The T&S Handbook · Glossary

Glossary

Acronyms and Trust & Safety jargon used across the handbook, in plain language.

Back to the handbook
33 terms
The first time a term appears in a chapter, it's underlined; hover or focus it for this same definition.
  • age assuranceAny method of estimating or verifying a user's age, from a typed birth date to an ID check.
  • ban evasionA banned user coming back with a new account.
  • chargebackWhen a card issuer reverses a payment, usually because the cardholder disputed it.
  • classifierA machine-learning model that labels content, for example spam or not spam.
  • confidence intervalThe range the true value probably sits in, given how big your sample was.
  • COPPAThe US Children's Online Privacy Protection Act. It requires verifiable parental consent before collecting personal data from children under 13.
  • CSAMChild sexual abuse material. Illegal to make, possess or share in almost every country. US providers must report it to NCMEC.
  • CyberTiplineNCMEC's system for reports of child sexual exploitation. US providers are required by law to report to it.
  • doxxingPublishing someone's private information, such as their home address, to harass or endanger them.
  • DSAThe EU Digital Services Act. It sets rules for how online services handle illegal content, complaints and transparency in the EU.
  • escalationMoving a case to a more senior or specialist reviewer, such as legal or child safety, because it's severe or unclear.
  • GDPRThe EU General Data Protection Regulation, which governs how personal data is collected, used and kept.
  • gpt-oss-safeguardAn open model from OpenAI that labels content against a policy you write. Free to download; you run it on your own hardware.
  • groomingWhen an adult builds a child's trust over time in order to sexually abuse or exploit them.
  • hash matchingComparing files against a database of known illegal images by their digital fingerprint.
  • incident commandA clear structure for running a crisis: one person in charge, defined roles and a regular rhythm of updates.
  • jailbreakA prompt written to trick an AI into ignoring its safety rules.
  • matched cohortA comparison group chosen to look like the group you're studying, so the main difference is the thing you're measuring.
  • north starOne of the few top-level numbers that show whether users are actually safer.
  • OfcomThe UK communications regulator, which enforces the Online Safety Act.
  • Online Safety ActThe UK law that requires online services to assess and reduce the risk of illegal content and, where children are likely to use them, harm to children. Ofcom enforces it.
  • p50The median: half of cases are faster than this.
  • p9090% of cases are faster than this. It shows your slow tail, which is where incidents come from.
  • precisionOf everything your system acted on, the share that really broke the rules.
  • prevalenceThe share of what users see that breaks a rule.
  • prompt injectionHidden instructions in a document, web page or tool result that try to take control of an AI.
  • recallOf all the rule-breaking content out there, the share your system caught.
  • red teamPeople who deliberately attack your system to find weaknesses before real attackers do.
  • RFPRequest for proposal: the document you send vendors asking how they would meet your requirements, and at what price.
  • sextortionThreatening to share someone's intimate images unless they pay or send more. Teenagers are frequent targets.
  • SLAService-level agreement: the response time you've committed to.
  • statement of reasonsThe explanation the EU DSA requires you to send a user when you restrict their content or account.
  • trusted flaggerAn organization officially recognized under the EU DSA to report illegal content. Its notices must be handled first.