Features
Part of Critiques of social media: six families, tests and cases
Why enforcement statistics measure the platform, not the harm itself
Statistics behind critiques of social media: why counting harm depends on four hidden decisions and why enforcement figures measure the service.
Counting harm looks like the most objective thing anyone could do about these services, and it is the least. Before a single count exists somebody has decided what an incident is, who gets to say one happened, and what it is being counted against. Each of those decisions changes the answer by more than any finding the count could produce.
This page works through the four decisions and explains why enforcement figures are not harm figures. It names the reports a reader can check — Meta's Community Standards Enforcement Report, TikTok's and YouTube's enforcement reports, the transparency reports kept by X and Discord — and the rules behind them, the European Union's Digital Services Act and the United Kingdom's Online Safety Act 2023.
What to take away
- An incident count is a count of reports that met a definition and reached someone who recorded them. Three filters sit in front of it.
- Reporting rates vary by the thing being reported and by who is affected, so a change in a count can be a change in reporting.
- Removals and enforcement actions measure what a service did, which is a measure of the service, not of what happened.
- Without a denominator, a count is a number with a word after it, and the denominators available are not the ones anybody wants.
The four decisions before any counting
The question
- Definition
- What counts as one incident?
- Who says
- Who is entitled to declare that it happened?
- Detection
- How does an incident come to anyone's attention?
- Denominator
- What is it counted against?
Why it moves the answer
- Definition
- A conversation, a campaign and a single message are all defensible units, and they differ by orders of magnitude
- Who says
- The affected person, a moderator, an automated classifier and a court would each produce a different set
- Detection
- Almost everything is invisible unless somebody reports, and reporting is effortful
- Denominator
- Users, accounts, sessions, messages and person-hours give completely different rates from the same numerator
The first decision alone can change a figure enough to reverse a comparison. Nothing about the counting that follows can repair it, and the decision is usually not stated, which means most published figures cannot be compared with each other even in principle. That is the operationalization problem, and here it lands on the most contested subject in the field.
Four decisions before counting
What counts as one incident?
Who can declare it happened?
How does it come to attention?
Why reporting rates ruin the time series
The count you get is reports, not events. Reports depend on whether people know how to report, whether they believe it accomplishes anything, and whether the process is safe for them.
All three change over time and none of them changes uniformly. Make reporting easier and counts rise, which reads as a worsening problem. Let the process become slow and unresponsive and counts fall, which reads as improvement. Neither reading is about how much is happening.
This affects groups differently, which does the most damage. People who expect the process to work report; those who expect exposure or retaliation do not.
A count under-represents harm against exactly the people most exposed to it, and no weighting corrects this because the gap size is unknown. The general form is participation bias where stakes are highest.
Enforcement figures measure enforcement
The figures that get read as harm figures are the transparency reports themselves: Meta's Community Standards Enforcement Report, TikTok's Community Guidelines Enforcement Report, YouTube's Community Guidelines Enforcement Report, and the transparency reports kept by X and Discord. Each one counts actions a company took on its own services, and each one says so.
In the European Union, the Digital Services Act compels the disclosure, under Articles 15 and 24 for platforms generally and Article 42 for the largest, and the European Commission supervises the largest platforms directly while national Digital Services Coordinators handle the rest. In the United Kingdom, Ofcom enforces the Online Safety Act 2023. The United States imposes no equivalent duty, which is why American figures are voluntary and vary from company to company.
Services publish figures on removals, actions taken and material detected. These are real counts of real operations, and they are frequently read as measurements of harm, which they are not.
A removal figure rises when detection improves, when a policy widens, when a classifier is retrained, or when a campaign floods the system. It falls when any of those reverses.
It measures the service's activity. Its main use is comparing a service with itself over a period when nothing else changed, a condition that essentially never holds.
There is also a reflexive problem: once a figure is published and treated as a performance indicator, it becomes a target, and it stops measuring what it measured. That is Goodhart's law in a setting where the measured party writes the measurement.
What the denominator does
A raw count grows with the population. Any comparison across services, countries or years needs a denominator, and every available denominator is defective.
- Accounts include duplicates, abandonments and automation.
- Active users depends on a definition of active that each service sets for itself.
- Messages or posts changes with formata service where people write more has a larger denominator and a smaller rate for the same amount of harm.
- Person-hours is closest to what anyone means and is measured by nobody.
So the rate depends on the denominator chosen, and the choice is rarely defended. Reading a rate without knowing the denominator produces conclusions about the wrong quantity, which is the ordinary base rate fallacy with the base rate hidden rather than ignored.
What can be said instead
- That a mechanism exists,described plainly. Mechanisms do not need magnitudes to be worth understanding, and the mechanisms here are well described.
- What a service's own process does and does not cover,from its published policy.
- Which body a reader should consultfor figures in their own country, since enforcement and reporting are jurisdictional.
- What would have to be measuredfor a figure to mean what people want it to mean. Saying that clearly is a real contribution and it is cheap.
This is the same position the site takes on counting attention, under benefits of social media, and on counting standing, under online status and identity. The sorting of which complaints could even in principle be answered by a count is set out under critiques of social media.
Common questions
Are transparency reports worthless, then?
No. They are informative about the service's operations and its policy in practice, which is worth knowing. They are not measurements of what happens to people.
Could an outside body count better?
It could count differently, with a stated definition, which would be an improvement in clarity. It would have worse detection, because it cannot see inside the service, and the same reporting problems.
Why does no single regulator cover all of this?
Because enforcement and reporting are jurisdictional. In the European Union the Digital Services Act is applied by national Digital Services Coordinators, with the European Commission supervising the largest platforms itself; in the United Kingdom, Ofcom enforces the Online Safety Act 2023; the United States has no equivalent duty, so its figures stay voluntary. The dates and the changes in all of this are set out under platform timelines.
Is any figure in this area trustworthy?
Figures about a single service's own operations, over a short period, with the definition published, are trustworthy about that. Nothing generalizes, and the confident general figures are the ones to distrust most.
