Decisions about what stays on a platform are made by a layered system rather than by anyone reading each post. Understanding the layers explains why outcomes appear inconsistent.
Automation handles the volume
The quantity of material uploaded makes individual human review impossible, so classifiers assess everything and act on high-confidence cases automatically.
These systems are strongest with material that is visually or textually distinctive, and weakest with anything depending on context or intent.
Satire, reclaimed language and documentation of harm all resemble the categories being blocked, which is where automated errors concentrate. Classifiers also perform unevenly across languages, since training material is far thinner outside the largest few.
Human review is a queue with a clock
Uncertain cases go to reviewers, largely employed by outsourced firms, who work through queues against targets measured in seconds per item.
At that pace a reviewer applies a decision tree rather than deliberating, which is what makes outcomes reproducible but also brittle.
The work involves sustained exposure to distressing material, and its psychological cost is now a recognised and litigated aspect of the industry.
Policy is written centrally and interpreted narrowly
Internal guidelines translate broad public rules into specific tests a reviewer can apply consistently, and those documents are far longer than the published policies.
Reviewers are expected to follow the tests rather than exercise judgement, since consistency across thousands of people is the objective.
The consequence is decisions that satisfy the letter of a rule while appearing plainly wrong to anyone who understands the situation. Guidelines are revised continually as edge cases surface, so the same post can be judged differently in successive months.
Escalation is uneven by design
Accounts with large followings, news organisations and political figures are routed to specialist teams with more time and more context.
That tiering is defensible on impact grounds and produces visible asymmetry, where similar posts receive different treatment based on who published them.
Publicity operates as an informal escalation path, which is why appeals that attract attention resolve faster than identical ones that do not.
Errors are unavoidable in both directions
Any classifier can be tuned toward removing more borderline material or leaving more of it, and no setting eliminates both types of error.
Platforms adjust that balance in response to public pressure, and the balance shifts by category and by region rather than being set once.
Users experience the result as arbitrary because the trade-off is invisible, though the underlying decision is deliberate and regularly revisited.