Meta’s claim that hate speech prevalence remains “consistently low” is undermined by its own transparency data.

Speaking to the Royal Commission into Antisemitism and Social Cohesion, Meta’s global policy director Benjamin Good  called prevalence the most important metric and claimed it remained low since the January 2025 changes, when Meta largely stopped its automated removal of hate speech. 

The same transparency report shows Facebook reduced the amount of hate speech it was removing by 79% as part of this change, yet with little to no impact on its prevalence metric.

This should not be possible.

Either Meta removed enormous volumes of hate speech that users would never have seen, or its prevalence metric is not properly capturing the scale of hate speech being shown to users.

Meta’s claim that prevalence remains low cannot be accepted at face value.

OHPI addressed these issues in a submission to the Royal Commission, accessible here.

Today Online Hate Prevention Institute CEO Dr Andre Oboler has responded to Meta’s testimony:

“What Meta is doing is like measuring Mount Everest in light-years and then saying it looks very small. The mountain has not changed, only the scale used to describe it. Meta’s prevalence metric has the same problem. It makes a vast amount of hate speech appear insignificant.

The better question is how many instances of hate speech are seen on Facebook per 100,000 people each year. That is closer to how we measure harm offline, including in crime statistics.

My calculations suggest Facebook users are seeing around 500,000 to 600,000 views of hate speech per 100,000 users each year. Across more than 3 billion Facebook users, that means hate speech is being shown roughly 5.8 billion times a year, yet Meta still calls this low.

This only counts the hate speech Meta recognises. Experts around the world have shown that much of the hate speech reported to Meta is not recognised by the company. The real level must be higher. The volume is almost too large to imagine, yet Meta’s metric makes it seem too small to matter.

In short, the real level of hate speech appears insignificant to Meta due to the scale they’re using to measure it”

 

Dr Oboler’s first report to the Royal Commission set out the calculation in detail. The results were:

• 577,430 views of hateful material per 100,000 Australian Facebook users per year
• 791,758 views of hateful material per 100,000 Australian Instagram users per year

Meta’s claim calls into question the extent to which the public are being given a meaningful account of online hate.

A prevalence figure expressed as a tiny percentage may make the problem sound small. But when applied across billions of users, it means billions of views of hate speech each year. To use Mount Everest again, we may make the world’s highest peak sound inconsequential to alien visitors by saying: “Mount Everest? Oh, it covers 0.0004% of the Earth’s surface.”

 

Media contact:
Online Hate Prevention Institute
Email: joshua@ohpi.org.au
Website: www.ohpi.org.au

Join us on social media
OHPI LinkedIn • OHPI Instagram • OHPI Facebook