Lecture 10 — Who's watching

Platforms, disclosure rules and content moderation at a scale no human team can read

26 April 2027

Trust and safety work used to mean a large team of human reviewers applying a policy document to a queue. That model is breaking under its own volume, and what’s replacing it is a hybrid most platforms are still improvising: AI-assisted triage feeding human judgement on the genuinely hard cases, a labelling regime for synthetic content that varies wildly between platforms, and a disclosure requirement that only works if creators comply with it voluntarily. A 2025 conference paper put the uncomfortable question in its title: is this “the end of trust and safety” as a distinct discipline, or the beginning of a more specialised one?

We will read one real platform’s synthetic-media labelling policy line by line, then write a better one for a case that policy doesn’t cover cleanly. YouTube’s actual disclosure policy is a good specimen: creators self-report “AI use” during upload, realistic synthetic content requires disclosure while non-realistic content or minor edits don’t, the platform can apply a label automatically and notify the creator when it does, and certain sensitive categories — elections, health, major public events — get a more prominent label than the rest. The rule of thumb underneath all of it: if a viewer could reasonably believe this really happened, it needs a label.

Outline#

• from scaled moderation to specialised review: what changed and why
• disclosure and labelling regimes, compared across platforms
• the cases every current policy handles badly: satire, research, consented synthetic likeness

Further reading#

YouTube’s actual disclosure policy — read this before the session, not a summary of it. This week’s casework asks you to find what it doesn’t cover, and that only works if you’ve read the real thing closely enough to know.