans on the basis of race, religion or similar grounds hateful, yet protected. Much content no brand would fund is legal, so the brand floor is the brand's to write.
Platform labels follow the same split. YouTube's advertiser friendly guidelines (opens in a new tab) give no ads to hate or harassment toward individuals or groups, and limited ads to content that may be offensive but is used for education, news or a documentary.
How do coded terms evade keyword rules?
Hateful communities change their vocabulary to avoid moderation. Numbers stand for slogans, ordinary words take on second meanings, digits replace letters and spaces split a word. A list written last year misses this year's terms, and a new term can also match innocent pages that use the word in its everyday sense. HateCheck tested such spelling variations; the two research models caught hate with spaces between the letters only 43.9 and 37.6 percent of the time.
Five ways to handle coded hate without a longer blocklist
- Judge the site, not the token. One coded term on a large news site is a quote; the same term across a small site's pages is a pattern.
- Read the surrounding sentences. Coded terms travel with mockery, grievance and a named target group, which a reader recognizes even when the term is new.
- Ask people who track the terms. Community groups, researchers and employee groups can know a phrase before a list does.
- Check symbols separately. Some hateful markers are images or numbers in a username, which text rules do not read.
- Retire stale terms. Remove words that now block more legitimate pages than hateful ones.
Our guide to keyword blocklists explains why word lists overblock. For hate speech the cost is sharper, because a list can end up naming the targeted groups themselves.
How should hate speech be reviewed in other languages?
Hate speech does not translate word for word. Slurs and idioms carry meanings a dictionary misses, and a term that is hateful in one dialect can be neutral in another. The Multilingual HateCheck study (opens in a new tab) (2022) built tests in ten languages: Arabic, Dutch, French, German, Hindi, Italian, Mandarin, Polish, Portuguese and Spanish, with native speaking experts adapting cases rather than translating them. A high performing multilingual model scored from 58.1 for Hindi to 71.5 for Mandarin overall, and did much worse on non hateful cases than hateful ones in every language.
Review checklist for hate speech in another language
- A fluent speaker of the page's language and dialect reviews borderline pages, not a translation tool alone.
- The reviewer knows the communities the brief reaches, including terms they have reclaimed.
- Community and diaspora news sites are judged by their own editorial record.
- Pages that quote hate to report or condemn it keep the news tolerance in every language.
- Overblocking is checked per language, since errors are not spread evenly.
- Each decision is recorded with its reason so another reviewer can repeat it.
Canadian plans can meet this question. A brand may reach readers through community publishers in languages other than English and French, and treating them with more suspicion because a reviewer cannot read them is itself a bias. Our pages on religious content and LGBTQ+ content cover how faith and identity publishing get caught by hate speech rules.
What does a hate speech review look like on a real plan?
Fictional example
A remittance service reaching newcomer communities in Manitoba
A fictional money transfer service wants display ads in Winnipeg and Brandon, Manitoba, beside newcomer guides and community news in English, plus a list of Tagalog and Punjabi community publishers. Its previous agency excluded the whole hate speech category and blocklisted the words immigrant, refugee and deportation.
The new planner excludes the floor and high risk tiers, accepts medium and low, and drops the three words. A community paper's report on racist flyers left at a mosque returns to the plan, with a settlement guide on reporting hate incidents. A forum posting jokes about newcomers stays out at high risk, with the reason recorded. Two staff members who read Tagalog and Punjabi review every flagged page in those languages.
- Floor and high risk excluded; medium and low accepted.
- Group names removed from the blocklist.
- Flagged pages in other languages read by fluent staff.
How does PlacementGraph grade hate speech on a site?
Frequently asked questions
Are slurs in song lyrics or stand up comedy treated as hate speech?
They are graded by treatment like any other content. The legacy framework places dramatic depiction of hate in entertainment at medium risk, not at the floor. YouTube's guidelines add that stating comedic intent is not sufficient, so a routine built on demeaning a protected group can still lose ads. A brand decides whether entertainment at medium risk fits its tolerance.
Does an article about a hate group's history count as hate speech?
It can count as low risk. Educational and informative treatment of hate, and news features on the subject, sit in the framework's lowest tier. The test is whether the page explains and contextualizes the group or recruits for it. Reproduced propaganda without commentary moves a page toward high risk.
Should one hateful article remove a whole website from a plan?
It depends on how the article got there. A hateful opinion piece commissioned and published by the editors says something about the site's standards. A single hateful post on a large site with clear editorial controls may be an exception that a page level exclusion handles. Sample several recent pages before deciding, and record the reason either way.
Do protected trait lists differ between frameworks and platforms?
Yes, slightly. The legacy framework names caste, immigration status, serious disease and victims of violent acts. Google's publisher policies name veteran status and traits tied to systemic discrimination. A brand policy should list the traits it protects in its own words rather than inherit a tool's list.
Sources
- Brand Safety Floor and Suitability Framework (opens in a new tab)Brand Safety Institute
- Google Publisher Policies (opens in a new tab)Google Publisher Policies Help
- Content Taxonomy 3.1 (TSV) (opens in a new tab)IAB Tech Lab
- HateCheck: Functional Tests for Hate Speech Detection Models (opens in a new tab)Röttger et al., ACL 2021
- Multilingual HateCheck: Functional Tests for Multilingual Hate Speech Detection Models (opens in a new tab)Röttger et al., WOAH 2022
- Criminal Code, section 319 (opens in a new tab)Justice Laws Website, Government of Canada
- Matal v. Tam, 582 U.S. (2017) (opens in a new tab)Supreme Court of the United States
- Advertiser-friendly content guidelines (opens in a new tab)YouTube Help
- WFA discontinues GARM (opens in a new tab)World Federation of Advertisers, August 9, 2024
Researched and written by PlacementGraph Research. Read how we research, source and update these pages. Corrections: diana@placementgraph.com.